bleve Documentation

Technical Documentation: blevesearch/bleve

ℹ️ Provenance: Multi-Tier Ingestion · CodeWiki Reference

Overview (`README.md`) Copy Section

bleve

A modern indexing + search library in GO

Features Copy Chapter

• Index any GO data structure or JSON
• Intelligent defaults backed up by powerful configuration (scorch)
• Supported field types:
  • text, number, datetime, boolean, geopoint, geoshape, IP, vector
• Supported query types:
  • term, phrase, match, match_phrase, prefix, regexp, wildcard, fuzzy
  • term range, numeric range, date range, boolean field
  • compound queries: conjuncts, disjuncts, boolean (must/should/must_not)
  • query string syntax
  • geo spatial search
  • approximate k-nearest neighbors via vector search
  • synonym search
  • hierarchical nested search
• tf-idf / bm25 scoring models
• Hybrid search: exact + semantic
  • Supports RRF (Reciprocal Rank Fusion) and RSF (Relative Score Fusion)
• Result pagination
• Query time boosting
• Search result match highlighting with document fragments
• Aggregations/faceting support:
  • terms facet
  • numeric range facet
  • date range facet

Indexing Copy Chapter

message := struct {
    Id   string
    From string
    Body string
}{
    Id:   "example",
    From: "[email protected]",
    Body: "bleve indexing is easy",
}

mapping := bleve.NewIndexMapping()
index, err := bleve.New("example.bleve", mapping)
if err != nil {
    panic(err)
}
index.Index(message.Id, message)

Querying Copy Chapter

index, _ := bleve.Open("example.bleve")
query := bleve.NewQueryStringQuery("bleve")
searchRequest := bleve.NewSearchRequest(query)
searchResult, _ := index.Search(searchRequest)

Command Line Interface Copy Chapter

To install the CLI for the latest release of bleve, run:

go install github.com/blevesearch/bleve/v2/cmd/bleve@latest

$ bleve --help
Bleve is a command-line tool to interact with a bleve index.

Usage:
  bleve [command]

Available Commands:
  bulk        bulk loads from newline delimited JSON files
  check       checks the contents of the index
  count       counts the number documents in the index
  create      creates a new index
  dictionary  prints the term dictionary for the specified field in the index
  dump        dumps the contents of the index
  fields      lists the fields in this index
  help        Help about any command
  index       adds the files to the index
  mapping     prints the mapping used for this index
  query       queries the index
  registry    registry lists the bleve components compiled into this executable
  scorch      command-line tool to interact with a scorch index

Flags:
  -h, --help   help for bleve

Use "bleve [command] --help" for more information about a command.

Text Analysis Copy Chapter

Bleve includes general-purpose analyzers (customizable) as well as pre-built text analyzers for the following languages:

Arabic (ar), Bulgarian (bg), Catalan (ca), Chinese-Japanese-Korean (cjk), Kurdish (ckb), Danish (da), German (de), Greek (el), English (en), Spanish - Castilian (es), Basque (eu), Persian (fa), Finnish (fi), French (fr), Gaelic (ga), Spanish - Galician (gl), Hindi (hi), Croatian (hr), Hungarian (hu), Armenian (hy), Indonesian (id, in), Italian (it), Dutch (nl), Norwegian (no), Polish (pl), Portuguese (pt), Romanian (ro), Russian (ru), Swedish (sv), Turkish (tr)

Text Analysis Wizard Copy Chapter

bleveanalysis.couchbase.com

Discussion/Issues Copy Chapter

Discuss usage/development of bleve and/or report issues here:

• Github issues
• Google group

License Copy Chapter

Apache License Version 2.0

Overview (`config/README.md`) Copy Section

Bleve Config

NOTE you probably do not need this package. It is only intended for general purpose applications that want to include large parts of Bleve regardless of whether or not the code is directly using it.

General Purpose Applications Copy Chapter

A general purpose application, that must allow users to express the need for Bleve components at runtime can accomplish this by:

import _ "github.com/blevesearch/bleve/config"

Overview (`geo/README.md`) Copy Section

Geo spatial search support in bleve

Latest bleve spatial capabilities are powered by spatial hierarchical tokens generated from s2geometry.
You can find more details about the s2geometry basics here, and explore the extended functionality of our forked golang port of s2geometry lib here.

Users can continue to index and query geopoint field type and the existing queries like,

• Point Distance
• Bounded Rectangle
• Bounded Polygon

as before.

New Spatial Field Type - geoshape Copy Chapter

We have introduced a field type (geoshape) for representing the new spatial types.

Using the new geoshape field type, users can unblock the spatial capabilities
for the geojson shapes like,

• Point
• LineString
• Polygon
• MultiPoint
• MultiLineString
• MultiPolygon
• GeometryCollection

In addition to these shapes, bleve will also support additional shapes like,

• Circle
• Envelope (Bounded box)

To specify GeoJSON data, use a nested field with:

• a field named type that specifies the GeoJSON object type and the type value will be case-insensitive.
• a field named coordinates that specifies the object's coordinates.

"fieldName": { 
              "type": "GeoJSON Type", 
              "coordinates": <coordinates> 
           }

• If specifying latitude and longitude coordinates, list the longitude first and then latitude.
• Valid longitude values are between -180 and 180, both inclusive.
• Valid latitude values are between -90 and 90, both inclusive.
• Shapes would be internally represented as geodesics.
• The GeoJSON specification strongly suggests splitting geometries so that neither of their parts crosses the antimeridian.

Examples for the various geojson shapes representations are as below.

Point Copy Chapter

The following specifies a Point field in a document:

{
  "type": "point",
  "coordinates": [75.05687713623047, 22.53539059204079]
}

Linestring Copy Chapter

The following specifies a Linestring field in a document:

{
  "type": "linestring",
  "coordinates": [
    [77.01416015625, 23.0797317624497],
    [78.134765625, 20.385825381874263]
  ]
}

Polygon Copy Chapter

The following specifies a Polygon field in a document:

{
  "type": "polygon",
  "coordinates": [
    [
      [85.605, 57.207],
      [86.396, 55.998],
      [87.033, 56.716],
      [85.605, 57.207]
    ]
  ]
}

The first and last coordinates must match in order to close the polygon.
And the exterior coordinates have to be in Counter Clockwise Order in a polygon. (CCW)

MultiPoint Copy Chapter

The following specifies a Multipoint field in a document:

{
  "type": "multipoint",
  "coordinates": [
    [-115.8343505859375, 38.45789034424927],
    [-115.81237792968749, 38.19502155795575],
    [-120.80017089843749, 36.54053616262899],
    [-120.67932128906249, 36.33725319397006]
  ]
}

MultiLineString Copy Chapter

The following specifies a MultiLineString field in a document:

{
  "type": "multilinestring",
  "coordinates": [
    [
      [-118.31726074, 35.250105158],
      [-117.509765624, 35.3756141]
    ],
    [
      [-118.696289, 34.624167789],
      [-118.317260742, 35.03899204]
    ],
    [
      [-117.9492187, 35.146862906],
      [-117.6745605, 34.41144164]
    ]
  ]
}

MultiPolygon Copy Chapter

The following specifies a MultiPolygon field in a document:

{
  "type": "multipolygon",
  "coordinates": [
    [
      [
        [-73.958, 40.8003],
        [-73.9498, 40.7968],
        [-73.9737, 40.7648],
        [-73.9814, 40.7681],
        [-73.958, 40.8003]
      ]
    ],
    [
      [
        [-73.958, 40.8003],
        [-73.9498, 40.7968],
        [-73.9737, 40.7648],
        [-73.958, 40.8003]
      ]
    ]
  ]
}

GeometryCollection Copy Chapter

The following specifies a GeometryCollection field in a document:

{
  "type": "geometrycollection",
  "geometries": [
    {
      "type": "multipoint",
      "coordinates": [
        [-73.958, 40.8003],
        [-73.9498, 40.7968],
        [-73.9737, 40.7648],
        [-73.9814, 40.7681]
      ]
    },
    {
      "type": "multilinestring",
      "coordinates": [
        [
          [-73.96943, 40.78519],
          [-73.96082, 40.78095]
        ],
        [
          [-73.96415, 40.79229],
          [-73.95544, 40.78854]
        ],
        [
          [-73.97162, 40.78205],
          [-73.96374, 40.77715]
        ],
        [
          [-73.9788, 40.77247],
          [-73.97036, 40.76811]
        ]
      ]
    },
    {
      "type": "polygon",
      "coordinates": [
        [
          [0, 0],
          [3, 6],
          [6, 1],
          [0, 0]
        ],
        [
          [2, 2],
          [3, 3],
          [4, 2],
          [2, 2]
        ]
      ]
    }
  ]
}

Circle Copy Chapter

If the user wishes to cover a circular region over the earth's surface, then they could use this shape.
A sample circular shape is as below.

{
  "type": "circle",
  "coordinates": [75.05687713623047, 22.53539059204079],
  "radius": "1000m"
}

Circle is specified over the center point coordinates along with the radius.
Example formats supported for radius are:
"5in" , "5inch" , "7yd" , "7yards", "9ft" , "9feet", "11km", "11kilometers", "3nm", "3nauticalmiles", "13mm" , "13millimeters", "15cm", "15centimeters", "17mi", "17miles", "19m" or "19meters".

If the unit cannot be determined, the entire string is parsed and the unit of meters is assumed.

Envelope Copy Chapter

Envelope type, which consists of coordinates for upper left and lower right points of the shape to represent a bounding rectangle in the format [[minLon, maxLat], [maxLon, minLat]].

{
  "type": "envelope",
  "coordinates": [
    [72.83, 18.979],
    [78.508, 17.4555]
  ]
}

GeoShape Query Copy Chapter

Geoshape query support three types/filters of spatial querying capability across those heterogeneous types of documents indexed.

Query Structure Copy Section

{
  "query": {
    "geometry": {
      "shape": {
        "type": "<shapeType>",
        "coordinates": [
          [[]]
        ]
      },
      "relation": "<<filterName>>"
    }
  }
}

shapeType => can be any of the aforementioned types like Point, LineString, Polygon, MultiPoint,
Geometrycollection, MultiLineString, MultiPolygon, Circle and Envelope.

filterName => can be any of the 3 types like intersects, contains and within.

Relation Copy Section

Older Implementation Copy Section

First, all of this geo code is a Go adaptation of the Lucene 5.3.2 sandbox geo support.

Notes Copy Chapter

• All of the APIs will use float64 for lon/lat values.
• When describing a point in function arguments or return values, we always use the order lon, lat.
• High level APIs will use TopLeft and BottomRight to describe bounding boxes. This may not map cleanly to min/max lon/lat when crossing the dateline. The lower level APIs will use min/max lon/lat and require the higher-level code to split boxes accordingly.
• Points and MultiPoints may only contain Points and MultiPoints.
• LineStrings and MultiLineStrings may only contain Points and MultiPoints.
• Polygons or MultiPolygons intersecting Polygons and MultiPolygons may return arbitrary results when the overlap is only an edge or a vertex.
• Circles containing polygon will return a false positive result if all of the vertices of the polygon are within the circle, but the orientation of those points are clock-wise.
• The edges of an Envelope follows the latitude and longitude lines instead of the shortest path on a globe.
• Envelope intersecting queries with LineStrings, MultiLineStrings, Polygons and MultiPolygons implicitly converts the Envelope into a Polygon which changes the curvature of the edges causing inaccurate results for few edge cases.

Overview (`index/scorch/README.md`) Copy Section

scorch

Definitions Copy Chapter

Batch

• A collection of Documents to mutate in the index.

Document

• Has a unique identifier (arbitrary bytes).
• Is comprised of a list of fields.

Field

• Has a name (string).
• Has a type (text, number, date, geopoint).
• Has a value (depending on type).
• Can be indexed, stored, or both.
• If indexed, can be analyzed.
  -m If indexed, can optionally store term vectors.

Scope Copy Chapter

Scorch MUST implement the bleve.index API without requiring any changes to this API.

Scorch MAY introduce new interfaces, which can be discovered to allow use of new capabilities not in the current API.

Implementation Copy Chapter

The scorch implementation starts with the concept of a segmented index.

A segment is simply a slice, subset, or portion of the entire index. A segmented index is one which is composed of one or more segments. Although segments are created in a particular order, knowing this ordering is not required to achieve correct semantics when querying. Because there is no ordering, this means that when searching an index, you can (and should) search all the segments concurrently.

Internal Wrapper Copy Section

In order to accommodate the existing APIs while also improving the implementation, the scorch implementation includes some wrapper functionality that must be described.

\_id field

In scorch, field 0 is prearranged to be named _id. All documents have a value for this field, which is the documents external identifier. In this version the field MUST be both indexed AND stored. The scorch wrapper adds this field, as it will not be present in the Document from the calling bleve code.

NOTE: If a document already contains a field _id, it will be replaced. If this is problematic, the caller must ensure such a scenario does not happen.

Proposed Structures Copy Section

type Segment interface {

  Dictionary(field string) TermDictionary

}

type TermDictionary interface {

  PostingsList(term string, excluding PostingsList) PostingsList

}

type PostingsList interface {

  Next() Posting

  And(other PostingsList) PostingsList
  Or(other PostingsList) PostingsList

}

type Posting interface {
  Number() uint64

  Frequency() uint64
  Norm() float64

  Locations() Locations
}

type Locations interface {
  Start() uint64
  End() uint64
  Pos() uint64
  ArrayPositions() ...
}

type DeletedDocs {

}

type SegmentSnapshot struct {
  segment Segment
  deleted PostingsList
}

type IndexSnapshot struct {
  segment []SegmentSnapshot
}

What about errors?
What about memory mgmnt or context?
Postings List separate iterator to separate stateful from stateless

Mutating the Index Copy Section

The bleve.index API has methods for directly making individual mutations (Update/Delete/SetInternal/DeleteInternal), however for this first implementation, we assume that all of these calls can simply be turned into a Batch of size 1. This may be highly inefficient, but it will be correct. This decision is made based on the fact that Couchbase FTS always uses Batches.

NOTE: As a side-effect of this decision, it should be clear that performance tuning may depend on the batch size, which may in-turn require changes in FTS.

From this point forward, only Batch mutations will be discussed.

Sequence of Operations:

1. For each document in the batch, search through all existing segments. The goal is to build up a per-segment bitset which tells us which documents in that segment are obsoleted by the addition of the new segment we're currently building. NOTE: we're not ready for this change to take effect yet, so rather than this operation mutating anything, they simply return bitsets, which we can apply later. Logically, this is something like:

   go Copy

   foreach segment {
       dict := segment.Dictionary("\_id")
       postings := empty postings list
       foreach docID {
         postings = postings.Or(dict.PostingsList(docID, nil))
       }
     }

   NOTE: it is illustrated above as nested for loops, but some or all of these could be concurrently. The end result is that for each segment, we have (possibly empty) bitset.

