UC Berkeley 开源 LEANN 向量索引：60M 分块 RAG 索引从 201 GB 降至 6 GB

Add as a preferredsource on Google

UC Berkeley’s Sky Computing Lab built LEANN, an open-source vector index that stores almost no embeddings. The LEANN vector index recomputes embeddings at query time instead of saving them to disk. The repo benchmark shows a 60 million chunk Wikipedia index dropping from 201 GB to 6 GB. The research paper won the Best Research Paper award at MLSys 2026. So, it brings semantic search and RAG over personal data, on a laptop.

TL;DR

Size: 6 GB for 60M chunks (repo); 4 GB in the paper’s 76 GB Wikipedia setup. Default embedder: Contriever (110M params). It is an index, not an LLM, so no context length applies.

Runs on: laptops and workstations. The paper tested an RTX 4090 box with 32 GB RAM and an M1 Ultra Mac. A CPU-only Linux install exists, with no published CPU latency.

Performance: matches HNSW downstream accuracy at 90% recall, using under 5% of raw-data storage.

Best: 25.5% NQ accuracy, identical to HNSW, using 4 GB instead of 188 GB.

Worst: HotpotQA retrieval takes 44.80 s on the Mac, 50.1% of total pipeline time.

What is LEANN?

LEANN is a low-storage approximate nearest neighbor (ANN) index built on top of FAISS. It targets one problem: vector indexes are often bigger than the data they index. The research paper shows a 76 GB corpus needs 173 GB of embeddings plus a 15 GB HNSW graph. That more than doubles the raw data.

TEXTTABLELAYOUTIs your PDF parser actually getting it right?omni_parse_bench · scores text, tables & layout · open sourceStar on GitHub→Sponsored

LEANN keeps a pruned graph, drops the full vectors, and rebuilds them on demand. The code is MIT licensed and installs with uv pip install leann.

How does LEANN recompute embeddings without stalling search?

Graph indexes like HNSW visit only a small set of nodes per query. LEANN exploits that. It stores the graph plus a compact product quantization (PQ) table. At query time, it re-embeds only nodes on the search path.

3 techniques keep this practical:

Two-level search: cheap PQ distances rank neighbors first. Only top candidates get exact recomputation.

Dynamic batching: candidates from several hops are grouped, for example 64 at once, for GPU efficiency.

High-degree preserving pruning: most nodes keep few edges, while the top 2% “hub” nodes keep more.

In ablations, two-level search gives an average 1.4x speedup. Adding batching raises it to 1.8x. Embedding recomputation still accounts for about 76% of query latency.

How does LEANN perform on benchmarks?

The research team tested NQ, TriviaQA, GPQA and HotpotQA against a 60M-passage RPJ-Wiki datastore. Latency is measured at 90% Recall@3. All figures below are self-reported in the research paper.

On NQ, with Qwen3-4B on an RTX 4090, LEANN and HNSW both reach 25.5% downstream accuracy. LEANN uses 4 GB of storage. HNSW uses 188 GB. BM25 scores 18.3%, and PQ compression scores 17.9%.

The cost is search time. LEANN retrieval takes 2.48 s on NQ, versus 0.05 s for HNSW. Generation takes about 20 s, so end-to-end latency rises from 20.95 s to 23.34 s.

Across datasets on the RTX 4090, LEANN retrieval ranges from 1.12 s (GPQA) to 7.12 s (HotpotQA). The IVF-based recomputation baseline, inspired by EdgeRAG, is up to 200x slower. LEANN improves exact match by up to 11.8% over BM25.

The GitHub repo reports similar savings on personal data. 400K chat messages shrink from 1.8 GB to 64 MB. 38K browser entries shrink from 130 MB to 6.4 MB.

LEANN vs other vector indexes

ApproachPruned graph, recompute on demandGraph plus all vectors in RAMGraph and vectors on SSD, PQ in RAMIVF clusters, recompute on demandCompressed vectors plus graph

Storage, 76 GB corpus4 GB188 GB270 GBBelow 5% of raw data20 GB

NQ retrieval at 90% recall (RTX 4090)2.48 s0.05 s0.03 s307.61 sFails to reach 90%

HotpotQA retrieval at 90% recall (RTX 4090)7.12 s0.05 s0.11 s429.46 sFails to reach 90%

NQ downstream accuracy (Qwen3-4B)25.5%25.5%Not disclosedNot disclosed17.9%

LicenseMITMITMITNot disclosedVaries by implementation

Storage, latency and accuracy figures come from the LEANN paper v2, Tables 1 and 2 and Section 7.2 (self-reported). License and design rows link to each project’s own source.

Where can developers use LEANN?

The repo ships ready apps for documents, Apple Mail, Chrome history, WeChat, iMessage, and ChatGPT or Claude exports. MCP readers pull live data from Slack and Twitter bookmarks. LEANN also runs as a semantic search MCP server for Claude Code.

On 30 SWE-Bench Pro tasks from ContextBench, the team reports 24.2% initial relevant-code recall versus 11.4% for BM25. The agent also used 8.4% fewer tokens.

HNSW is the default backend, with DiskANN for faster search. Generation works with Ollama, Hugging Face, Anthropic, or any OpenAI-compatible API. It supports Ubuntu, Arch, WSL, macOS and Windows. Some apps, like Apple Mail search, are macOS only.

Key Takeaways

LEANN stores a pruned graph and recomputes embeddings at query time.

60M chunks: 201 GB down to 6 GB in the repo benchmark.

Matches HNSW accuracy (25.5% on NQ) at 90% recall.

Search costs seconds: 2.48 s vs 0.05 s for HNSW on NQ.

MIT licensed and an MLSys 2026 Best Research Paper.

Asif RazzaqWebsite | + postsBio ⮌

Asif RazzaqNace AI Releases Drex for Search for AI Agents: 92.05% on BrowseComp-Plus With 12.6 Search Calls

$uv pip install omni-parse-bench✓text✓tables✓layout→ scored per pageStar on GitHub→Sponsored

RELATED ARTICLESMORE FROM AUTHOR

Mirendil Launches Fern 0.1, Its First Model, With 10M-Token Context

OpenAI Launches GPT-6.1 Terra, Halving GPT-6.1 Sol’s API Price for Agent Work

Nace AI Releases Drex for Search for AI Agents: 92.05% on BrowseComp-Plus With 12.6 Search Calls