Skip to content

rag/: Vector similarity search operates directly on high-dimensional embeddings in mem #466

@aniket866

Description

@aniket866

Location

rag/ -> Retrieval-Augmented Generation pipeline


Issue

Vector similarity search operates directly on high-dimensional embeddings in memory, causing OOM errors as the knowledge base scales.


Proposed Fix

Migrate the vector index to a dedicated vector database (e.g., Pinecone, Milvus, Qdrant) or pgvector, and implement a hybrid search (sparse + dense) mechanism.


Why This Matters

In-memory FAISS/Annoy indices are not horizontally scalable and risk bringing down the entire API server.


Difficulty

Hard( Data migration, schema design for vectors, and tuning retrieval metrics (MRR, NDCG) across a distributed database.)


I would like to work on this issue.

@Eshajha19 /assign

Metadata

Metadata

Assignees

Projects

No projects

Milestone

No milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions