1. Project Overview & Quickstart (yichuan-w/LEANN)
llms.txt — LEANN MCP and Agent Integration
product: LEANN
homepage: https://github.com/yichuan-w/LEANN
contact: https://github.com/yichuan-w/LEANN/issues
Installation
install: uv tool install leann-core --with leann
MCP Server Entry Point
mcp.server: leann_mcp
mcp.protocol_version: 2024-11-05
Tools
mcp.tools: leann_list, leann_search
mcp.tool.leann_list.description: List available LEANN indexes
mcp.tool.leann_list.input: {}
mcp.tool.leann_search.description: Semantic search across a named LEANN index
mcp.tool.leann_search.input.index_name: string, required
mcp.tool.leann_search.input.query: string, required
mcp.tool.leann_search.input.top_k: integer, optional, default=5, min=1, max=20
mcp.tool.leann_search.input.complexity: integer, optional, default=32, min=16, max=128
Notes
note: Build indexes with leann build <name> --docs <files...> before searching.
example.add: claude mcp add --scope user leann-server -- leann_mcp
example.verify: claude mcp list | cat
2. Official Technical Reference & Guides (yichuan-w/yichuan-w.github.io)
File: README.md
Yichuan Wang's Personal Website
[click here to browse](http://yichuan-w.github.io/)
[click here to edit](https://github.com/yichuan-w/yichuan-w.github.io/edit/master/docs/index.md)
File: docs/index.md
hide:
- navigation
- toc
Yichuan Wang
Bio {align=right style="width:7.5em; margin-left: 0.5em; margin-top: 0.5em; border-radius: 1em;"}
I am a second-year PhD student fortunate to be advised by Joseph E. Gonzalez and Matei Zaharia at UC Berkeley in SkyLab. I also collaborate closely with Sewon Min and Ion Stoica. I received my B.E. in Computer Science and Technology from the ACM Honored Class at Shanghai Jiao Tong University (SJTU). My research focuses on machine learning systems, particularly agent infrastructure, especially search infrastructure, LLM serving, and LLM post-training. My research has received best-paper awards at MLSys. If you share similar research interests, I'd love to connect and chat!
During my undergraduate studies, I collaborated with Prof. Jinyang Li at NYU and Dr. Minjie Wang on projects related to scaling up GNN training. At SJTU, I was supervised by Prof. Quan Chen at SJTU.
I enjoy connecting with researchers and practitioners in the field. Feel free to reach out via email, WeChat (15858459091), or Twitter. For more details about my background, you can view my CV and Statement of Purpose from my graduate school applications.
In my spare time, I contribute to open-source projects like SGLang and torchtitan.
More recently, I have dedicated significant time to maintaining LEANN , a system that manages every piece of your context on device and enables seamless interaction with LLMs. It has evolved from a research prototype into an open-source community project with over 20 active contributors.
I also maintain PixelRAG , a multimodal RAG system that retrieves directly over web screenshots instead of extracted text — keeping the layout, tables, and figures that text-only pipelines throw away. You can try it in the live playground.
:fontawesome-solid-inbox: Work Email: [email protected]
:fontawesome-solid-inbox: Personal Email: [email protected]
:fontawesome-solid-envelope:
:academicons-google-scholar:
:fontawesome-brands-github:
:fontawesome-brands-linkedin:
:fontawesome-brands-twitter:
:academicons-cv:
News
[08/2026] We wrote a blog post on killing the training-inference numeric mismatch in RL for linear attention, and on whether it actually helps async RL. There is also a thread on it.
[06/2026] Excited to release PixelRAG , a next-generation search assistant for LLMs: it retrieves directly over web screenshots instead of extracted text, keeping the layout, tables and figures that text-only pipelines throw away. Read the paper, or try it in the live playground!
[05/25/2026] I joined the Meta MSL Infra PyTorch team as a Research Scientist Intern, working on post-training infrastructure.
[05/20/2026] LEANN received the MLSys 2026 Best Paper Award!
[01/26/2026] LEANN is accepted in MLSys, see you in Bellevue!
[12/12/2025] Happy to release DS SERVE, a framework for efficient and scalable neural retrieval and the largest open vector store over pre-trained data. Check out our live demo! Fun fact: Jinjian is the first undergraduate I mentored at Berkeley—it was a wonderful collaboration!
[09/22/2025] DLPM is verified to work on real workload at Yandex LLM serving stack!
