Chroma (ChromaDB)

Chroma (ChromaDB) est un vector database open source Python-first, lancé en 2023 par Chroma (YC W23). Conçu pour developer experience exceptionnelle — installation pip simple, API intuitive, idéal prototypage RAG et applications small/medium scale. Modes embedded (local) ou client-server.

Installation : `pip install chromadb` puis `import chromadb ; client = chromadb.Client()`. C'est tout pour démarrer en local. Mode persistent : `chromadb.PersistentClient(path="./chroma_db")`. Mode server : `chroma run --path /chroma_data`.

Features : (1) **Simple API** — `collection.add(documents=["text1", "text2"], ids=["id1", "id2"])`, `collection.query(query_texts=["search"], n_results=5)` ; (2) **Automatic embedding** — Chroma embed avec default model (all-MiniLM-L6-v2) ou custom embedding function (OpenAI, Cohere, HuggingFace) ; (3) **Metadata filtering** ; (4) **Document storage** alongside vectors (no separate storage) ; (5) **Multimodal** support via custom embedding ; (6) **HNSW** indexing ; (7) **Telemetry** opt-out ; (8) **Integration** native LangChain, LlamaIndex, HayStack ; (9) **Chroma Cloud** (private preview / launching) — managed service.

Limitations vs production vector DBs : (1) scaling limité (single-node primarily, distributed in Chroma Cloud) ; (2) less rich filtering (vs Qdrant payload conditions) ; (3) HA et replication absent (in open source self-hosted) ; (4) performance moins optimisée que Rust-based Qdrant ou C++-based Milvus pour 10M+ vectors.

Use cases idéaux : (1) prototypage RAG apps (Jupyter notebook → embed knowledge base → query) ; (2) hackathons et POCs ; (3) personal projects ; (4) educational/learning vector DB concepts ; (5) small business RAG sub-1M documents.

Production at scale : migrate to Pinecone (managed) ou Qdrant (self-hosted) ou Weaviate. Chroma stays leader for dev-friendly prototyping. Compétences AI-102, AIF-C01.