Imported from faionfaion/faion-network (skills/faion/knowledge/ml-engineering/db-chroma/AGENTS.md). Install upstream with npx skills add faionfaion/faion-network --skill db-chroma. Copyright stays with the author.

Chroma Vector Database

Summary

One-sentence: Embedded SQLite/HNSW vector store for RAG prototyping with zero infrastructure and idempotent get_or_create_collection ingestion; single-process only.

One-paragraph: Chroma runs in-process (PersistentClient backed by SQLite + HNSW). Distance metric is set at collection creation and cannot be changed without recreating. IDs are strings (integers must be stringified). get_or_create_collection makes re-runs idempotent. Integrates directly with LangChain and LlamaIndex. Use Chroma for dev / eval / single-developer prototypes; promote to Qdrant or Pinecone before crossing 1M vectors, multi-tenant SaaS, or multi-process write workloads.

Ефективно для: Solopreneur prototyping a RAG pipeline locally — closes the gap between "open notebook" and a persistent vector store without Docker overhead.

Applies If (ALL must hold)

• Single-developer prototype OR CI eval harness OR notebook-based research.
• Corpus < 1M vectors.
• Single-process write workload; no concurrent writers.
• Acceptable to migrate to a production vector store before scale-up.

Skip If (ANY kills it)

• Production deployment with >1M vectors — load [[db-qdrant]] or [[db-weaviate]].
• Multi-tenant SaaS — Chroma lacks tenant isolation.
• High-concurrency writes — SQLite corrupts under concurrent writers.

Prerequisites

Assumes Loaded

Content (load on demand)

Task Routing

Templates

Scripts

Related

• [[db-qdrant]] · [[db-weaviate]] · [[db-comparison]] · [[rag-architecture]]

Decision tree

The mandatory tree at content/06-decision-tree.xml routes by corpus size, tenancy, and concurrency to Chroma, Qdrant, Weaviate, or pgvector. Use it before instantiating a client so the scale-up path is explicit.