I’ve been experimenting with Pinecone for a semantic search tool, and I’m curious—has anyone successfully implemented hybrid search, combining both vector similarity and structured filters (like metadata tags or timestamps)?
I’m trying to prioritize recent results without sacrificing relevance from embeddings. Any best practices on tuning score weights, or how to keep latency low when filters are stacked on top of vector queries?
Would love to see how others are approaching this—especially at scale.
Hi @ziddihoon228 , thanks for writing in! We hear this from many customers, and the current approach is to perform multiple queries with different time ranges and merge/re-rank the results. We’re actively working on some improvements in this area, so will update you here with more when that’s available.
FWIW, typically we refer to “hybrid search” as the combination of dense (i.e. semantic) and sparse (i.e. keyword/token)
Hi @ziddihoon228,
I have been using an hybrid index from pinecone with metadata as filters.
Hybrid Index from Pinecone has several guardrails (such as the mandatory use dotproduct metric).
Sparse vector mainly used to emulate a fuzzy search without the constraints of deploying too much processes. Workspretty well except the sparse embedding model from Pinecone which only delas with english for now.