Chroma

Picture the workflow: you're building an AI app that answers questions over your company's docs, support tickets or product catalog. You need semantic search for fuzzy questions, keyword matching for exact product names and error codes, and metadata filters so one customer never sees another customer's data. That's the job Chroma is built for. Its homepage demo shows a knowledge_base collection with 1,277,467 records waiting for a query, and the code sample adds documents, runs a BM25 sparse vector query and returns ranked results in 18ms. You can start locally, then move the same setup to Chroma Cloud when you're ready for production.

What makes Chroma interesting for this use case is that it doesn't force you to stitch together a vector store and a separate keyword engine. Dense vector search, sparse lexical search (BM25 and SPLADE), trigram and regex full-text search, and metadata filtering with faceting all live in one place. There's also a Forking feature for dataset versioning, A/B testing and roll-outs, which is useful when you're testing a new embedding model against your current one. The core is Apache 2.0, it has 27k GitHub stars and claims 15M+ monthly downloads, and you get TypeScript, Python and Rust clients plus a CLI.

The Cloud version is where the architecture gets clever. Chroma stores all vectors, metadata and indexes in S3 or GCS and caches hot data in memory and warm data on SSD. The pitch is simple math: memory runs about $5/GB/mo while object storage is $0.02/GB/mo, and since 1GB of text can turn into 15GB of vectors, that gap matters. Chroma says this makes it up to 10x cheaper. The tradeoff shows up in its own published numbers: warm queries at 384 dimensions over 100k vectors hit 20ms at p50, but cold queries sit at 650ms p50 and 1.5s p99. If your app touches lots of rarely queried tenants, you'll feel that. For most RAG products with a steady working set, it's a sensible deal.