2. Also concurrent with 1, the documents in the batch are analyzed. This analysis proceeds using the existing analyzer pool.

3. (after 2 completes) Analyzed documents are fed into a function which builds a new Segment representing this information.

4. We now have everything we need to update the state of the system to include this new snapshot.

   • Acquire a lock
   • Create a new IndexSnapshot
   • For each SegmentSnapshot in the IndexSnapshot, take the deleted PostingsList and OR it with the new postings list for this Segment. Construct a new SegmentSnapshot for the segment using this new deleted PostingsList. Append this SegmentSnapshot to the IndexSnapshot.
   • Create a new SegmentSnapshot wrapping our new segment with nil deleted docs.
   • Append the new SegmentSnapshot to the IndexSnapshot
   • Release the lock

An ASCII art example:

0 - Empty Index

  No segments

  IndexSnapshot
    segments []
    deleted []


  1 - Index Batch [ A B C ]

  segment       0
  numbers   [ 1 2 3 ]
  \_id      [ A B C ]

  IndexSnapshot
    segments [ 0 ]
    deleted [ nil ]


  2 - Index Batch [ B' ]

  segment       0           1
  numbers   [ 1 2 3 ]     [ 1 ]
  \_id      [ A B C ]     [ B ]

  Compute bitset segment-0-deleted-by-1:
            [ 0 1 0 ]

  OR it with previous (nil) (call it 0-1)
            [ 0 1 0 ]

  IndexSnapshot
    segments [  0    1 ]
    deleted  [ 0-1 nil ]

  3 - Index Batch [ C' ]

    segment       0           1      2
    numbers   [ 1 2 3 ]     [ 1 ]  [ 1 ]
    \_id      [ A B C ]     [ B ]  [ C ]

    Compute bitset segment-0-deleted-by-2:
              [ 0 0 1 ]

    OR it with previous ([ 0 1 0 ]) (call it 0-12)
              [ 0 1 1 ]

  Compute bitset segment-1-deleted-by-2:
              [ 0 ]

  OR it with previous (nil)
              still just nil


    IndexSnapshot
      segments [  0    1    2 ]
      deleted  [ 0-12 nil  nil ]

is there opportunity to stop early when doc is found in one segment
also, more efficient way to find bits for long lists of ids?

Searching Copy Section

In the bleve.index API all searching starts by getting an IndexReader, which represents a snapshot of the index at a point in time.

As described in the section above, our index implementation maintains a pointer to the current IndexSnapshot. When a caller gets an IndexReader, they get a copy of this pointer, and can use it as long as they like. The IndexSnapshot contains SegmentSnapshots, which only contain pointers to immutable segments. The deleted posting lists associated with a segment change over time, but the particular deleted posting list in YOUR snapshot is immutable. This gives a stable view of the data.

Term Search

Term search is the only searching primitive exposed in today's bleve.index API. This ultimately could limit our ability to take advantage of the indexing improvements, but it also means it will be easier to get a first version of this working.

A term search for term T in field F will look something like this:

searchResultPostings = empty
  foreach segment {
    dict := segment.Dictionary(F)
    segmentResultPostings = dict.PostingsList(T, segmentSnapshotDeleted)
    // make segmentLocal numbers into global numbers, and flip bits in searchResultPostings
  }

The searchResultPostings will be a new implementation of the TermFieldReader interface.

As a reminder this interface is:

// TermFieldReader is the interface exposing the enumeration of documents
// containing a given term in a given field. Documents are returned in byte
// lexicographic order over their identifiers.
type TermFieldReader interface {
  // Next returns the next document containing the term in this field, or nil
  // when it reaches the end of the enumeration.  The preAlloced TermFieldDoc
  // is optional, and when non-nil, will be used instead of allocating memory.
  Next(preAlloced *TermFieldDoc) (*TermFieldDoc, error)

  // Advance resets the enumeration at specified document or its immediate
  // follower.
  Advance(ID IndexInternalID, preAlloced *TermFieldDoc) (*TermFieldDoc, error)

  // Count returns the number of documents contains the term in this field.
  Count() uint64
  Close() error
}

At first glance this appears problematic, we have no way to return documents in order of their identifiers. But it turns out the wording of this perhaps too strong, or a bit ambiguous. Originally, this referred to the external identifiers, but with the introduction of a distinction between internal/external identifiers, returning them in order of their internal identifiers is also acceptable. ASIDE: the reason for this is that most callers just use Next() and literally don't care what the order is, they could be in any order and it would be fine. There is only one search that cares and that is the ConjunctionSearcher, which relies on Next/Advance having very specific semantics. Later in this document we will have a proposal to split into multiple interfaces:

• The weakest interface, only supports Next() no ordering at all.
• Ordered, supporting Advance()
• And/Or'able capable of internally efficiently doing these ops with like interfaces (if not capable then can always fall back to external walking)

But, the good news is that we don't even have to do that for our first implementation. As long as the global numbers we use for internal identifiers are consistent within this IndexSnapshot, then Next() will be ordered by ascending document number, and Advance() will still work correctly.

NOTE: there is another place where we rely on the ordering of these hits, and that is in the "_id" sort order. Previously this was the natural order, and a NOOP for the collector, now it must be implemented by actually sorting on the "_id" field. We probably should introduce at least a marker interface to detect this.

An ASCII art example:

/* Detailed source-code truncated for AI context efficiency. */

Future improvements

In the future, interfaces to detect these non-serially operating TermFieldReaders could expose their own And() and Or() up to the higher level Conjunction/Disjunction searchers. Doing this alone offers some win, but also means there would be greater burden on the Searcher code rewriting logical expressions for maximum performance.

Another related topic is that of peak memory usage. With serially operating TermFieldReaders it was necessary to start them all at the same time and operate in unison. However, with these non-serially operating TermFieldReaders we have the option of doing a few at a time, consolidating them, dispoting the intermediaries, and then doing a few more. For very complex queries with many clauses this could reduce peak memory usage.

Memory Tracking Copy Section

All segments must be able to produce two statistics, an estimate of their explicit memory usage, and their actual size on disk (if any). For in-memory segments, disk usage could be zero, and the memory usage represents the entire information content. For mmap-based disk segments, the memory could be as low as the size of tracking structure itself (say just a few pointers).

This would allow the implementation to throttle or block incoming mutations when a threshold memory usage has (or would be) exceeded.

Persistence Copy Section

Obviously, we want to support (but maybe not require) asynchronous persistence of segments. My expectation is that segments are initially built in memory. At some point they are persisted to disk. This poses some interesting challenges.

At runtime, the state of an index (it's IndexSnapshot) is not only the contents of the segments, but also the bitmasks of deleted documents. These bitmasks indirectly encode an ordering in which the segments were added. The reason is that the bitmasks encode which items have been obsoleted by other (subsequent or more future) segments. In the runtime implementation we compute bitmask deltas and then merge them at the same time we bring the new segment in. One idea is that we could take a similar approach on disk. When we persist a segment, we persist the bitmask deltas of segments known to exist at that time, and eventually these can get merged up into a base segment deleted bitmask.

This also relates to the topic rollback, addressed next...

Rollback Copy Section

One desirable property in the Couchbase ecosystem is the ability to rollback to some previous (though typically not long ago) state. One idea for keeping this property in this design is to protect some of the most recent segments from merging. Then, if necessary, they could be "undone" to reveal previous states of the system. In these scenarios "undone" has to properly undo the deleted bitmasks on the other segments. Again, the current thinking is that rather than "undo" anything, it could be work that was deferred in the first place, thus making it easier to logically undo.

Another possibly related approach would be to tie this into our existing snapshot mechanism. Perhaps simulating a slow reader (holding onto index snapshots) for some period of time, can be the mechanism to achieve the desired end goal.

Internal Storage Copy Section

The bleve.index API has support for "internal storage". The ability to store information under a separate name space.

This is not used for high volume storage, so it is tempting to think we could just put a small k/v store alongside the rest of the index. But, the reality is that this storage is used to maintain key information related to the rollback scenario. Because of this, its crucial that ordering and overwriting of key/value pairs correspond with actual segment persistence in the index. Based on this, I believe its important to put the internal key/value pairs inside the segments themselves. But, this also means that they must follow a similar "deleted" bitmask approach to obsolete values in older segments. But, this also seems to substantially increase the complexity of the solution because of the separate name space, it would appear to require its own bitmask. Further keys aren't numeric, which then implies yet another mapping from internal key to number, etc.

More thought is required here.

Merging Copy Section

The segmented index approach requires merging to prevent the number of segments from growing too large.

Recent experience with LSMs has taught us that having the correct merge strategy can make a huge difference in the overall performance of the system. In particular, a simple merge strategy which merges segments too aggressively can lead to high write amplification and unnecessarily rendering cached data useless.

A few simple principles have been identified.

• Roughly we merge multiple smaller segments into a single larger one.
• The larger a segment gets the less likely we should be to ever merge it.
• Segments with large numbers of deleted/obsoleted items are good candidates as the merge will result in a space savings.
• Segments with all items deleted/obsoleted can be dropped.

Merging of a segment should be able to proceed even if that segment is held by an ongoing snapshot, it should only delay the removal of it.

Overview (`index/upsidedown/protoc-README.md`) Copy Section

Instructions for generating new go stubs using upsidedown.proto Copy Chapter

1. Download latest of protoc-gen-go

go install google.golang.org/protobuf/cmd/protoc-gen-go@latest

2. To generate upsidedown.pb.go using upsdidedown.proto:

protoc --go_out=. --go_opt=Mindex/upsidedown/upsidedown.proto=index/upsidedown/ index/upsidedown/upsidedown.proto

3. Manually add back Size and MarshalTo methods for BackIndexRowValue, BackIndexTermsEntry, BackIndexStoreEntry to support upside_down.

Overview (`analysis/lang/pl/stempel/javadata/README.md`) Copy Section

javadata

Go library to read data written with java.io.DataOutput

Create And Search Your First Index (`docs/create_and_search_your_first_index.md`) Copy Section

Creating a Bleve Index

A simple how-to example using Bleve in Go to create an index, add documents, and run search queries with results.

package main

import (
    "fmt"
    "log"

    bleve "github.com/blevesearch/bleve/v2"
)

type Document struct {
    ID      string `json:"id"`
    Title   string `json:"title"`
    Content string `json:"content"`
}

func main() {
    indexPath := "example.bleve"
    // Create a new index
    mapping := bleve.NewIndexMapping()
    index, err := bleve.New(indexPath, mapping)
    if err != nil {
        log.Fatal(err)
    }
    defer index.Close()

    // Add documents
    documents := []Document{
        {
            ID:      "doc",
            Title:   "Bleve documentation",
            Content: "Bleve provides full-text search capabilities.",
        },
        {
            ID:      "doc1",
            Title:   "Elasticsearch documentation",
            Content: "Elasticsearch provides full-text search capabilities as well.",
        },
    }

    // Iterate and index the documents
    batch := index.NewBatch()
    for _, doc := range documents {
        batch.Index(doc.ID, doc)
    }
    if err := index.Batch(batch); err != nil {
        log.Fatal(err)
    }

    // Search the created index
    query := bleve.NewQueryStringQuery("bleve")
    searchRequest := bleve.NewSearchRequest(query)
    searchRequest.Explain = true
    searchRequest.Fields = []string{"title", "content"}
    searchResult, err := index.Search(searchRequest)
    if err != nil {
        log.Fatal(err)
    }
    fmt.Println(searchResult)
}

Output Copy Chapter

$ go run main.go

1 matches, showing 1 through 1, took 262.333µs
    1. doc (0.471405)
        title
                Bleve documentation
        content
                Bleve provides full-text search capabilities.

Step-by-Step Breakdown Copy Chapter

1. Index Creation Copy Section

// Create a new index mapping
mapping := bleve.NewIndexMapping()
// Create a new index (this creates a directory on disk)
index, err := bleve.New(indexPath, mapping)

What happens:

• Creates an index mapping with default settings
• Creates a new index directory example.bleve/
• Sets up the underlying storage (Scorch engine by default)

2. Document Indexing Copy Section

// Index a document with a unique ID
err := index.Index("doc", map[string]interface{}{
    "title":   "My Document",
    "content": "This is the document content",
    "author":  "John Doe",
})

What happens:

• Document gets a unique ID (doc)
• Fields are automatically mapped based on their Go types
• Text fields are analyzed (tokenized, lowercased, etc.) based on the chosen mapping (here, the default one)
• Document is stored in the search index

3. Searching Copy Section

// Create a query
query := bleve.NewQueryStringQuery("search terms")
request := bleve.NewSearchRequest(query)

// Execute search
results, err := index.Search(request)

What happens:

• Query string is parsed and analyzed
• Index is searched for matching documents
• Results are scored and ranked by relevance by the algorithm used
• Document metadata and highlights are returned

Working with Existing Indexes Copy Chapter

To open an existing index instead of creating a new one:

// Open existing index
index, err := bleve.Open("example.bleve")
if err != nil {
    log.Fatal(err)
}
defer index.Close()

Different Query Types Copy Chapter

1. Query String Query (Simple) Copy Section

query := bleve.NewQueryStringQuery("golang programming")

2. Match Query (Exact Field) Copy Section

query := bleve.NewMatchQuery("bleve")
query.SetField("title")  // Search only in title field

3. Boolean Query (Complex) Copy Section

mustQuery := bleve.NewMatchQuery("golang")
shouldQuery := bleve.NewMatchQuery("programming")

boolQuery := bleve.NewBooleanQuery()
boolQuery.AddMust(mustQuery)
boolQuery.AddShould(shouldQuery)

4. Range Query (Numeric/Date) Copy Section

minPrice := 20.50
maxPrice := 40.75
query := bleve.NewNumericRangeQuery(&minPrice, &maxPrice)
query.SetField("price")

Advanced Index Configuration Copy Chapter

Custom Field Mapping Copy Section

// We can create customized mappings as well with configurable analyzers
mapping := bleve.NewIndexMapping()

// Text field with custom analyzer
titleMapping := bleve.NewTextFieldMapping()
titleMapping.Analyzer = "en"  // English analyzer

// Numeric field
priceMapping := bleve.NewNumericFieldMapping()

// Date field
dateMapping := bleve.NewDateTimeFieldMapping()

// Document mapping
docMapping := bleve.NewDocumentMapping()
docMapping.AddFieldMappingsAt("title", titleMapping)
docMapping.AddFieldMappingsAt("price", priceMapping)
docMapping.AddFieldMappingsAt("created_at", dateMapping)

// Add to index mapping
mapping.AddDocumentMapping("product", docMapping)

// Create index with custom mapping
index, err := bleve.New("products.bleve", mapping)

Batch Operations Copy Section

For better performance when indexing many documents, we can do indexing in batches:

batch := index.NewBatch()

for _, doc := range documents {
    batch.Index(doc.ID, doc)
}

// Execute batch
err := index.Batch(batch)

Custom Query (`docs/custom_query.md`) Copy Section

Custom Queries: `custom_filter` and `custom_score`

Bleve exposes two query nodes for embedding-defined per-hit logic:

• custom_filter: keep or drop each candidate hit.
• custom_score: mutate each candidate hit score.