[08/10/2025] We release LEANN , a fully local RAG with 97% storage savings. One of the first open-source engines to bring semantic search to Claude Code, all with zero cloud cost and full privacy. Drop me an email if you’d like to contribute — we’re on a mission to build an engine that understands every piece of data on your PC.
[06/20/2025] I will attend SIGMOD in Berlin this year. Let's chat about vector database, RAG, LLM and we can hang around!
[01/01/2025] Happy to share my first work at Berkeley, DLPM — an efficient, fair, and load-balanced LLM serving engine. Thanks to all collaborators from LMSys!
[08/29/2024] After being rejected once, DiskGNN has finally been accepted by SIGMOD25. See you in Berlin! :smile:
[08/13/2024] Starting my PhD journey at SkyLab—excited. :party_popper:
[05/08/2024] We put DiskGNN on Arxiv If you want to train a super-large-scale GNN while balancing speed and accuracy, you should try it! We will integrate it into DGL asap. :fire:
Education
UC Berkeley { align=right style="height:6em; border-radius: 0.5em;"}
Starting from Sept. 2024
PhD in EECS
Shanghai Jiao Tong University{ align=right style="height:6em; border-radius: 0.5em;"}
Sept. 2020 -- June 2024
B.Eng. in Computer Science at ACM Honors Class, advised by Prof. Quan Chen, Prof. Yong Yu
New York University Courant Institute{ align=right style="height:6em; border-radius: 0.5em;"}
July. 2023 -- Dec. 2023
Research assistant, advised by Prof. Jinyang Li
Publications
1.LEANN: A Low-Storage Vector Index pdf poster repo
Yichuan Wang, Zhifei Li, Shu Liu, Yongji Wu, Ziming Mao, Yilong Zhao, Xiao Yan, Zhiying Xu, Yang Zhou, Ion Stoica, Sewon Min, Matei Zaharia, Joseph E. Gonzalez
MLSys 2026, Best Paper Award (Short version in VecDB@ICML2025)
2.PIXELRAG: Web Screenshots Beat Text for
Retrieval-Augmented Generation pdf repo playground
Yichuan Wang*, Zhifei Li*, Zirui Wang, Paul Teiletche, Lesheng Jin, Matei Zaharia†, Joseph E. Gonzalez†, Sewon Min†
*indicates equal contribution, †indicates equal advising
Preprint (Short version accepted by EMM-QA@ICML2026: Efficient Multimodal Question Answering)
3.DS SERVE: A Framework for Efficient and Scalable Neural Retrieval pdf website poster repo
Jinjian Liu*, Yichuan Wang*, Xinxi Lyu, Rulin Shao, Joseph E. Gonzalez, Matei Zaharia, Sewon Min
*indicates equal contribution
Accepted by AAAI 2026 Demo
4.Locality-aware Fair Scheduling in LLM Serving pdf poster
Shiyi Cao*, Yichuan Wang*, Ziming Mao, Pin-Lun Hsu, Liangsheng Yin, Tian Xia, Dacheng Li, Shu Liu, Yineng Zhang, Yang Zhou, Ying Sheng, Joseph E. Gonzalez, Ion Stoica
*indicates equal contribution
Preprint
5.DiskGNN: Bridging I/O Efficiency and Model Accuracy for Out-of-Core GNN Training pdf poster slides
Renjie Liu*, Yichuan Wang*, Xiao Yan, Zhenkun Cai, Minjie Wang, Haitian Jiang, Bo Tang, Jinyang Li
*indicates equal contribution
Accepted by SIGMOD 2025(Oral)
6.Optimizing Dynamic Neural Networks with Brainstorm pdf
Weihao Cui, Zhenhua Han, Lingji Ouyang, Yichuan Wang, Ningxin Zheng, Lingxiao Ma, Yuqing Yang, Fan Yang, Jilong Xue, Lili Qiu, Lidong Zhou, Quan Chen, Haisheng Tan, Minyi Guo.
