pinecone

Create a Pinecone index by initializing the Pinecone client with your API key and calling pc.create_index with a name, dimension matching your embedding model, metric such as cosine, and a ServerlessSpec specifying cloud and region.

Hybrid search in Pinecone works by upserting vectors with both values and sparse_values fields, then querying with vector and sparse_vector parameters plus an alpha weight where 0 is sparse only, 1 is dense only, and 0.5 is balanced.

Pinecone is a fully managed serverless SaaS with auto-scaling and a 99.9% uptime SLA, while Chroma is self-hosted and open-source, and FAISS is an offline library for pure similarity search. Choose Pinecone when you do not want to manage infrastructure.

Pinecone supports metadata filtering using operators like $eq, $gte, $lte, $in, $and, and $or passed via the filter parameter in index.query. Filtering adds roughly 10-20ms of latency overhead per query.

Pinecone integrates with LangChain through PineconeVectorStore for similarity search and retrievers, and with LlamaIndex through its PineconeVectorStore connected to a StorageContext. Both require an existing Pinecone index.

Pinecone serverless is less suitable when you need consistent p95 latency or high sustained throughput, where pod-based indexes with dedicated resources perform better. Self-hosted alternatives fit teams avoiding SaaS dependencies.