Bleve itself only executes callbacks that were already provided/attached by the
embedding application. It does not interpret any
embedder-specific payload such as callback source, "params", or requested fields.

Query Objects Copy Chapter

CustomFilterQuery contains:

• a child query
• a searcher.CustomFilterFunc callback attached by the embedder

CustomScoreQuery contains:

• a child query
• a searcher.CustomScoreFunc callback attached by the embedder

Constructors Copy Chapter

Embedding applications can construct the nodes directly with callbacks:

filterQuery := query.NewCustomFilterQueryWithFilter(childQuery,
    func(d *search.DocumentMatch) bool {
        return true
    },
    nil, // fields; nil when callback does not need field values
    nil, // payload; nil when no source/params round-trip is needed
)

scoreQuery := query.NewCustomScoreQueryWithScorer(childQuery,
    func(d *search.DocumentMatch) float64 {
        return d.Score
    },
    nil, // fields
    nil, // payload
)

Constructor arguments after the callback are:

• fields: field names to preload into d.Fields via doc values before the callback runs
• payload: non-query custom attributes that must survive JSON round-trips (for example distributed fan-out)

Pass nil for either when not needed:

payload := map[string]interface{}{
    "params": map[string]interface{}{"weight": 0.05},
    "source": "function score(doc, params) { ... }",
}

scoreQuery := query.NewCustomScoreQueryWithScorer(childQuery,
    func(d *search.DocumentMatch) float64 {
        f := d.Fields
        return d.Score + (f["abv"].(float64) * 0.05)
    },
    []string{"abv", "ibu"},
    payload,
)

Payload rules:

• pass only non-query custom-node attributes like source and params
• do not include query in payload; query always comes from the child query argument
• do not include fields in payload; fields come from the dedicated constructor argument
• pass nil as the fourth argument when payload round-trip is not needed

Field Access and Doc Values Requirement Copy Chapter

When the fields constructor argument is non-empty, bleve loads the listed
field values into DocumentMatch.Fields before invoking the callback. Field
values are read from doc values (column-oriented, memory-mapped), not
stored fields.

Doc values must be enabled for every field listed in "fields". If a field
does not have doc values enabled in the index mapping, its value will be silently
absent from DocumentMatch.Fields at callback time — no error is raised, the
field simply won't appear. This can cause filters to drop all hits or scores to
compute incorrectly if the callback assumes the field is present.

Doc values are enabled by default for all field types (DocValues: true) and
for dynamic mappings (DocValuesDynamic: true), so no extra configuration is
needed unless these defaults have been explicitly disabled. If they have been
disabled, re-enable them:

fm := mapping.NewTextFieldMapping()
fm.DocValues = true                    // default is true; only needed if previously disabled
imap.DefaultMapping.AddFieldMappingsAt("genre", fm)

For dynamic mappings, ensure docvalues_dynamic has not been set to false on
the document mapping.

Datetime fields require an explicit datetime mapping. Dynamic mapping indexes
date-like strings as text, which means the standard analyzer tokenizes the value
(e.g. "2010-07-22 20:00:20" becomes ["2010", "07", "22", ...]). The doc
value then contains tokens instead of the original date, making comparisons
fail. To use datetime fields in custom_filter or custom_score callbacks, add an
explicit field mapping with "type": "datetime" in the index definition.

Field type decoding Copy Section

At the bleve layer, doc values are decoded based on the field's mapping type:

Runtime Behavior Copy Chapter

• Bleve builds the child query searcher first.
• If fields is non-empty, a DocValueReader is created for those fields.
• custom_filter wraps the child with CustomFilterSearcher.
• custom_score wraps the child with CustomScoreSearcher.
• On each hit, doc values are loaded into DocumentMatch.Fields, then the callback is invoked.
• The callback receives only *search.DocumentMatch; it does not receive SearchContext.

Error Cases Copy Chapter

• missing child query:
  • "custom filter query must have a query"
  • "custom score query must have a query"
• missing bound callback:
  • "custom filter query must have a filter callback"
  • "custom score query must have a score callback"

Fast Merge (`docs/fast_merge.md`) Copy Section

Fast Merge

• v2.6.0 comes with the support for a feature called "fast merge" where we first train the index on a vector dataset and then build the vector index using this trained information of the centroid layout.
• This is a separate feature when compared to the existing behavior where we were performing the merge in a very naive fashion of reconstructing the participating vector indexes, re-training and then adding back the vectors into the index. Fast merge essentially merges the corresponding centroid cells' data vectors in a block-wise fashion without doing the expensive operations. For this, a centroid/trained segment template needs to be built using enough random samples from a dataset prior to building segments with the actual index data.
• A trade-off with this feature is if new data were introduced that is not available at the time of index introduction, a data drift can develop resulting in reduced recall.
• This feature underneath the hood utilizes the existing merge_from API in our fork of the FAISS codebase.

Support Copy Chapter

• This feature is supported primarily for the IVF family of indexes. This happens when
  • the field mapping has optimization type as ivf,rabitq or when using bivf-sq8/bivf-flat for binary quantization.
  • when the above optimizations aren't used but the scale of data exceeds 10K vectors.

Usage Copy Chapter

The feature can be enabled by first passing a key-value pair in the config part while creating a new index. If the flag is false, then the behavior falls back to the more expensive naive merge.

kvConfig := map[string]interface{}{
  "vector_index_fast_merge": true,
}

index, err := bleve.NewUsing("example.bleve", bleve.NewIndexMapping(), bleve.Config.DefaultIndexType, bleve.Config.DefaultMemKVStore, kvConfig)
if err != nil {
  log.Fatal(err)
}

The user should now "train" the index on a random sample of the vector dataset they're planning to index and search.

• It's completely up to the user as to how much data they want to use for training, controlling the batch size used while training and also marking whether the training is complete.
• NOTE: User must index their data only after marking the training as complete, otherwise the batch won't be indexed.

The training is done using the new Train() API which takes the existing Batch construct we have

batch := index.NewBatch()
for _, doc := range trainingDocuments {
  batch.Index(doc.ID, doc)
}

// train the index on the batch of data
// NOTE: the training can be done in an incremental manner as well, by using the same Train() API but repeatedly calling it on a particular batch of data.
if err := index.Train(batch); err != nil {
  log.Fatal(err)
}

batch.Reset()
batch.SetInternal(util.BoltTrainCompleteKey, []byte("true"))
if err := index.Train(batch); err != nil {
  log.Fatal(err)
}

// at this point the index is trained completely, and any training related constructs in the background will be cleaned up

Disclaimer Copy Chapter

• This feature is primarily meant for the use case where the user is aware of how much data they want to index and also for a read-heavy workload and little to no updates on the index itself.
• The intention of the feature is to be able to quickly index a massive scale of data on an index in an efficient manner and perform search on it.
• Without this feature, i.e. when the index build happens without a prior training phase
  • The user wouldn't have to worry about use cases where the dataset is continuously updated with new "type" of vector. This is because each merge cycle would do the training afresh.
  • The user doesn't have a lag in indexing the data either, they can start ingesting the data immediately.
• Based on what's mentioned above, when it comes to update and delete type of workloads on the dataset it's extremely difficult to detect when the data drift will occur. So we end up falling back to the naive way of reconstructing + re-training.

Geo (`docs/geo.md`) Copy Section

Geo spatial search

Redirect to geo/README.md

Hierarchy (`docs/hierarchy.md`) Copy Section

Hierarchical nested search

• v2.6.0 (and after) will come with support for array indexing and hierarchical nested search.
• We've achieved this by embedding nested documents within our bleve (scorch) indexes.
• Usage of zap file format: v17. Here we preserve hierarchical document relationships within segments, continuing to conform to the segmented architecture of scorch.

Supported Copy Chapter

• Indexing Arrays allows specifying fields that contain arrays of objects. Each object in the array can have its own set of fields, enabling the representation of hierarchical data structures within a single document.

{
    "id": "1",
    "name": "John Doe",
    "addresses": [
        {
            "type": "home",
            "street": "123 Main St",
            "city": "Hometown",
            "zip": "12345"
        },
        {
            "type": "work",
            "street": "456 Corporate Blvd",
            "city": "Metropolis",
            "zip": "67890"
        }
    ]
}

• Multi-level arrays: arrays can contain objects that themselves have array fields, allowing for deeply nested structures, such as a list of projects, each with its own list of tasks.

{
    "id": "2",
    "name": "Jane Smith",
    "projects": [
        {
            "name": "Project Alpha",
            "tasks": [
                {"title": "Task 1", "status": "completed"},
                {"title": "Task 2", "status": "in-progress"}
            ]
        },
        {
            "name": "Project Beta",
            "tasks": [
                {"title": "Task A", "status": "not-started"},
                {"title": "Task B", "status": "completed"}
            ]
        }
    ]
}

• Multiple arrays: A document can have multiple fields that are arrays, each representing different hierarchical data, such as a list of phone numbers and a list of email addresses.

{
    "id": "3",
    "name": "Alice Johnson",
    "phones": [
        {"type": "mobile", "number": "555-1234"},
        {"type": "home", "number": "555-5678"}
    ],
    "emails": [
        {"type": "personal", "address": "[email protected]"},
        {"type": "work", "address": "[email protected]"}
    ]
}

• Hybrid arrays: Multi-level and multiple arrays can be combined within the same document to represent complex hierarchical data structures, such as a company with multiple departments, each having its own list of employees and projects.

{
    "id": "doc1",
    "company": {
        "id": "c1",
        "name": "TechCorp",
        "departments": [
            {
                "name": "Engineering",
                "budget": 2000000,
                "employees": [
                    {"name": "Alice", "role": "Engineer"},
                    {"name": "Bob", "role": "Manager"}
                ],
                "projects": [
                    {"title": "Project X", "status": "ongoing"},
                    {"title": "Project Y", "status": "completed"}
                ]
            },
            {
                "name": "Sales",
                "budget": 300000,
                "employees": [
                    {"name": "Eve", "role": "Salesperson"},
                    {"name": "Mallory", "role": "Manager"}
                ],
                "projects": [
                    {"title": "Project A", "status": "completed"},
                    {"title": "Project B", "status": "ongoing"}
                ]
            }
        ],
        "locations": [
            {"city": "Athens","country": "Greece"},
            {"city": "Berlin", "country": "USA"}
        ]
    }
}

• Earlier versions of Bleve only supported flat arrays of primitive types (e.g., strings, numbers), and would flatten nested structures, losing the hierarchical relationships, so the above complex documents could not be accurately represented or queried. For example, the "employees" and "projects" fields within each department would be flattened, making it impossible to associate employees with their respective departments.

• From v2.6.0 onwards, Bleve allows for accurate representation and querying of complex nested structures, preserving the relationships between different levels of the hierarchy, across multi-level, multiple and hybrid arrays.

• The addition of nested document mappings enables the definition of fields that contain arrays of objects, giving the option to preserve the hierarchical relationships within the array during indexing. Having nested as false (default) will flatten the objects within the array, losing the hierarchy, which was the earlier behavior.

{
    "departments": {  
        "dynamic": false,
        "enabled": true,
        "nested": true,
        "properties": {
            "employees": {
                "dynamic": false,
                "enabled": true,
                "nested": true 
            },
            "projects": {
                "dynamic": false,
                "enabled": true,
                "nested": true
            }
        }
    },
    "locations": {
        "dynamic": false,
        "enabled": true,
        "nested": true
    }
}

• Any Bleve query (e.g., match, phrase, term, fuzzy, numeric/date range, etc.) can be executed against fields within nested documents, with no special handling required. The query processor will automatically traverse the nested structures to find matches. Additional search constructs
  like vector search, synonym search, hybrid and pre-filtered vector search integrate seamlessly with hierarchy search.

• Conjunction Queries (AND queries) and other queries that depend on term co-occurrence within the same hierarchical context will respect the boundaries of nested documents. This means that terms must appear within the same nested object to be considered a match. For example, a conjunction query searching for an employee named "Alice" with the role "Engineer" within the "Engineering" department will only return results where both name and role terms are found within the same employee object, which is itself within an "Engineering" department object.

• Some other search constructs will have enhanced precision with hierarchy search.

  • Field-Level Highlighting: Only fields within the matched nested object are retrieved and highlighted, ensuring highlights appear in the correct hierarchical context. For example, a match in departments[name=Engineering].employees highlights only employees in that department.

  • Nested Faceting / Aggregations: Facets are computed within matched nested objects, producing context-aware buckets. E.g., a facet on departments.projects.status returns ongoing or completed only for projects in matched departments.

  • Sorting by Nested Fields: Sorting can use fields from the relevant nested object, e.g., ordering companies by departments.budget sorts based on the budget of the specific matched department, not unrelated departments.

• Vector Search (KNN / Multi-KNN): When a document contains an array of objects with vector/multi-vector fields, the final document score and ranking are identical whether or not the array is marked as nested. In both cases, the highest-scoring vector is selected; either directly from the array (non-nested) or from the best-matching nested object with its score bubbled up to the parent document.

• Pre-Filtered Vector Search: When vector search is combined with filters on fields inside a nested array, the filters are applied first to pick which nested items are eligible. Vector similarity is then computed only on the vector fields of those filtered nested objects. For example, if departments.employees is a nested array, a pre-filtered KNN query for employees with a skills_vector matching machine learning engineer, a role of Manager, and belonging to the Sales department will first narrow the candidate set to only employees who meet the requirement, and then compute vector similarity on the skills_vector of that filtered subset. This ensures that vector search results come only from the employees that satisfy the filter, and not from unrelated employees in other departments.

Indexing Copy Chapter

Below is an example of using the Bleve API to index documents with hierarchical structures, using hybrid arrays and nested mappings.

/* Detailed source-code truncated for AI context efficiency. */

Querying Copy Chapter

/* Detailed source-code truncated for AI context efficiency. */

Index Update (`docs/index_update.md`) Copy Section

Ability to reduce downtime during index mapping updates

• v2.5.4 (and after) will come with support for deleting or modifying any field mapping in the index mapping without requiring a full rebuild of the index
• We do this by storing which portions of the field that have to be deleted within zap and then lazily executing the deletion during subsequent merging of the segments

Usage Copy Chapter

While opening an index, if an updated mapping is provided as a string under the key updated_mapping within the runtimeConfig parameter of OpenUsing, then we open the index and try to update it to use the new mapping provided.