Accepted by OSDI 2023
7.Forming Scalable, Convergent GNN Layers that Minimize a Sampling-Based Energy pdf
Haitian Jiang, Renjie Liu, Zengfeng Huang, Yichuan Wang, Xiao Yan, Zhenkun Cai, Minjie Wang, David Wipf
Accepted by ICLR 2025
8.The Danger of Overthinking: Examining the Reasoning-Action Dilemma in Agentic Tasks pdf
Alejandro Cuadron, Dacheng Li, Wenjie Ma, Xingyao Wang, Yichuan Wang, Siyuan Zhuang, Shu Liu, Luis Gaspar Schroeder, Tian Xia, Huanzhi Mao, Nicholas Thumiger, Aditya Desai, Ion Stoica, Ana Klimovic, Graham Neubig, Joseph E. Gonzalez
Preprint
9.Autellix: An Efficient Serving Engine for LLM Agents as General Programs pdf
Michael Luo, Xiaoxiang Shi, Colin Cai, Tianjun Zhang, Justin Wong, Yichuan Wang, Chi Wang, Yanping Huang, Zhifeng Chen, Joseph E. Gonzalez, Ion Stoica
Accepted by NSDI 2026
Services
- [2026] ACM CAIS 2026 Workshop AID-Wild Reviewer
- [2026] NeurIPS 2026 Conference Reviewer
- [2026] ICML 2026 Workshop EMM-QA Reviewer
- [04/2025] SIGMOD'25 Artifact Evaluation Committee
- [03/2025] MLsys'25 Artifact Evaluation Committee
- [02/2025] Eurosys'25 Artifact Evaluation Committee
- [07/2024] SIGCOMM'24 Artifact Evaluation Committee
- [04/2024] OSDI'24 Artifact Evaluation Committee
- [04/2024] USENIX ATC'24 Artifact Evaluation Committee
Invited Talks
- [10/2025] Lightspeed interview - LEANN: Towards Lightweight Vector Search and RAG Everything on PC
- [09/2025] Bytedance Data Team - LEANN: Towards Lightweight Vector Search and RAG Everything on PC
- [07/2025] SIGMOD 2025 - DiskGNN: Bridging I/O Efficiency and Model Accuracy for Out-of-Core GNN Training
- [01/2025] LMsys - SGLang-FLPM [Video]
Mentorship
- Zhifei Li → Princeton CS PhD
- Jinjian Liu → UIUC MSCS
Misc/Hobbies
I like playing basketball! (But unfortunately, it seems that there aren't any good basketball courts near Berkeley.) So now, I might (or rather, I can only) play more badminton(I still enjoy). Besides that, I've been an NBA fan for 15 years, and I like Chris Paul! In recent years, I've also become very interested in the Premier League(I love Manchester City, come on Blue Moon!). If you share the same interests, we should chat/hang out!
Blog Posts
Here are my thoughts and comments on various research papers and topics in systems and ML, along with practical lessons learned from my development experience:
Research Vision
- Rethinking the Search Stack for the AI Era - Why we need to move beyond traditional search APIs for LLM agents and build native retrieval stacks.
Production & Entrepreneurship Thoughts
- Rethinking RL as a Service and the Position of RL Infrastructure - Why RL infrastructure is a compelling systems direction, but RL as a generic enterprise service is still a difficult business in May 2026.
Tech Development Experience
- Mastering DiskANN: Practical Lessons from Building Large-Scale Vector Search Systems - Practical recipes for configuring DiskANN, from optimal compression rates to avoiding hidden metric pitfalls, based on experience building LEANN and DS-serve.
- Lessons Learned in Development - LEANN Project - Practical insights from building RAG systems, including chunk overlap strategies, data format optimization, and embedding model comparisons
- Defending Against the Training-Inference Numeric Mismatch in RL (Especially Linear Attention) - Aug 2026. Driving the trainer/generator logprob gap to exactly zero for a Gated DeltaNet model on TorchTitan RL + vLLM, and measuring what that is actually worth under async off-policy RL.
Misc
- Opportunities in the Open-Source Agentic RL Landscape - The missing end-to-end open-source pipeline: domain-specialized RL plus on-policy distillation, and where the highest-impact contributions are.
Paper Reads
- My Thoughts on ReasonIR paper - My thoughts on ReasonIR paper, reasoning-based information retrieval
- My Thoughts on Qwen3 Embedding paper - My thoughts on Qwen3 Embedding paper, advancing text embedding through foundation models
- My Thoughts on RAGCache paper - My thoughts on RAGCache paper arxiv
- My Thoughts on Exploring Orak: A Unified Benchmark for LLM Agents in Games paper - My thoughts on Orak paper, a comprehensive evaluation on Game Plantform arxiv