RAGatouille API Server

[0-hero]

No changes on core RAGatouille module. Just a fastapi addition

Looking for suggestion on the implementation, routes etc...

API Endpoints

Index

• POST /api/v1/index/: Create an index with documents.

Document Management

• POST /api/v1/documents/search/: Search documents.
• [ISSUE] POST /api/v1/documents/add/: Add documents to an existing index.
• [ISSUE] DELETE /api/v1/documents/delete/: Delete documents from an index.

Encode

• POST /api/v1/encode/: Encode documents.
• POST /api/v1/encode/search/: Search through encoded documents.
• POST /api/v1/encode/clear/: Clear all encoded documents.

Rerank

• POST /api/v1/rerank/: Rerank a set of documents based on a query.

Known Issues. Still exploring the cause

1. Add to index doesn’t work

2024-01-31 00:15:41,992 - app.api.endpoints.document_management - ERROR - Failed to add documents to index: Error in void faiss::Clustering::train_encoded(faiss::idx_t, const uint8_t *, const faiss::Index *, faiss::Index &, const float *) at /Users/runner/work/faiss-wheels/faiss-wheels/faiss/faiss/Clustering.cpp:281: Error: 'nx >= k' failed: Number of training points (114) should be at least as large as number of clusters (128)

2. Search after delete & Search after failed "add" document throws below error.

2024-01-30 18:11:28,756 - app.api.endpoints.search - INFO - Initiating search with query: 'egypt'. Index: 'string', k: 10
2024-01-30 18:11:28,848 - app.api.endpoints.search - ERROR - Search operation failed for query: 'egypt': 0
Traceback (most recent call last):
  File “…/RAGatouille-experiments/serve/app/api/endpoints/search.py", line 13, in search
    results = rag.search(
  File "…/RAGatouille-experiments/ragatouille/RAGPretrainedModel.py", line 292, in search
    return self.model.search(
  File "…/RAGatouille-experiments/ragatouille/models/colbert.py", line 458, in search
    document_id = self.pid_docid_map[id_]
KeyError: 0
INFO:     127.0.0.1:64170 - "POST /api/v1/search/ HTTP/1.1" 500 Internal Server Error

[0-hero]

@bclavie @anirudhdharmarajan Any thoughts on the above issues?

EDIT: Just saw issue 1 might be related to #71

[bclavie]

Thanks a lot for this! I’m likely not going to be able to review bigger PRs in the next few days but will do so ASAP (or defer to @Anmol6 / @anirudhdharmarajan if they have the bandwidth to do so!)

[0-hero]

@bclavie No problem, thanks! I’m running additional tests, while trying to fix the above

[0-hero]

@bclavie gentle bump!

[bclavie]

Hey, sorry for the delay in reviewing this, still getting some random bouts of poor health (COVID really is tricky!)

As for add_to_index, I think it's fine for it to have this known issue, as the issue is being fixed by @anirudhdharmarajan and isn't related to this service (the function itself has a fundamental issue atm). Moreover, I think this style of on-the-fly serving would work better with an HNSW style index than the optimised one currently used if CRUD is an issue!

Things are looking pretty good at first glance and I really do appreciate it, it lowers the barrier to entry to even more!

A nitpick: could you migrate the dependency management to poetry, or, at least (I don't mind this option!) pin the versions in requirements.txt? Some of those libs move fast and it'd be best to have them pinned!

[0-hero]

Sorry to hear that @bclavie. Updated the requirements & migrated the dependency management to poetry

[0-hero]

@bclavie do you think this is ready for the merge?

[hiepxanh]

@bclavie hi, what do you think, I love this feature so I can implement it and deploy on docker <3