POOR context retrieval!!

[reqyou]

Hey everyone 👋

I'm running a HYBRID search (can be replaced by any other SEARCH MODE) with top_k=40 using nanovector DB and testing queries on a small/medium-sized vector DB (~400 entities). My query is a single word, specifically an entity_name like "ID_B_123". This query/prompt is also a db node. This niode has some random node description that includes the entity_name again.

However, none of the search methods (global, hybrid, or mix) will return this relevant node as part of the results (especially when the db nodes are growing).

Has anyone else faced this issue?

I guess thats because the embedding of the prompt (just the entity name) isn't semantically rich enough to match well with the node description — even though the user is clearly requesting a very specific entity. In practice, this node should be ranked very high, possibly even with an importance score of 1.0.

I'm curious how others are solving this type of edge case.

My ideas so far:

1. Before running lightrag.query, check if the user input matches nodes entity_name. If so, use lightrag.get_entity_info(..), extract the description and inject it into the prompt — so the scoring reflects that direct match from similarity of 0.1 to 0.95 (a wild guess).
2. Try a different DB that supports better keyword or hybrid ranking behavior?

Would love to hear your thoughts or solutions if you've dealt with something similar.

Best,
req.

[ArindamRoy23]

Whats your embedding model?

[reqyou]

openai text-embedding-3-large

[ArindamRoy23]

That might be the issue. I would suggest using a smaller embedding model and checking