ejaasaari

ejaasaari / lemur

Public

LEMUR reduces multi-vector retrieval for late interaction models such as ColBERT into regular single-vector retrieval.

approximate-nearest-neighbor-search colbert embeddings multi-vector multi-vector-embeddings
25
3
100% credibility
Found Feb 04, 2026 at 21 stars -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
Python
AI Summary

LEMUR is a Python library implementing a fast learned method for multi-vector retrieval to approximate and rerank maximum similarity scores between queries and large corpora of token embeddings.

How It Works

1
📚 Discover Lemur

You learn about Lemur, a speedy tool that helps find the best matches in huge collections of documents or items super fast.

2
🛠️ Set it up

You add Lemur to your computer with a simple command, and it's ready to go in moments.

3
💾 Prepare your collection

You gather your big list of items, like documents, each described by numbers from their key parts, and note how many parts each has.

4
🧠 Train the smart finder

You teach Lemur about your collection by letting it learn patterns – it practices on samples to get really good at spotting similarities.

5
🔍 Ask your questions

You describe what you're looking for with similar number descriptions, and Lemur scans the whole collection lightning-fast.

6
⚡ Get top matches

Lemur quickly narrows down thousands of possibilities to your best handful of matches, perfectly ranked.

✅ Search mastery achieved

Now you can zip through massive libraries or datasets, finding exactly what you need every time, feeling like a pro.

Sign up to see the full architecture

5 more

Sign Up Free

Star Growth

See how this repo grew from 21 to 25 stars Sign Up Free
Repurpose This Repo

Repurpose is a Pro feature

Generate ready-to-use prompts for X threads, LinkedIn posts, blog posts, YouTube scripts, and more -- with full repo context baked in.

Unlock Repurpose
AI-Generated Review

What is lemur?

LEMUR delivers fast approximate retrieval for multi-vector embeddings, like token-level reps from data lemur github repos or learned lemur setups. You feed it corpus token embeddings and doc lengths, it trains a model to score queries against docs via max inner products, grabs candidates, then reranks exactly. Python-based with Torch and NumPy, it runs on AVX-512 CPUs and pairs with pyglass for ANN on big indexes.

Why is it gaining traction?

It crushes naive similarity on variable-length docs without GPU dependency, indexing 100k+ items via cheap matrix ops or MIP-ANN. Devs love the fit-once, query-fast flow: compute query features, topk candidates, precise MaxSim rerank in one pass. Beats vanilla ColBERT-style retrieval in speed for CPU-bound servers.

Who should use this?

RAG engineers tuning semantic search on token embeddings from long docs. Backend teams at learned lemur denver startups or colfax labs handling unchunked passages. Anyone optimizing retrieval without scaling to GPU clusters.

Verdict

Grab it for prototyping learned multi-vector search—simple API shines on mid-scale data. But 24 stars and 1.0% credibility score scream early days; sparse tests and docs mean validate hard before prod.

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.