If the update fails, the index is unchanged and an error is returned explaining why the update was unsuccessful.

What can be deleted and what can't be deleted? Copy Chapter

Fields can be partially deleted by changing their Index, Store, and DocValues parameters from true to false, or completely removed by deleting the field itself.

Additionally, document mappings can be deleted either by fully removing them from the index mapping or by setting the Enabled value to false, which deletes all fields defined within that mapping.

However, if any of the following conditions are met, the index is considered non-updatable.

• Any additional fields or enabled document mappings in the new index mapping
• Any changes to IncludeInAll, type, IncludeTermVectors and SkipFreqNorm
• Any document mapping having its enabled value changing from false to true
• Text fields with a different analyzer or date time fields with a different date time format
• Vector and VectorBase64 fields changing dims, similarity or vectorIndexOptimizedFor
• Any changes when field is part of _all
• Full field deletions when it is covered by any dynamic setting (Index, Store or DocValues Dynamic)
• Any changes to dynamic settings at the top level or any enabled document mapping
• If multiple fields sharing the same field name either from different type mappings or aliases are present, then any non-compatible changes across all of these fields

How to enforce immediate deletion? Copy Chapter

Since the deletion is only done during merging, a force merge may be used to completely remove the stale data.

Sample code to update an existing index Copy Chapter

newMapping := `<Updated Index Mapping>`
config := map[string]interface{}{
    "updated_mapping": newMapping,
}
index, err := bleve.OpenUsing("<Path to Index>", config)
if err != nil {
    return err
}

Pagination (`docs/pagination.md`) Copy Section

Pagination

Why pagination matters Copy Chapter

Search queries can match many documents. Pagination lets you fetch and display results in chunks, keeping responses small and fast.

By default, Bleve returns the first 10 hits sorted by relevance (score), highest first.

Two pagination modes Copy Chapter

• From/Size: simple and stateless; cost grows with page depth.
• SearchAfter/SearchBefore: efficient for deep paging; requires passing sort keys from the previous page.

Both modes can be used with any valid sort.

`Size`/`From` Copy Chapter

Offset-based pagination uses Size (page length) and From (number of hits to skip). Bleve collects at least Size + From ordered results, then returns the Size slice starting at From.

JSON example:

{
  "query": { "match": "California" },
  "sort": ["-_score"],
  "size": 5,
  "from": 10
}

The result would be 5 hits starting from the 11th hit.

When to use:

• Simple, stateless pagination for shallow pages.
• Avoid for deep pages, as memory grows with From for deeper pages.

`SearchAfter` and `SearchBefore` Copy Chapter

This returns the next (or previous) page based on a boundary defined by the sort keys of a specific hit. This keeps resource usage proportional to the page size, even for deep pages.

Rules:

• Use either SearchAfter (forward) or SearchBefore (backward), not both at once.
• The length of SearchAfter/SearchBefore must match the length of Sort.
• Values are strings representing the sort keys, in the same order as Sort.
• Keep the same query and sort across pages for consistent navigation.

Where do sort keys come from?

• Each hit includes Sort (and DecodedSort from Bleve v2.5.2). Take the last hit's sort keys for SearchAfter, or the first hit's sort keys for SearchBefore.
• If the field/fields to be searched over are numeric, datetime or geo, the values in the Sort field may have garbled values; this is because of how Bleve represents such data types internally. To use such fields as sort keys, use the DecodedSort field, which decodes the internal representations. This feature is available from Bleve v2.5.4.

When using DecodedSort, the Sort array in the search request needs to explicitly declare the type of the field for proper decoding. Hence, the Sort array must contain either SortField objects (for numeric and datetime) or SortGeoDistance objects (for geo) rather than just the field names. More info on SortField and SortGeoDistance can be found in sort_facet.md.

Forward pagination over _id and _score:

{
  "query": { "match": "California" },
  "sort": ["_id", "_score"],
  "search_after": ["hotel_10180", "0.998"],
  "size": 3
}

Backward pagination over _id and _score:

{
  "query": { "match": "California" },
  "sort": ["_id", "_score"],
  "search_before": ["hotel_17595", "0.623"],
  "size": 4
}

Pagination using numeric, datetime, and geo fields. Notice how we specify the sort objects, with the "type" field explicitly declared in case of numeric and datetime:

{
  "query": {
    "match_all": {}
  },
  "size": 10,
  "sort": [
    {"by": "field", "field": "price", "type": "number"},
    {"by": "field", "field": "created_at", "type": "date"},
    {"by": "geo_distance", "field": "location", "location": {"lat": 40.7128, "lon": -74.0060}}
  ],
  "search_after": ["99.99", "2023-10-15T10:30:00Z", "5.2"]
}

Total Sort Order Copy Chapter

Pagination is deterministic. Ensure your Sort defines a total order, so that documents with the same sort keys are not left out:

• Sort strings can be field names (prefix with - for descending), "_score", or "_id".
• Always include a stable tie-breaker as the last key, typically "_id".
• Examples:
  • ["country", "-age", "_id"]
  • ["-_score", "_id"] (default score desc with a tie-breaker)

Performance guidance Copy Chapter

• Offset pagination cost grows with From (collects at least Size + From results before slicing).
• SearchAfter/SearchBefore keeps memory and network proportional to Size.
• For large datasets and deep navigation, prefer using SearchAfter and SearchBefore.

Persister (`docs/persister.md`) Copy Section

Scorch Index Memory and File Management

Memory Management Copy Chapter

When data is indexed in Scorch — using either the index.Index() or index.Batch() API — it is added as part of an in-memory "segment". Memory management in Scorch indexing mainly relates to handling these in-memory segments during workloads that involve inserts or updates.

In scenarios with a continuous stream of incoming data, a large number of in-memory segments can accumulate over time. This is where the persister component comes into play—its job is to flush these in-memory segments to disk.

Starting with v2.5.0, Scorch supports parallel flushing of in-memory segments to disk, where the persister checks the total in-memory data and distributes the flush across multiple workers. This feature is disabled by default and can be enabled using two configuration options:

• NumPersisterWorkers: This factor decides how many maximum workers can be spawned to flush out the in-memory segments. Each worker will work on a disjoint subset of segments, merge them, and flush them out to the disk. By default the persister deploys only one worker.
• MaxSizeInMemoryMergePerWorker: This config decides what's the maximum amount of input data in bytes a single worker can work upon. By default this value is equal to 0 which means that this config is disabled and the worker tries to merge all the data in one shot. Also note that it's imperative that the user set this config if NumPersisterWorkers > 1.

If the index is tuned to have a higher NumPersisterWorkers value, the memory can potentially drain out faster and ensure stronger consistency behavior — but there would be a lot of on-disk files, and the background merger would experience the pressure of managing this large number of files, which can be resource-intensive.

• Tuning this config is very dependent on the available CPU resources. Something to keep in mind here is that the process's RSS can increase if the number of workers — and each of them working upon a large amount of data — is high.

Increasing the MaxSizeInMemoryMergePerWorker value would mean that each worker acts upon a larger amount of data and spends more time merging and flushing it out to disk — which can be healthy behavior in terms of I/O, although it comes at the cost of time.

• Changing this config is use-case dependent; for example, in use cases where the payload or per-doc size is generally large (for example vector use cases), it would be beneficial to have a larger value for this.

So, having the ideal values for these two configs is definitely dependent on the use case and can involve a bunch of experiments, keeping the resource usage in mind.

File Management Copy Chapter

The persister introducing some number of file segments into the system would change the state of the system, and the merger would wake up and try to manage these on-disk files.

Management of these files is crucial when it comes to query latency because a higher number of files would dictate searching through a larger number of files and also higher read amplification to some extent, because the backing data structures can potentially be compacted in size across files.

The merger sees the files on disk and plans out which segments to merge so that the final layout of segment tiers (each tier having multiple files), which grow in a logarithmic way (the chances of larger tiers growing in number would decrease), is maintained. This also implies that deciding this first-tier size becomes important in deciding the number of segment files across all tiers.

Starting with v2.5.0, this first-tier size is dependent on the file size using the FloorSegmentFileSize config, because that's a better metric to consider (unlike the legacy live doc count metric) in order to ensure that the behavior is in line with the use case and aware of the payload/doc size.

• This config can also be tuned to dictate how the I/O behaviour should be within an index. While tuning this config, it should be in proportion to the MaxSizeInMemoryMergePerWorker since that dictates the amount of data flushed out per flush.
• The observation here is that FloorSegmentFileSize is less than MaxSizeInMemoryMergePerWorker and for an optimal I/O during indexing, this value can be set close to MaxSizeInMemoryMergePerWorker/6.

Setting a Persister/Merger Config in Index Copy Chapter

The configs are set via the kvConfig parameter in the NewUsing() or OpenUsing() API:

// setting the persister and merger configs
kvConfig := map[string]interface{}{
    "scorchPersisterOptions": map[string]interface{}{
        "numPersisterWorkers":           4,
        "maxSizeInMemoryMergePerWorker": 20000000,
    },
    "scorchMergePlanOptions": map[string]interface{}{
        "floorSegmentFileSize": 10000000,
    },
}
// passing the config to the index
index, err := bleve.NewUsing("example.bleve", bleve.NewIndexMapping(), bleve.Config.DefaultIndexType, bleve.Config.DefaultMemKVStore, kvConfig)
if err != nil {
    panic(err)
}

Query Openapi Spec.Yaml (`docs/query-openapi-spec.yaml`) Copy Section

openapi: 3.0.3
info:
title: Bleve JSON Query Language
A comprehensive specification for the Bleve Query JSON Language.

Bleve is a text indexing library for Go that provides full-text search capabilities.
This specification defines the JSON structure for all supported query types in Bleve.

## Query Types Supported:

### Basic Queries
- **Term Query**: Exact term matching
- **Match Query**: Full-text search with analysis
- **Match Phrase Query**: Phrase matching with proximity
- **Prefix Query**: Prefix matching
- **Wildcard Query**: Pattern matching with * and ?
- **Regexp Query**: Regular expression matching
- **Fuzzy Query**: Fuzzy matching with edit distance
- **Query String Query**: Human-readable query syntax

### Range Queries
- **Numeric Range Query**: Numeric value ranges
- **Date Range Query**: Date/time ranges
- **Term Range Query**: Lexicographic term ranges

### Boolean Queries
- **Boolean Query**: Must/should/must_not combinations
- **Conjunction Query**: All queries must match
- **Disjunction Query**: Any query can match

### Special Queries
- **Match All Query**: Matches all documents
- **Match None Query**: Matches no documents
- **Bool Field Query**: Boolean field matching
- **Doc ID Query**: Document ID matching

### Geographic Queries
- **Geo Distance Query**: Distance-based geographic search
- **Geo Bounding Box Query**: Bounding box geographic search
- **Geo Bounding Polygon Query**: Polygon geographic search
- **Geo Shape Query**: Complex geographic shape matching

### Vector Queries
- **KNN Query**: K-nearest neighbors vector search

### Network Queries
- **IP Range Query**: IP address range matching

## Common Properties

All queries support:
  - **boost**: Query boost value (float, default 1.0)
  - **field**: Field to search (string, optional)

## Query Parsing

Bleve automatically detects query types based on the presence of specific JSON fields.
The query parser uses a heuristic approach to determine the query type from the JSON structure.

version: 2.0.0
contact:
name: Bleve Search
url: https://github.com/blevesearch/bleve
license:
name: Apache 2.0
url: https://www.apache.org/licenses/LICENSE-2.0

tags:

• name: Basic Queries

• name: Range Queries

• name: Boolean Queries

• name: Special Queries

• name: Geographic Queries

• name: Vector Queries

• name: Network Queries
  components:
  schemas:

  Base Query Schema

  Query:
  type: object
  properties:
  boost:
  type: number
  format: float
  default: 1.0
  minimum: 0
  field:
  type: string
  
  additionalProperties: true

  Basic Queries

  TermQuery:
  allOf:
  - type: object
  properties:
  term:
  type: string
  example: "search"
  required:
  - term
  example:
  term: "search"
  field: "title"
  boost: 2.0

  MatchQuery:
  allOf:
  - type: object
  properties:
  match:
  type: string
  example: "full text search"
  operator:
  type: string
  enum: [and, or]
  default: "or"
  fuzziness:
  oneOf:
  - type: integer
  minimum: 0
  maximum: 2
  - type: string
  enum: ["auto"]
  default: 0
  prefix_length:
  type: integer
  minimum: 0
  default: 0
  analyzer:
  type: string
  required:
  - match
  example:
  match: "full text search"
  field: "content"
  operator: "and"
  fuzziness: 1

  MatchPhraseQuery:
  allOf:
  - type: object
  properties:
  match_phrase:
  type: string
  example: "exact phrase match"
  fuzziness:
  oneOf:
  - type: integer
  minimum: 0
  maximum: 2
  - type: string
  enum: ["auto"]
  default: 0
  analyzer:
  type: string
  required:
  - match_phrase
  example:
  match_phrase: "exact phrase match"
  field: "content"
  fuzziness: "auto"

  PrefixQuery:
  allOf:
  - type: object
  properties:
  prefix:
  type: string
  example: "pref"
  required:
  - prefix
  example:
  prefix: "pref"
  field: "title"

  WildcardQuery:
  allOf:
  - type: object
  properties:
  wildcard:
  type: string
  example: "tet"
  required:
  - wildcard
  example:
  wildcard: "tet"
  field: "title"

  RegexpQuery:
  allOf:
  - type: object
  properties:
  regexp:
  type: string
  example: "te[st]"
  required:
  - regexp
  example:
  regexp: "te[st]"
  field: "title"

  FuzzyQuery:
  allOf:
  - type: object
  properties:
  term:
  type: string
  example: "search"
  fuzziness:
  oneOf:
  - type: integer
  minimum: 0
  maximum: 2
  - type: string
  enum: ["auto"]
  default: 1
  prefix_length:
  type: integer
  minimum: 0
  default: 0
  required:
  - term
  example:
  term: "search"
  field: "title"
  fuzziness: "auto"

  QueryStringQuery:
  allOf:
  - type: object
  properties:
  query:
  type: string
  example: "title:search AND content:full text"
  required:
  - query
  example:
  query: "title:search AND content:full text"

  Range Queries

  NumericRangeQuery:
  allOf:
  - type: object
  properties:
  min:
  type: number
  format: float
  max:
  type: number
  format: float
  inclusive_min:
  type: boolean
  default: true
  inclusive_max:
  type: boolean
  default: false
  example:
  min: 10.5
  max: 100.0
  inclusive_min: true
  inclusive_max: false
  field: "price"

