# stanford-futuredata/ColBERT

ColBERT: state-of-the-art neural search (SIGIR'20, TACL'21, NeurIPS'21, NAACL'22, CIKM'22, ACL'23, EMNLP'23)

Repository: https://github.com/stanford-futuredata/ColBERT
Canonical: https://ross.abutalabs.com/products/colbert
Language: Python
License: MIT
License Family: permissive
Last push: 2025-10-14T18:50:11+00:00

## Health v2 (maintenance only)
Score: 45/100 (v2, computed 2026-09-03T02:20:16.233290+00:00)
- activity 47, release rhythm 10, longevity 100
- inputs: {"age_days": 2291, "days_push": 323, "days_rel": 387, "gap_med": null, "n_releases_24m": 1}
- flags: none
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 3924, forks 473 (observed 2026-08-28T04:08:29.834026+00:00)

## What it is
ColBERT is a fast and accurate neural retrieval model that encodes passages and queries into token-level embedding matrices and scores them with contextual late interaction (MaxSim), enabling BERT-quality search over large corpora in tens of milliseconds. This repository is the reference Python implementation (ColBERTv2) with indexing and retrieval tooling, widely used for semantic search and RAG pipelines.

## Use cases
- build semantic search over a large text corpus
- retrieve relevant passages for a RAG pipeline
- rank documents by fine-grained relevance to a query
- index millions of passages for fast neural search
- improve retrieval quality beyond single-vector embeddings
- fine-tune a neural retriever on domain-specific data
- add neural retrieval to a question answering system

## When to choose
- you need high-accuracy neural retrieval where token-level late interaction outperforms single-vector dense models
- you want low query latency over large corpora using the PLAID engine
- you need a strong retriever component for RAG or open-domain QA systems
- you want to train or fine-tune a BERT-based retrieval model on your own data

## When to avoid
- you only need simple keyword search such as BM25 or SQL full-text without ML dependencies
- you need a managed, horizontally scalable production vector database rather than a research-grade retrieval library
- you lack GPU resources for encoding and indexing documents
- you want an off-the-shelf application instead of a Python library to integrate into your code

## Facets
- artifact type: library
- maturity: active
- function: search-engine, rag, machine-learning, deep-learning, transformers
- domain: machine-learning, deep-learning
- platform: python
- tags: neural-search, late-interaction, bert, information-retrieval, dense-retrieval, passage-ranking, semantic-search, maxsim, plaid, research, search, retrieval-augmented-generation, natural-language-processing, gpu

## Member repositories
- stanford-futuredata/ColBERT (main) score 45

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:08:29.834026+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-29T18:24:44.912241+00:00, confidence not recorded.
  - readme: https://github.com/stanford-futuredata/ColBERT (fetched 2026-08-28T04:08:29.834026+00:00, sha ce4b15d905d8)
- Data as of 2026-08-30T08:39:29.467469+00:00.