Support two indexing styles: ColBERT/PLAID style optimisation and HNSW-style uncompressed indexes

AnswerDotAI / RAGatouille

Easily use and train state of the art late-interaction retrieval methods (ColBERT) in any RAG pipeline. Designed for modularity and ease-of-use, backed by research.

Apache License 2.0

2.8k stars 196 forks source link

Support two indexing styles: ColBERT/PLAID style optimisation and HNSW-style uncompressed indexes #110

Open bclavie opened 7 months ago

bclavie commented 7 months ago

Currently, we only use the ColBERT optimised indexes, or index-free in-memory encodings.

For low-to-medium volumes of documents, not using the ColBERT optimisation can have advantages: easier CRUD, potentially better performance, the concept is easier to grasp, quicker indexing time... (and tradeoffs: much bigger footprint, no constant-time querying, etc...)

We should provide the ability to create both kinds of indexes, and let the user decide which indexing style they wish to use.

0-hero commented 7 months ago

Are you looking to use something like FAISS or https://github.com/nmslib/hnswlib ?