  DateRangeQuery:
  allOf:
  - type: object
  properties:
  start:
  type: string
  format: date-time
  end:
  type: string
  format: date-time
  inclusive_start:
  type: boolean
  default: true
  inclusive_end:
  type: boolean
  default: false
  example:
  start: "2023-01-01T00:00:00Z"
  end: "2023-12-31T23:59:59Z"
  inclusive_start: true
  inclusive_end: true
  field: "created_at"

  TermRangeQuery:
  allOf:
  - type: object
  properties:
  min:
  type: string
  max:
  type: string
  inclusive_min:
  type: boolean
  default: true
  inclusive_max:
  type: boolean
  default: false
  example:
  min: "a"
  max: "m"
  inclusive_min: true
  inclusive_max: false
  field: "name"

  Boolean Queries

  BooleanQuery:
  allOf:
  - type: object
  properties:
  must:
  oneOf:
  - type: array
  items:
  $ref: '#/components/schemas/Query'
  - $ref: '#/components/schemas/Query'
  should:
  oneOf:
  - type: array
  items:
  $ref: '#/components/schemas/Query'
  - $ref: '#/components/schemas/Query'
  must_not:
  oneOf:
  - type: array
  items:
  $ref: '#/components/schemas/Query'
  - $ref: '#/components/schemas/Query'
  filter:
  oneOf:
  - type: array
  items:
  $ref: '#/components/schemas/Query'
  - $ref: '#/components/schemas/Query'
  min_should:
  type: integer
  minimum: 0
  example:
  must:
  - term: "important"
  field: "status"
  - match: "urgent"
  field: "title"
  should:
  - match: "critical"
  field: "content"
  - match: "priority"
  field: "tags"
  must_not:
  - term: "archived"
  field: "status"
  filter:
  - term: "published"
  field: "status"
  boost: 1.5

  ConjunctionQuery:
  allOf:
  - type: object
  properties:
  conjuncts:
  type: array
  items:
  $ref: '#/components/schemas/Query'
  minItems: 1
  required:
  - conjuncts
  example:
  conjuncts:
  - term: "search"
  field: "title"
  - match: "full text"
  field: "content"

  DisjunctionQuery:
  allOf:
  - type: object
  properties:
  disjuncts:
  type: array
  items:
  $ref: '#/components/schemas/Query'
  minItems: 1
  min:
  type: number
  format: float
  default: 1
  minimum: 0
  required:
  - disjuncts
  example:
  disjuncts:
  - term: "urgent"
  field: "title"
  - term: "critical"
  field: "content"
  min: 1

  Special Queries

  MatchAllQuery:
  allOf:
  - type: object
  properties:
  match_all:
  type: object
  example:
  match_all: {}
  boost: 1.0

  MatchNoneQuery:
  allOf:
  - type: object
  properties:
  match_none:
  type: object
  example:
  match_none: {}

  BoolFieldQuery:
  allOf:
  - type: object
  properties:
  bool:
  type: boolean
  required:
  - bool
  example:
  bool: true
  field: "published"

  DocIDQuery:
  allOf:
  - type: object
  properties:
  ids:
  type: array
  items:
  type: string
  minItems: 1
  required:
  - ids
  example:
  ids: ["doc1", "doc2", "doc3"]

  Geographic Queries

  GeoDistanceQuery:
  allOf:
  - type: object
  properties:
  location:
  oneOf:
  - type: array
  items:
  type: number
  format: float
  minItems: 2
  maxItems: 2
  - type: object
  properties:
  lon:
  type: number
  format: float
  lat:
  type: number
  format: float
  distance:
  type: string
  required:
  - location
  - distance
  example:
  location: [-122.4194, 37.7749]
  distance: "10km"
  field: "location"

  GeoBoundingBoxQuery:
  allOf:
  - type: object
  properties:
  top_left:
  oneOf:
  - type: array
  items:
  type: number
  format: float
  minItems: 2
  maxItems: 2
  - type: object
  properties:
  lon:
  type: number
  format: float
  lat:
  type: number
  format: float
  bottom_right:
  oneOf:
  - type: array
  items:
  type: number
  format: float
  minItems: 2
  maxItems: 2
  - type: object
  properties:
  lon:
  type: number
  format: float
  lat:
  type: number
  format: float
  required:
  - top_left
  - bottom_right
  example:
  top_left: [-122.5, 37.8]
  bottom_right: [-122.3, 37.7]
  field: "location"

  GeoBoundingPolygonQuery:
  allOf:
  - type: object
  properties:
  polygon_points:
  type: array
  items:
  oneOf:
  - type: array
  items:
  type: number
  format: float
  minItems: 2
  maxItems: 2
  - type: object
  properties:
  lon:
  type: number
  format: float
  lat:
  type: number
  format: float
  minItems: 3
  required:
  - polygon_points
  example:
  polygon_points:
  - [-122.5, 37.8]
  - [-122.3, 37.8]
  - [-122.3, 37.7]
  - [-122.5, 37.7]
  field: "location"

  GeoShapeQuery:
  allOf:
  - type: object
  properties:
  geometry:
  type: object
  properties:
  type:
  type: string
  enum: [Point, LineString, Polygon, MultiPoint, MultiLineString, MultiPolygon, GeometryCollection]
  coordinates:
  oneOf:
  - type: array
  - type: array
  items:
  type: array
  - type: array
  items:
  type: array
  items:
  type: array
  required:
  - type
  - coordinates
  required:
  - geometry
  example:
  geometry:
  type: "Polygon"
  coordinates:
  - [[-122.5, 37.8], [-122.3, 37.8], [-122.3, 37.7], [-122.5, 37.7], [-122.5, 37.8]]
  field: "location"

  Vector Queries

  KNNQuery:
  allOf:
  - type: object
  properties:
  field:
  type: string
  vector:
  type: array
  items:
  type: number
  format: float
  minItems: 1
  k:
  type: integer
  format: int64
  minimum: 1
  maximum: 10000
  params:
  type: object
  required:
  - field
  - vector
  - k
  example:
  field: "embedding"
  vector: [0.1, 0.2, 0.3, 0.4, 0.5]
  k: 10
  boost: 1.0

  Network Queries

  IPRangeQuery:
  allOf:
  - type: object
  properties:
  cidr:
  type: string
  required:
  - cidr
  example:
  cidr: "192.168.1.0/24"
  field: "ip_address"

  Phrase Query (internal use)

  PhraseQuery:
  allOf:
  - type: object
  properties:
  terms:
  type: array
  items:
  type: string
  required:
  - terms
  example:
  terms: ["exact", "phrase", "match"]
  field: "content"

  MultiPhraseQuery:
  allOf:
  - type: object
  properties:
  terms:
  type: array
  items:
  type: array
  items:
  type: string
  required:
  - terms
  example:
  terms:
  - ["exact", "precise"]
  - ["phrase"]
  - ["match", "search"]

examples:
# Basic Query Examples
simple_term:
summary: Simple Term Query
value:
term: "search"
field: "title"
boost: 2.0

simple_match:
  summary: Simple Match Query
  value:
    match: "full text search"
    field: "content"
    operator: "and"

fuzzy_search:
  summary: Fuzzy Search
  value:
    term: "search"
    field: "title"
    fuzziness: "auto"
    prefix_length: 2

# Boolean Query Examples
complex_boolean:
  summary: Complex Boolean Query
  value:
    must:
      - term: "important"
        field: "status"
      - match: "urgent"
        field: "title"
    should:
      - match: "critical"
        field: "content"
      - match: "priority"
        field: "tags"
    must_not:
      - term: "archived"
        field: "status"
    filter:
      - term: "published"
        field: "status"
    boost: 1.5

# Range Query Examples
price_range:
  summary: Price Range Query
  value:
    min: 10.0
    max: 100.0
    inclusive_min: true
    inclusive_max: false
    field: "price"

date_range:
  summary: Date Range Query
  value:
    start: "2023-01-01T00:00:00Z"
    end: "2023-12-31T23:59:59Z"
    inclusive_start: true
    inclusive_end: true
    field: "created_at"

# Geographic Query Examples
geo_distance:
  summary: Geographic Distance Query
  value:
    location: [-122.4194, 37.7749]
    distance: "10km"
    field: "location"

geo_bounding_box:
  summary: Geographic Bounding Box Query
  value:
    top_left: [-122.5, 37.8]
    bottom_right: [-122.3, 37.7]
    field: "location"

# Vector Query Examples
knn_search:
  summary: K-Nearest Neighbors Query
  value:
    field: "embedding"
    vector: [0.1, 0.2, 0.3, 0.4, 0.5]
    k: 10
    boost: 1.0

# Query String Examples
query_string:
  summary: Query String Query
  value:
    query: "title:search AND (content:full text OR content:search) AND -status:archived"

responses:
ValidationSuccess:
content:
application/json:
schema:
type: object
properties:
valid:
type: boolean
example: true
query_type:
type: string
example: "term"
message:
type: string
example: "Query is valid"

ValidationError:
  content:
    application/json:
      schema:
        type: object
        properties:
          valid:
            type: boolean
            example: false
          error:
            type: string
            example: "unknown query type"
          details:
            type: string
            example: "No recognized query fields found"

externalDocs:
url: https://github.com/blevesearch/bleve

Score Fusion (`docs/score_fusion.md`) Copy Section

Score Fusion for Hybrid Search

Bleve supports hybrid search that combines full-text search (FTS) with vector (kNN) search to leverage the strengths of both approaches:

• Full-text search excels at exact keyword matching, filtering, and structured queries
• Vector search captures semantic similarity and handles synonyms and paraphrasing naturally

With v2.5.4 onwards - when using hybrid search, you can choose different score fusion strategies to combine results from both search methods. This document describes the available fusion strategies and how to use them.

Fusion Strategies Copy Chapter

Additive Score Fusion (Default) Copy Section

By default, Bleve combines FTS and kNN scores using a simple weighted addition. See the Vector Search documentation for details on the default hybrid search behavior and examples.

While this approach works well with proper boost tuning, it can be sensitive to different score scales and distributions. The fusion strategies below (RRF and RSF) provide more robust alternatives that handle score normalization automatically.

Reciprocal Rank Fusion (RRF) Copy Section

Reciprocal Rank Fusion is a rank-based algorithm that combines results based on their position in each result list, rather than their raw scores. This makes it robust to different score scales and distributions.

Algorithm:

For each document appearing in FTS or kNN results, the RRF score is calculated as:

RRF\_score = w_{\text{fts}} \cdot \frac{1}{k + \text{rank}_{\text{fts}}} + \sum_{i=1}^{n} w_{\text{knn}_i} \cdot \frac{1}{k + \text{rank}_{\text{knn}_i}}

Where:

• $\text{rank}_{\text{fts}}$: 1-indexed rank of the document in the FTS result list (or 0 if not present)
• $\text{rank}_{\text{knn}_i}$: 1-indexed rank of the document in the i-th kNN result list (or 0 if not present)
• $k$: rank constant (default: 60) that dampens the impact of rank differences
• $w_{\text{fts}}$: weight from the FTS query boost value
• $w_{\text{knn}_i}$: weight from the i-th kNN query boost value
• $\sum_{i=1}^{n}$: summation over all kNN queries (you can add multiple kNN queries)

Advantages:

• Distribution-agnostic - no need for score normalization
• Works out of the box with minimal tuning
• Prioritizes documents appearing in both result lists
• Robust to outliers since only ranks matter

Disadvantages:

• Ignores score magnitude (loses some information)
• May be sensitive to imbalanced result list sizes

Usage:

// Create a hybrid search with RRF fusion
searchRequest := bleve.NewSearchRequest(bleve.NewMatchQuery("dark chocolate"))
searchRequest.Score = bleve.ScoreRRF  // Alternatively, set to "rrf"

// Add first kNN component
searchRequest.AddKNN(
    "embedding",                             // Vector field
    []float32{0.1, 0.2, 0.3, 0.4},          // Query vector
    30,                                      // k neighbors
    1.0,                                     // kNN weight (boost)
)

// Add second kNN component (optional - you can add multiple)
searchRequest.AddKNN(
    "image_embedding",                       // Different vector field
    []float32{0.5, 0.3, 0.1, 0.8},          // Query vector
    20,                                      // k neighbors
    0.5,                                     // kNN weight (boost)
)

// Optional: Configure RRF parameters
params := bleve.RequestParams{
    ScoreRankConstant: 60,                   // Rank constant (default: 60)
    ScoreWindowSize: 150                     // Window size (default: size)
}
searchRequest.AddParams(params)

searchResult, err := index.Search(searchRequest)

Relative Score Fusion (RSF) Copy Section

Relative Score Fusion is a score-based strategy that normalizes scores from both modalities into a common [0, 1] range using min-max normalization before combining them.

Algorithm:

1. Min-max normalize each result set independently:

   math Copy

   \text{normalized\_score} = \frac{\text{score} - \text{min\_score}}{\text{max\_score} - \text{min\_score}}

2. Combine normalized scores using weighted addition:

   math Copy

   RSF\_score = w_{\text{fts}} \cdot \text{normalized\_score\_fts} + \sum_{i=1}^{n} w_{\text{knn}_i} \cdot \text{normalized\_score\_knn}_i

Where:

• $w_{\text{fts}}$: weight from the FTS query boost value
• $w_{\text{knn}_i}$: weight from the i-th kNN query boost value
• $\sum_{i=1}^{n}$: summation over all kNN queries (you can add multiple kNN queries)

Advantages:

• Score-aware - retains relevance magnitude information
• Resolves incompatible score ranges
• Easy to understand

Disadvantages:

• Sensitive to outliers - a single extreme score can skew normalization
• Doesn't account for the shape or distribution of scores

Usage:

// Create a hybrid search with RSF fusion
searchRequest := bleve.NewSearchRequest(bleve.NewMatchQuery("machine learning"))
searchRequest.Score = bleve.ScoreRSF  // Or set to "rsf"

// Add first kNN component
searchRequest.AddKNN(
    "content_vector",                        // Vector field
    []float32{0.5, 0.3, 0.1, 0.8},          // Query vector
    20,                                      // k neighbors
    1.0,                                     // kNN weight (boost)
)

// Add second kNN component (optional - you can add multiple)
searchRequest.AddKNN(
    "title_vector",                          // Different vector field
    []float32{0.2, 0.7, 0.4, 0.1},          // Query vector
    15,                                      // k neighbors
    0.8,                                     // kNN weight (boost)
)

// Optional: Configure RSF parameters
params := bleve.RequestParams{
    ScoreWindowSize: 150                     // Window size (default: size)
}
searchRequest.AddParams(params)

searchResult, err := index.Search(searchRequest)

Parameters Copy Chapter

Score Copy Section

The Score field in your search request specifies which fusion strategy to use:

• ScoreRRF ("rrf"): Reciprocal Rank Fusion
• ScoreRSF ("rsf"): Relative Score Fusion
• Omitted or empty: Default additive fusion with scores returned

Params Copy Section

The Params object contains additional parameters for score fusion:

Score Window Size

ScoreWindowSize is the maximum number of results to consider from each result list for fusion.

• Default: Same as Size parameter
• Minimum: Must be ≥ Size and ≥ 1
• Purpose: Controls the tradeoff between relevance and performance

A larger window size increases the chance of finding relevant results but requires more computation. For pagination to work consistently, ensure:

From + Size <= ScoreWindowSize

Example:

{
  "score": "rrf",
  "params": {
    "score_window_size": 150
  },
  "size": 10,
  "from": 0
}

With window size set to 150, you can paginate through up to 150 results. If you try to access results beyond this (e.g., from=160), you'll get an empty result set.

Score Rank Constant

Only applicable for RRF

ScoreRankConstant controls how much the rank position affects the reciprocal rank score.

• Default: 60
• Range: Any positive integer
• Effect: Higher values dampen the impact of rank differences

Example:

{
  "score": "rrf",
  "params": {
    "score_rank_constant": 60
  }
}

Weighting Queries Copy Chapter

The boost value in your query components controls their relative importance in hybrid search:

// FTS query with boost 2.0
query := bleve.NewMatchQuery("search term")
query.SetBoost(2.0)

searchRequest := bleve.NewSearchRequest(query)

// kNN query with boost 1.0
searchRequest.AddKNN("vec", queryVector, 10, 1.0)

For RRF and RSF, weights determine the relative importance of each component's contribution, rather than scaling raw scores.

Example: If fts_boost = 2.0 and knn_boost = 1.0, the FTS contribution is twice as important as the kNN contribution in the final ranking in RRF or RSF.

Restrictions Copy Chapter

When using score fusion (Score set to "rrf" or "rsf"), certain features are not supported:

• SearchAfter/SearchBefore: Not compatible with score fusion. For pagination, use From and Size only.
• Sort: Only descending score sort (-_score) or default sorting is allowed
• Faceting: Only documents included in the FTS result list are considered. Documents that appear exclusively in the kNN result list are ignored during faceting.

Choosing a Fusion Strategy Copy Chapter

Scoring (`docs/scoring.md`) Copy Section

Scoring models for document hits

• Search is performed on a collection of fields using compound queries such as conjunction/disjunction/boolean, etc. However, the scoring itself is done independently for each field and then aggregated to get the final score for a document hit.
• Default scoring scheme for document hits involving text hits: tf-idf.
• Nearest-neighbor/vector hits scoring depends on chosen knn distance metric, highlighted here.
• Hybrid search scoring will combine tf-idf scores with knn distance numbers.
• v2.5.0 (and after) will come with support for bm25 scoring for exact searches.

BM25 Copy Chapter

When it comes to scoring a document hit for a specific field, BM25 scoring mechanism requires the following stats:

• fieldLen - The number of analyzed terms in the current document's field.
• avgFieldLen - The average number of analyzed terms in the field across all the documents.
• docTotal - The total number of documents in the index.
• docTerm - The total number of documents containing the query term within the index.

The scoring formula followed in BM25 is

\sum_{i}^n IDF(q_i) {{f(q_i,D) * (k1 + 1)}\over{f(q_i,D) + k1 * (1-b+b*{{fieldLen}\over{avgFieldLen}})}}

$IDF(q_i)$ here refers to Inverse Document Frequency, which talks about how rare (and hence rich in information) is a particular query term $q_i$ across all the documents in the index, which is calculated as

\ln(1 + {{docTotal - docTerm + 0.5}\over{docTerm + 0.5}})

Coming back to the BM25 scoring, $f(q_i,D)$ refers to the frequency of the query term in document $D$. The entire equation has certain multipliers

• $k1$ - helps in controlling the saturation of the score with respect to query term in a document. Basically if the query term's frequency is too high, the score value gets saturated and doesn't increase beyond a certain point.
• $b$ - controls the extent to which the $fieldLen$ normalizes the term's frequency.

How to enable and use BM25 Copy Section

Bleve v2.5.0 updated the indexMapping construct with the concept of scoringModel. This is a global (meaning applicable to all fields) setting which drives which scoring algorithm to apply while scoring the document hits. Supported scoring models are defined here

For instance, while defining the index mapping for the data model that's been decided by the user, following snippet can be referred to enable BM25

indexMapping := bleve.NewIndexMapping()
indexMapping.TypeField = "type"
indexMapping.DefaultAnalyzer = "en"
indexMapping.ScoringModel = "bm25"

During search time there's no explicit change involved, unless the user wants to perform a global scoring.

Global Scoring Copy Section

Let's say that the user has a dataset which is quite large (let's say 3 million) and to have good throughput, they create 3 shards (with the same index mapping) for the "index". Each of these shards can be bleve.Index type and while performing a search over the entire dataset, a bleve.IndexAlias can be created over which a search can be performed. This parallelizes things pretty well, both on the indexing path and the search path.

The concept of global scoring is applicable when the index is "sharded" (similar to above situation). This is because each index has data which is disjoint, and thereby while performing the scoring on document hits on each of them, the value of stats is not complete at a global level, since we're doing a search over the entire dataset using the bleve.IndexAlias. For eg: docTotal value while scoring the document hits would be 1 million which is incorrect at a global level.

So in order to keep the scoring roughly same across varying count of the number of shards involved, we provide a mechanism to enable "global scoring". In this type of search, an initial roundtrip is performed to gather and aggregate the stats necessary for the scoring mechanism and in the second phase, the actual search is performed. So naturally this comes at a cost of latency. As a reference here's how the user can go about with it

multiPartIndex := bleve.NewIndexAlias(shard1, shard2)
// set the alias with the same index mapping which both the shards use.
err = multiPartIndex.SetIndexMapping(indexMapping)
if err != nil {
    return err
}

ctx := context.Background()
ctx = context.WithValue(ctx, search.SearchTypeKey, search.GlobalScoring)

res, err := multiPartIndex.SearchInContext(ctx, searchRequest)

A note here is that, this would only matter if the relative order of the document hits vary quite a bit (vs single shard case). This would be possible when the shard count increases quite a bit, in low doc count situations or if there is a heavy skew in the data distribution amongst the shards for some reason. Ideally the shards are created when the data is quite large and each of them index the same amount of data - in which case the scores won't fluctuate much to affect the relative hit order and the user can choose to avoid the global scoring mechanism altogether.

TF-IDF Copy Chapter

TF-IDF is the default scoring mechanism involved (for backward compatibility reasons) and requires no change from the user at index or search time to avail it.

The scoring formula involved is

\sum_{i}^n f(q_i, D) * {{1}\over{\sqrt{fieldLen}}} * IDF(q_i)

where $IDF(q_i)$ is

1 + {{docTotal}\over{1 + docTerm}}

Note: TF-IDF formula doesn't accommodate logic for score saturation due to term frequency or fieldLen. So, it's recommended to use BM25 scoring by explicitly setting it in the index mapping.

Search Autocomplete (`docs/search_autocomplete.md`) Copy Section

Edge N-gram Autocomplete in Bleve

Search autocomplete is a feature which we see in search boxes when suggestions appear while we type.
So when we type jav, we see suggestions like: java javascript javascript programming etc.
This is helpful because it saves users time in finding what they are looking for.

2. How Does It Work? Copy Chapter

Autocomplete generally works in three steps:

1. Index Time: Breaking text into searchable pieces (tokens)
2. Query Time: Matching user input against indexed tokens
3. Results: Returning relevant suggestions quickly

But before we jump into the flow, let's understand different methods to achieve this and why edge n-grams are the most efficient approach.

3. Different Tokenization Methods Copy Chapter

There are several tokenization approaches, each with its own strengths and weaknesses:

3.1 Single Token Tokenizer Copy Section

// analysis/tokenizer/single/single.go
func (t *SingleTokenTokenizer) Tokenize(input []byte) analysis.TokenStream {
    return analysis.TokenStream{
        &analysis.Token{
            Term:     input,    // Here entire input as one token
            Position: 1,
            Start:    0,
            End:      len(input),
            Type:     analysis.AlphaNumeric,
        },
    }
}

How it works: Treats the entire input as a single token.

Example: "JavaScript Programming" → ["JavaScript Programming"]

Pros:

• Simple and fast
• Perfect for exact phrase matching
• Minimal index size

Cons:

• No autocomplete support (can't match partial text)
• Not flexible for search

Use case: Keyword fields, IDs, exact phrase matching

3.2 Whitespace Token Tokenizer Copy Section

// analysis/tokenizer/whitespace/whitespace.go  
func TokenizerConstructor(config map[string]interface{}, cache *registry.Cache) (analysis.Tokenizer, error) {
    return character.NewCharacterTokenizer(notSpace), nil
}

func notSpace(r rune) bool {
    return !unicode.IsSpace(r)  // Split on whitespace
}

How it works: Splits text on whitespace characters.

Example: "JavaScript Programming" → ["JavaScript", "Programming"]

Pros:

• Simple word-based tokenization
• Works well for basic prefix search

Cons:

• Only matches from word beginnings
• No support for partial word matching
• Limited autocomplete capabilities

Use case: Basic search, word-level indexing

3.3 N-gram Method Copy Section

// analysis/token/ngram/ngram.go
func (s *NgramFilter) Filter(input analysis.TokenStream) analysis.TokenStream {
    rv := make(analysis.TokenStream, 0, len(input))
    
    for _, token := range input {
        runeCount := utf8.RuneCount(token.Term)
        runes := bytes.Runes(token.Term)
        
        // ..generate all possible n-grams
    }
    return rv
}

How it works: Creates ALL possible substrings of specified lengths.

Example: "java" with 2-3 grams → ["ja", "av", "va", "jav", "ava"]

Pros:

• Supports partial matching anywhere in the text
• Coverage of all possible substrings

Cons:

• MASSIVE index size (exponential growth)
• Brings more noise as irrelevant matches (e.g., "av" matching "java")
• Poor performance for autocomplete
• High memory usage

Use case: Full-text substring search (not ideal for autocomplete)

3.4 Edge N-gram Method (preferred for autocomplete) Copy Section

// analysis/token/edgengram/edgengram.go
func (s *EdgeNgramFilter) Filter(input analysis.TokenStream) analysis.TokenStream {
    rv := make(analysis.TokenStream, 0, len(input))
    
    for _, token := range input {
        runeCount := utf8.RuneCount(token.Term)
        runes := bytes.Runes(token.Term)
        // ..builds tokens based on either end, specified in the input
    }
    return rv
}

How it works: Creates substrings only from the beginning (or end) of each word.

Example: "javascript" with front edge n-grams (1-5) → ["j", "ja", "jav", "java", "javas"]

Pros:

• Perfect for autocomplete (matches prefixes naturally)
• Efficient index size (linear growth vs exponential)
• Fast queries (direct term matching, no complex processing)
• Intuitive results (matches what users expect)
• Highly scalable for large datasets

Cons:

• Only supports prefix matching (but that's preferred for autocomplete!)
• Slightly larger index than basic tokenization

Use case: Search autocomplete, prefix-based search

4. Why Edge N-grams Are Most Efficient for Autocomplete Copy Chapter

Let's see what happens when a user types "java" with edge_ngram tokenizer:

Index contains: ["j", "ja", "jav", "java", "javas", "javasc", "javascr", ...]
User types: "java"  
Query: ExactTermQuery("java")
Process: Direct hash table lookup for term "java"
Result: Instant match, then retrieve documents containing this term

Key advantages of Edge N-gram approach:

1. O(1) lookup: Direct term matching
2. No query-time processing: Terms are pre-computed at index time
3. Better caching: Exact term queries cache better than prefix queries
4. Consistent performance: Query time doesn't increase with index size

5. Step-by-Step Implementation: Building Search Autocomplete Copy Chapter

Now let's see how to implement this in practice using our exact configuration. We'll build a complete autocomplete system step by step.

Step 1: Create Custom Edge N-gram Token Filter Copy Section

First, we need to create a custom token filter for edge n-grams. Here's how it looks in our configuration:

// Create a new index mapping
indexMapping := mapping.NewIndexMapping()

// 1. Define the edgeGram token filter
edgeGramFilter := map[string]interface{}{
    "type": edgengram.Name,
    "min":  2.0,
    "max":  4.0,
    "back": false,
}

// Register the token filter
if err := indexMapping.AddCustomTokenFilter("Engram", edgeGramFilter); err != nil {
    log.Fatal(err)
}

What each setting does:

• "type": "edge_ngram" - Tells Bleve to use the edge n-gram filter
• "min": 2 - Start creating tokens from 2 characters ("ja", "sc", etc.)
• "max": 4 - Stop at 4 characters ("java", "scri", etc.)
• "back": false - Create tokens from the front (beginning) of words

Step 2: Create Custom Analyzer Copy Section

Next, we create an analyzer that uses our custom token filter along with other helpful filters:

// 2. Define a custom analyzer that uses it
customAnalyzer := map[string]interface{}{
    "type":      custom.Name,
    "tokenizer": "unicode",
    "char_filters": []string{
         zerowidthnonjoiner.Name,
    },
    "token_filters": []string{
        "Engram", // our custom edge_ngram filter
        "to_lower",
        "stop_en",
    },
}

if err := indexMapping.AddCustomAnalyzer("edgeGramAnalyzer", customAnalyzer); err != nil {
    log.Fatal(err)
}

The pipeline works like this:

1. Input Text: "Schaumbergfest Event"

2. Tokenizer (unicode): Splits into words

   text Copy

   ["Schaumbergfest", "Event"]

3. Character Filter (zero_width_spaces): Removes invisible characters

   text Copy

   ["Schaumbergfest", "Event"] (cleaned)

4. Token Filter 1 (Engram): Creates edge n-grams (2-4 chars)

   text Copy

   ["Sc", "Sch", "Scha", "Ev", "Eve", "Even"]

5. Token Filter 2 (to_lower): Makes everything lowercase

   text Copy

   ["sc", "sch", "scha", "ev", "eve", "even"]

6. Token Filter 3 (stop_en): Removes common words (none in this case)

   text Copy

   ["sc", "sch", "scha", "ev", "eve", "even"] (final tokens)

Step 3: Configure Field Mapping Copy Section

Now we tell Bleve which fields to apply our autocomplete analyzer to:

// 3. Assign analyzer to a field mapping
    fieldMapping := mapping.NewTextFieldMapping()
    fieldMapping.Analyzer = "edgeGramAnalyzer"
    
    indexMapping.DefaultMapping.AddFieldMappingsAt("title", fieldMapping)
    
    indexPath := "example.bleve"
    index, err := bleve.New(indexPath, indexMapping)
    if err != nil {
        log.Fatal(err)
    }

Step 4: How It Works in Real Search Copy Section

When someone searches for "sc", here's what happens:

Index contains these tokens:

"sc" → [document1: "Schaumbergfest", document2: "Script", ...]
"sch" → [document1: "Schaumbergfest", ...]  
"scha" → [document1: "Schaumbergfest", ...]

User types "sc":

1. Query: name:sc
2. Bleve looks up exact term "sc" in the index
3. Finds document with "Schaumbergfest" and "Script"
4. Returns suggestion instantly

type Document struct {
  ID    string `json:"id"`
  Title string `json:"title"`
}
// 4. Index Documents
documents := []Document{
  {
    ID:    "doc1",
    Title: "Schaumbergfest",
  },
  {
    ID:    "doc2",
    Title: "Script",
  },
}

batch := index.NewBatch()
for _, doc := range documents {
  batch.Index(doc.ID, doc)
}
if err := index.Batch(batch); err != nil {
  log.Fatal(err)
}

// 5. Search the created index
query := bleve.NewMatchQuery("sc")
query.SetField("title")
searchRequest := bleve.NewSearchRequest(query)
searchRequest.Explain = true
searchRequest.Fields = []string{"title"}
searchResult, err := index.Search(searchRequest)
if err != nil {
  log.Fatal(err)
}
fmt.Println(searchResult)

Output:

$ go run main.go

2 matches, showing 1 through 2, took 189.041µs
    1. doc2 (0.343255)
        title
                Script
    2. doc1 (0.343255)
        title
                Schaumbergfest

Note: To run code, enclose code starting from Step 1 in func main.

Sort Facet (`docs/sort_facet.md`) Copy Section

Purpose of Docvalues

Background

What are docValues? In the index mapping, there is an option to enable or disable docValues for a specific field mapping. However, what does it actually mean to activate or deactivate docValues, and how does it impact the end user? This document aims to address these questions.

"default_mapping": {
	"dynamic": true,
	"enabled": true,
	"properties": {
		"loremIpsum": {
		"enabled": true,
		"dynamic": false,
		"fields": [
			{
				"name": "loremIpsum",
				"type": "text",
				"store": false,
				"index": true,
				"include_term_vectors": false,
				"include_in_all": false,
				"docvalues": true
			}
		]
	}
}

Enabling docValues will always result in an increase in the size of your Bleve index, leading to a corresponding increase in disk usage. But what advantages can you expect in return? This document also quantitatively assesses this trade-off with a test case.

In a more general sense, we recommend enabling docValues on a field mapping if you anticipate queries that involve sorting and/or facet operations on that field. It's important to note, though, that sorting and faceting will work irrespective of whether docValues are enabled or not. This may lead you to wonder if there's any real benefit to enabling docValues since you're allocating extra disk space without an apparent return. The real advantage, however, becomes evident in enhanced query response times and reduced memory consumption during active usage. By accepting a minor increase in the disk space used by your Full-Text Search (FTS) index, you can anticipate better performance in handling search requests that involve sorting and faceting.

Usage

The initial use of docValues comes into play when sorting is involved. In the search request JSON, there is a field named "sort." This optional "sort" field can have a slice of JSON objects as its value. Each JSON object must belong to one of the following types:

<li>SortDocID</li>
	<li>SortScore (which is the default if none is specified)</li>
	<li>SortGeoDistance</li> 
	<li>SortField</li>

DocValues are relevant only when any of the JSON objects in the "sort" field are of type SortGeoDistance or SortField. This means that if you expect queries on a field F, where the queries either do not specify a value for the "sort" field or provide a JSON object of type SortDocID or SortScore, enabling docValues will not improve sorting operations, and as a result, query latency will remain unchanged. It's worth noting that the default sorting object, SortScore, does not require docValues to be enabled for any of the field mappings. Therefore, a search request without a sorting operation will not utilize docValues at all.

"*"

],
"highlight": {},
"query": {
"match": "lorem ipsum",
"field":"dolor"
},
"size": 10,
"from": 0
} | {
"explain": true,
"fields": [
""
],
"highlight": {},
"query": {
"match": "lorem ipsum",
"field":"sit_amet"
},
"sort":[
{
"by":"id",
"desc":true
}
],
"size": 10,
"from": 0
} | {
"explain": true,
"fields": [
""
],
"highlight": {},
"query": {
"match": "lorem ipsum",
"field":"sit_amet"
},
"sort":[
{
"by":"score",
}
],
"size": 10,
"from": 0
} | {
"explain": true,
"fields": [
""
],
"highlight": {},
"query": {
"match": "lorem ipsum",
"field":"sit_amet"
},
"sort":[
{
"by":"field",
"field":"dolor",
"type":"auto",
"mode":"min",
"missing":"last"
}
],
"size": 10,
"from": 0
} | {
"explain": true,
"fields": [
""
],
"highlight": {},
"query": {
"match": "lorem ipsum",
"field": "dolor"
},
"sort": [
{
"by": "geo_distance",
"field": "sit_amet",
"location": [
123.223,
34.33
],
"unit": "km"
}
],
"size": 10,
"from": 0
} |
| No DocValues used | No DocValues used | No DocValues used | DocValues used for field "dolor". Field Mapping for "dolor" may enable docValues. | DocValues used for field "dolor". Field Mapping for "dolor" may enable docValues. |

Now, let's consider faceting. The search request object also includes another field called "facets," where you can specify a collection of facet requests, with each request being associated with a unique name. Each of these facet requests can fall into one of three types:

<li>Date range</li> 
<li>Numeric range</li>
<li>Term facet</li>

Enabling docValues for the fields associated with such facet requests might provide benefits in this context.

"*"

],
"highlight": {},
"query": {
"match": "lorem ipsum",
"field": "dolor"
},
"size": 10,
"from": 0
} | {
"explain": true,
"fields": [
""
],
"highlight": {},
"query": {
"match": "lorem ipsum",
"field": "sit_amet"
},
"facet": {
"facetA": {
"size": 1,
"field": "dolor",
"date_ranges": [
{
"name": "lorem",
"start": "20/August/2001",
"end": "22/August/2002",
"datetime_parser": "custDT"
}
]
}
},
"size": 10,
"from": 0
} | {
"explain": true,
"fields": [
""
],
"highlight": {},
"query": {
"match": "lorem ipsum",
"field": "sit_amet"
},
"facet": {
"facetA": {
"size": 1,
"field": "dolor",
"numeric_ranges":[
{
"name":"lorem",
"min":22,
"max":34
}
]
}
},
"size": 10,
"from": 0
} | {
"explain": true,
"fields": [
"*"
],
"highlight": {},
"query": {
"match": "lorem ipsum",
"field": "sit_amet"
},
"facet": {
"facetA": {
"size": 1,
"field": "dolor"
}
},
"size": 10,
"from": 0
} |
| No DocValues used | DocValues used for field "dolor". Field Mapping for "dolor" may enable docValues. | | |

In summary, when a search request is received by the Bleve index, it extracts all the fields from the sort objects and facet objects. To potentially benefit from docValues, you should consider enabling docValues for the fields mentioned in SortField and SortGeoDistance sort objects, as well as the fields associated with all the facet objects. By doing so, you can optimize sorting and faceting operations in your search queries.

"*"

],
"highlight": {},
"query": {
"match": "lorem ipsum",
"field": "sit_amet"
},
"facet": {
"facetA": {
"size": 1,
"field": "dolor",
"date_ranges": [
{
"name": "lorem",
"start": "20/August/2001",
"end": "22/August/2002",
"datetime_parser": "custDT"
}
]
}
},
"sort":[
{
"by":"field",
"field":"lorem",
"type":"auto",
"mode":"min",
"missing":"last"
}
],
"size": 10,
"from": 0
} | {
"explain": true,
"fields": [
"*"
],
"highlight": {},
"query": {
"match": "lorem ipsum",
"field": "sit_amet"
},
"facet": {
"facetA": {
"size": 1,
"field": "dolor",
"numeric_ranges":[
{
"name":"lorem",
"min":22,
"max":34
}
]
}
},
"sort": [
{
"by": "geo_distance",
"field": "ipsum",
"location": [
123.223,
34.33
],
"unit": "km"
}
],
"size": 10,
"from": 0
} |
| DocValues used for fields "dolor" and "lorem". Field Mapping for "dolor" and "lorem" may enable docValues. | DocValues used for fields "dolor" and "ipsum". Field Mapping for "dolor" and "ipsum" may enable docValues. |

Empirical Analysis

To evaluate our hypothesis, I've set up a sample dataset on my personal computer and I've created two Bleve indexes: one with docvalues enabled for three fields (dummyDate, dummyNumber, and dummyTerm), and another where I've disabled docValues for the same three fields. These field mappings were incorporated into the Default Mapping. It's important to mention that for both indexes, DocValues for dynamic fields were enabled, as the default mapping is dynamic.

The values for dummyDate and dummyNumber were configured to increase monotonically, with dummyDate representing a date value and dummyNumber representing a numeric value. This setup was intentional to ensure that facet aggregation would consistently result in cache hits and misses, providing a useful testing scenario.

"dynamic": true,
"enabled": true,
"properties": {
 "dummyNumber": {
  "enabled": true,
  "dynamic": false,
  "fields": [
   {
    "name": "dummyNumber",
    "type": "text",
    "store": false,
    "index": true,
    "include_term_vectors": false,
    "include_in_all": false,
    "docvalues": true
   }
  ]
 },
 "dummyTerm": {
  "enabled": true,
  "dynamic": false,
  "fields": [
   {
    "name": "dummyTerm",
    "type": "text",
    "store": false,
    "index": true,
    "include_term_vectors": false,
    "include_in_all": false,
    "docvalues": true
   }
  ]
 },
 "dummyDate": {
  "enabled": true,
  "dynamic": false,
  "fields": [
   {
    "name": "dummyDate",
    "type": "text",
    "store": false,
    "index": true,
    "include_term_vectors": false,
    "include_in_all": false,
    "docvalues": true
   }
  ]
 }
}

} | "default_mapping": {
"dynamic": true,
"enabled": true,
"properties": {
"dummyNumber": {
"enabled": true,
"dynamic": false,
"fields": [
{
"name": "dummyNumber",
"type": "text",
"store": false,
"index": true,
"include_term_vectors": false,
"include_in_all": false,
"docvalues": false
}
]
},
"dummyTerm": {
"enabled": true,
"dynamic": false,
"fields": [
{
"name": "dummyTerm",
"type": "text",
"store": false,
"index": true,
"include_term_vectors": false,
"include_in_all": false,
"docvalues": false
}
]
},
"dummyDate": {
"enabled": true,
"dynamic": false,
"fields": [
{
"name": "dummyDate",
"type": "text",
"store": false,
"index": true,
"include_term_vectors": false,
"include_in_all": false,
"docvalues": false
}
]
}
}
} |
| Docvalues enabled across all three field mappings | Docvalues disabled across all three field mappings |

Document Format used for the test scenario:

Now I ran the following set of search requests across both the indexes, while increasing the number of documents indexed from 2000 to 4000.

Even at this small scale, with a small document size and a very limited number of indexed documents, we still observe a noticeable tradeoff. With just a slight increase in the index size (an average of 7KB), we obtain a 20ms reduction in the total execution time, on average, for only 1000 queries.

Technical Information

When a search request involves facet or sorting operations on a field F, these operations occur after the main search query is executed. For instance, if the main query yields a result of 200 documents, the sorting and faceting processes will be applied to these 200 documents. However, the main query result only provides a set of document IDs, not the actual document contents.

Here's where docValues become essential. If the field mapping for F is docValue enabled, the system can directly access the values for the field from the stored docValue part in the index file. This means that for each document ID returned in the search result, the field values are readily available.

However, if docValues are not enabled for field F, the system must take a different approach. It needs to "fetch the document" from the index file, read the value for field F, and cache this field-document pair in memory for further processing. The issue becomes apparent in the latter scenario. By not enabling docValues for field F, you essentially retrieve all the documents in the search result (at the worst case), which can be a substantial amount of data. Moreover, you have to cache this information in memory, leading to increased memory usage. As a result, query latency significantly suffers because you're essentially fetching and processing all documents, which can be both time-consuming and resource-intensive. Enabling docValues for the relevant fields is, therefore, a crucial optimization to enhance query performance and reduce memory overhead in such situations.

Synonyms (`docs/synonyms.md`) Copy Section

Synonym search

• v2.5.0 (and after) will come with support for synonym definition indexing and search.
• We've achieved this by embedding synonym indexes within our bleve (scorch) indexes.
• Usage of zap file format: v16. Here we co-locate text, vector, and synonym indexes as neighbors within segments, continuing to conform to the segmented architecture of scorch.

Supported Copy Chapter

• Indexing Synonym Definitions allows specifying equivalent terms that will be used to construct the synonym index. There are currently two types of Synonym Definitions supported:

  1. Equivalent Mapping:

     In this type, all terms in the synonyms list are considered equal and can replace one another. Any of these terms can match a query or document containing any other term in the group, ensuring full synonym coverage.

     json Copy

     {
         "synonyms": [
             "tranquil",
             "peaceful",
             "calm",
             "relaxed",
             "unruffled"
         ]
     }

  2. Explicit Mapping:

     In this mapping, only the terms in the input list ("blazing") will have the terms in synonyms as their synonyms. The input terms are not equivalent to each other, and the synonym relationship is explicitly directional, applying only from the input to the synonyms.

     json Copy

     {
         "input": [
             "blazing"
         ],
         "synonyms": [
             "intense",
             "radiant",
             "burning",
             "fiery",
             "glowing"
         ]
     }

• The addition of Synonym Sources in the index mapping enables associating a set of synonym definitions (called a synonym collection) with a specific analyzer. This allows for preprocessing of terms in both the "input" and "synonyms" lists before the synonym index is created. By using an analyzer, you can normalize or transform terms (e.g., case folding, stemming) to improve synonym matching.

  json Copy

  {
      "analysis": {
          "synonym_sources": {
              "english": {
                  "collection": "en_thesaurus",
                  "analyzer": "en"
              },
              "german": {
                  "collection": "de_thesaurus",
                  "analyzer": "de"
              }
          }
      }
   }

  There are two synonym sources named "english" and "german," each associated with its respective synonym collection and analyzer. In any text field mapping, a synonym source can be specified to enable synonym expansion when the field is queried. The analyzer of the synonym source must match the analyzer of the field mapping to which it is applied.

• Any text-based Bleve query (e.g., match, phrase, term, fuzzy, etc.) will use the synonym source (if available) for the queried field to expand the search terms using the thesaurus created from user-defined synonym definitions. The behavior for specific query types is as follows:

  1. Queries with fuzziness parameter: For queries like match, phrase, and match-phrase that support the fuzziness parameter, the queried terms are fuzzily matched with the thesaurus's LHS terms to generate candidate terms. These terms are then combined with the results of fuzzy matching against the field dictionary, which contains the terms present in the queried field.

  2. Wildcard, Regexp, and Prefix queries: These queries follow a similar approach. First, the thesaurus is used to expand terms (e.g., LHS terms that match the prefix or regex). The resulting terms are then combined with candidate terms from dictionary expansion.

Indexing Copy Chapter

Below is an example of using the Bleve API to define synonym sources, index synonym definitions, and associate them with a text field mapping:

/* Detailed source-code truncated for AI context efficiency. */

Querying Copy Chapter

// Query the index created above.
// Create a match query for the term "persistent".
query := bleve.NewMatchQuery("persistent")

// Specify the field to search within, in this case, the "text" field.
query.SetField("text")

// Create a search request with the query and enable explanation to understand how results are scored.
searchRequest := bleve.NewSearchRequest(query)
searchRequest.Explain = true

// Execute the search on the index.
searchResult, err := index.Search(searchRequest)
if err != nil {
    // Handle any errors that occur during the search.
    panic(err)
}

// The search result will contain one match: "doc1". This document includes the term "hardworking",
// which is a synonym for the queried term "persistent". The synonym relationship is based on 
// the user-defined thesaurus associated with the index.
// Print the search results, which will include the explanation for the match.
fmt.Println(searchResult)

Vectors (`docs/vectors.md`) Copy Section

Approximate Nearest Neighbor Search (over vectors)

• v2.4.0 (and after) will come with support for vector indexing and search.
• We've achieved this by embedding FAISS indexes within our bleve (scorch) indexes.
• Introduction of new zap file formats: v16, v17 .. to co-locate text and vector indexes as neighbors within segments, continuing to conform to the segmented architecture of scorch.

Pre-requisite(s) Copy Chapter

• Induction of FAISS into our ecosystem, which is a fork of the original facebookresearch/faiss

• FAISS is a C++ library that needs to be compiled and its shared libraries need to be situated at an accessible path for your application.

• A vectors GO TAG needs to be set for bleve to access all the supporting code. This TAG must be set only after the FAISS shared library is made available. Failure to do either will inhibit you from using this feature.

• Please follow these instructions below for any assistance in the area.

• Releases of blevesearch/bleve work with select checkpoints of blevesearch/faiss owing to API changes and improvements (tracking over the bleve branch):

  bleve version(s)	blevesearch/faiss version
  v2.4.0	blevesearch/faiss@7b119f4 (modified v1.7.4)
  v2.4.1, v2.4.2	blevesearch/faiss@d9db66a (modified v1.7.4)
  v2.4.3, v2.4.4	blevesearch/faiss@b747c55 (modified v1.8.0)
  v2.5.0, v2.5.1	blevesearch/faiss@352484e (modified v1.10.0)
  v2.5.2, v2.5.3, v2.5.4	blevesearch/faiss@b3d4e00 (modified v1.11.0)
  v2.5.5, v2.5.6, v2.5.7	blevesearch/faiss@8a59a0c (modified v1.12.0)
  v2.6.0	blevesearch/faiss@fff814d (modified v1.13.2)

Features Copy Chapter

• The vector field type is an array that is to hold float32 values only.
• The vector_base64 field type to support base64 encoded strings using little endian byte ordering (v2.4.1+)
• Supported similarity metrics are: ["cosine" (v2.4.3+), "dot_product", "l2_norm"].
  • cosine paths will additionally normalize vectors before indexing and search.
• Supported dimensionality is between 1 and 2048 (v2.4.0), and up to 4096 (v2.4.1+).
• Supported vector index optimizations:
  • recall, latency (v2.4.0+)
    • Combination of Flat and IVF indexes with SQ8 quantization.
  • memory_efficient (v2.4.1+)
    • Combination of Flat and IVF indexes with SQ4 quantization.
  • bivf-flat, bivf-sq8 (v2.6.0+)
    • Combination of BFlat and BIVF indexes with Binary quantization.
    • Uses an additional index which is either a Flat index (bivf-flat) or an SQ8 index (bivf-sq8) for re-ranking.
  • ivf,rabitq (v2.6.0+)
    • Combination of Flat and IVF indexes with RaBitQ quantization.
    • Works with Fast Merge - where a centroid index needs to be built/trained on a sample dataset prior to building the index for segments that deploy IVF indexes to reflect.

• Vectors from documents that do not conform to the index mapping dimensionality are simply discarded at index time.
• The dimensionality of the query vector must match the dimensionality of the indexed vectors to obtain any results.
• Pure kNN searches can be performed, but the query attribute within the search request must be set -- to {"match_none": {}} in this case. The query attribute is made optional when knn is available with v2.4.1+.
• Hybrid searches are supported, where results from query are unioned with results from knn. The FTS scores from exact searches are simply summed with the similarity distances to determine the aggregate scores.

aggregate_score = (query_boost * query_hit_score) + (knn_boost * knn_hit_distance)

• Advanced score fusion strategies (v2.5.4+) are available if requested for - see score fusion.
• Multi kNN searches are supported - the knn object within the search request accepts an array of requests. These sub objects are unioned by default but this behavior can be overridden by setting knn_operator to "and".
• Previously supported pagination settings will work as they were, with size/limit being applied over the top-K hits combined with any exact search hits.
• Pre-filtered vector and hybrid search (v2.4.3+): Apply any Bleve filter query first to narrow down candidates before running kNN search, making vector and hybrid searches faster and more relevant.
• Fields containing multiple vectors (v2.5.7+):
  • A single document may contain multiple vectors within the same field, in the form of either:
    • an array of vectors (multi-vector field)
    • an array of objects each containing a vector (nested-vector field)
  • For single-kNN queries, each document is scored using its single best-matching vector.
  • For multi-kNN queries, the system selects the best-matching vector for each query vector within the document.

• GPU-Accelerated vector search (v2.6.0+):
  • Requires FAISS built with -DFAISS_ENABLE_GPU=ON CMake option (needs NVIDIA CUDA toolkit).
  • Requires the gpu go tag in addition to the vectors tag when building bleve.
  • GPU acceleration is enabled per vector field via the field mapping's GPU option.
  • Bleve will use any available GPUs to offload training, indexing, and kNN searches for GPU-enabled vector fields.
  • GPUs will be used for vector fields which are optimized for recall, latency, and memory_efficient.
  • Multi-GPU support: when multiple GPUs are available, a load balancer distributes vector search workloads across devices.
  • See GPU setup instructions below for building FAISS with GPU support.

Indexing Copy Chapter

/* Detailed source-code truncated for AI context efficiency. */

Querying Copy Chapter

// ------------------------------------
// Single-vector field search (v2.4.0+)
// ------------------------------------
searchRequest := bleve.NewSearchRequest(bleve.NewMatchNoneQuery())
searchRequest.AddKNN(
    "vec",                                   // Vector field
    []float32{0, 1, 1, 4, 4, 5, 7, 6, 8, 9}, // Query vector
    5,                                       // top-k
    1,                                       // boost
)
searchResult, err := index.Search(searchRequest)
if err != nil {
    panic(err)
}
// Scores are 1 / squared L2 distance, e.g., score = 0.25 for squared distance of 4
fmt.Printf("Single-vector field kNN result:\n%s\n", searchResult)

// -----------------------------------
// Multi-vector field search (v2.5.7+)
// -----------------------------------
searchRequest = bleve.NewSearchRequest(bleve.NewMatchNoneQuery())
searchRequest.AddKNN(
    "embeddings",
    []float32{0, 1, 1, 4, 4, 5, 7, 6, 8, 9},
    5,
    1.0,
)
searchResult, err = index.Search(searchRequest)
if err != nil {
    panic(err)
}
// Scores are based on the **best-matching vector** from the multi-vector field.
// Example: distances to doc vectors {10..19} and {20..29} → pick the closer one (smaller squared L2),
// then score = 1 / squared L2 distance.
fmt.Printf("Multi-vector field kNN result:\n%s\n", searchResult)

// ------------------------------------
// Nested-vector field search (v2.5.7+)
// ------------------------------------
searchRequest = bleve.NewSearchRequest(bleve.NewMatchNoneQuery())
searchRequest.AddKNN(
    "sections.vec",
    []float32{0, 1, 1, 4, 4, 5, 7, 6, 8, 9},
    5,
    1.0,
)
searchResult, err = index.Search(searchRequest)
if err != nil {
    panic(err)
}
// Scores are based on the **best-matching vector** from the nested-vector field.
// Example: distances to doc vectors {30..39} and {40..49} → pick the closer one (smaller squared L2),
// then score = 1 / squared L2 distance.
fmt.Printf("Nested-vector field kNN result:\n%s\n", searchResult)

// -----------------------------------------------------
// Multi kNN queries on multi-vector documents (v2.5.7+)
// -----------------------------------------------------
searchRequest = bleve.NewSearchRequest(bleve.NewMatchNoneQuery())
searchRequest.AddKNN(
    "embeddings",
    []float32{0, 1, 1, 4, 4, 5, 7, 6, 8, 9},
    5,
    1.0,
)
searchRequest.AddKNN(
    "embeddings",
    []float32{1, 2, 2, 5, 5, 6, 8, 7, 9, 10},
    8,
    1.0,
)
searchResult, err = index.Search(searchRequest)
if err != nil {
    panic(err)
}
// Document score explanation:
// - For each query vector, Bleve selects the **closest vector** in the multi-vector field.
// - Scores from multiple queries are then **normalized and summed** to get the final document score.
// For example, if the closest vector to the first query has squared L2 distance 4 (score 0.25)
// and the closest vector to the second query has squared L2 distance 1 (score 1.0),
// and both queries use equal boost values of 1.0, the normalization factor is 1/√2 (where 2 is the number of kNN queries).
// Then the total document score = 1/√2 * 0.25 + 1/√2 * 1.0 = 0.1768 + 0.7071 = 0.8839.
// Note: If the boost values differ, or if more queries are used, the normalization factor and score calculation will change accordingly.
fmt.Printf("Multi kNN queries result:\n%s\n", searchResult)

// --------------------------------------
// Hybrid search: text + vector (v2.4.0+)
// --------------------------------------
searchRequest = bleve.NewSearchRequest(bleve.NewMatchQuery("united states"))
searchRequest.AddKNN(
    "vec",
    []float32{0, 1, 1, 4, 4, 5, 7, 6, 8, 9},
    5,
    1,
)
searchResult, err = index.Search(searchRequest)
if err != nil {
    panic(err)
}
// Score = sum of text relevance score + kNN vector score
// Example: text score 0.5 + vector score 0.25 = total score 0.75
fmt.Printf("Hybrid search result:\n%s\n", searchResult)

Querying with filters (v2.4.3+) Copy Chapter

// Pre-filtered vector/hybrid search: filter query narrows candidates before KNN search
searchRequest = bleve.NewSearchRequest(bleve.NewMatchNoneQuery()) // replace with any Bleve query for Pre-filtered Hybrid Search
filterQuery := bleve.NewTermQuery("hello")                        // Filter query to narrow candidates
searchRequest.AddKNNWithFilter(
    "vec",                                   // Vector field name
    []float32{0, 1, 1, 4, 4, 5, 7, 6, 8, 9}, // Query vector (must match indexed vector dims)
    5,                                       // Number of nearest neighbors to return (k)
    1,                                       // Boost factor for KNN score
    filterQuery,                             // Filter query applied before KNN search
)
searchResult, err = index.Search(searchRequest)
if err != nil {
    panic(err)
}
// Scores are computed only among documents matching the filter query
// Example: if only one document matches the filter and has squared L2 distance 4 to the query vector,
// its score will be 0.25 (1 / 4) and returned as the top result.
fmt.Printf("Pre-filtered kNN search result:\n%s\n", searchResult)

Setup Instructions Copy Chapter

• Using cmake is a recommended approach by FAISS authors.
• More details here - faiss/INSTALL.

Linux Copy Section

Also documented here - go-faiss/README.

git clone https://github.com/blevesearch/faiss.git
cd faiss
cmake -B build -DFAISS_ENABLE_GPU=OFF -DFAISS_ENABLE_C_API=ON -DBUILD_SHARED_LIBS=ON .
make -C build
sudo make -C build install

Building will produce the dynamic library faiss_c. You will need to install it in a place where your system will find it (e.g. /usr/lib). You can do this with:

sudo cp build/c_api/libfaiss_c.so /usr/local/lib

OSX Copy Section

While you shouldn't need to do anything different for macOS x86_64, with aarch64 - some instructions need adjusting (see facebookresearch/faiss#2111) ..

LDFLAGS="-L/opt/homebrew/opt/llvm/lib" CPPFLAGS="-I/opt/homebrew/opt/llvm/include" CXX=/opt/homebrew/opt/llvm/bin/clang++ CC=/opt/homebrew/opt/llvm/bin/clang cmake -B build -DFAISS_ENABLE_GPU=OFF -DFAISS_ENABLE_C_API=ON -DBUILD_SHARED_LIBS=ON -DFAISS_ENABLE_PYTHON=OFF .
make -C build
sudo make -C build install
sudo cp build/c_api/libfaiss_c.dylib /usr/local/lib

Sanity check Copy Section

Once the supporting library is built and made available, a sanity run is recommended to make sure all unit tests and especially those accessing the vectors' code pass. Here's how ..

go test -ldflags "-r /usr/local/lib" ./... -tags=vectors

GPU Setup Instructions (v2.6.0+) Copy Chapter

GPU-accelerated vector search requires FAISS to be compiled with CUDA support and the gpu go tag to be set when building bleve.

Pre-requisites Copy Section

• An NVIDIA GPU with CUDA support (Pascal architecture / CC 6.0+ recommended).
• NVIDIA CUDA Toolkit installed and available on your system.
• The FAISS C API shared library built with GPU support.

Linux (GPU) Copy Section

git clone https://github.com/blevesearch/faiss.git
cd faiss
cmake -B build -DFAISS_ENABLE_GPU=ON -DFAISS_ENABLE_C_API=ON -DBUILD_SHARED_LIBS=ON .
make -C build
sudo make -C build install
sudo cp build/c_api/libfaiss_c.so /usr/local/lib

Sanity check (GPU) Copy Section

Run the full test suite with both the vectors and gpu build tags:

go test -ldflags "-r /usr/local/lib" ./... -tags=vectors,gpu

--- METRICS ---

• Files Extracted: 21
• Estimated Token Budget: ~38126 tokens
• Canonical Reference: https://codewiki.google/github.com/blevesearch/bleve