Chroma: Embedding Database for AI & RAG Applications

Setup: Ephemeral, Persistent & HTTP Client

Three client modes for different use cases — from in-memory testing to full HTTP server deployment.

pip install chromadb openai sentence-transformers

import chromadb
from chromadb.utils import embedding_functions

# 1. Ephemeral (in-memory, lost on exit) — great for tests
client = chromadb.Client()

# 2. Persistent (saves to disk as SQLite + parquet)
client = chromadb.PersistentClient(path="./chroma_db")

# 3. HTTP Server mode — run server first: chroma run --path ./chroma_db
client = chromadb.HttpClient(host="localhost", port=8000)

# Create a collection with OpenAI embedding function
openai_ef = embedding_functions.OpenAIEmbeddingFunction(
    api_key="YOUR_OPENAI_API_KEY",
    model_name="text-embedding-3-small"
)

collection = client.create_collection(
    name="my_rag_docs",
    embedding_function=openai_ef,
    metadata={"hnsw:space": "cosine"}   # cosine | l2 | ip
)
print(f"Collection: {collection.name}, count: {collection.count()}")

Add, Query, Update & Delete

The complete CRUD API for Chroma collections — Chroma auto-embeds text using the collection's embedding function.

# ADD documents — embedding auto-generated by the embedding function
collection.add(
    documents=[
        "LangChain is a framework for building LLM applications",
        "Chroma stores embeddings for similarity search",
        "RAG combines retrieval with language model generation",
    ],
    metadatas=[
        {"source": "blog", "year": 2024, "category": "llm"},
        {"source": "docs", "year": 2024, "category": "database"},
        {"source": "paper", "year": 2023, "category": "ai"},
    ],
    ids=["doc-1", "doc-2", "doc-3"]
)

# QUERY — similarity search with metadata filter
results = collection.query(
    query_texts=["how does retrieval augmented generation work?"],
    n_results=2,
    where={"category": {"$in": ["ai", "llm"]}},    # metadata filter
    include=["documents", "metadatas", "distances"]
)
for doc, meta, dist in zip(
    results["documents"][0],
    results["metadatas"][0],
    results["distances"][0]
):
    print(f"dist={dist:.4f} | {doc[:60]} | {meta}")

# UPDATE metadata or documents
collection.update(
    ids=["doc-1"],
    metadatas=[{"source": "blog", "year": 2025, "category": "llm"}]
)

# UPSERT (insert or update)
collection.upsert(
    documents=["Updated content about LangChain LCEL"],
    ids=["doc-1"]
)

# DELETE by ID or filter
collection.delete(ids=["doc-3"])
collection.delete(where={"year": {"$lt": 2024}})

Local Embeddings with SentenceTransformers

Use a local embedding model for zero API cost and data privacy — no data leaves your machine.

from chromadb.utils.embedding_functions import SentenceTransformerEmbeddingFunction

# all-MiniLM-L6-v2: 384 dims, very fast, great for local RAG
local_ef = SentenceTransformerEmbeddingFunction(
    model_name="all-MiniLM-L6-v2"
)

local_collection = client.create_collection(
    name="local_docs",
    embedding_function=local_ef
)

# Works identically — model runs locally, no API calls
local_collection.add(
    documents=["Local embedding with no API cost"],
    ids=["local-1"]
)

results = local_collection.query(
    query_texts=["free embeddings"],
    n_results=1
)
print(results["documents"])

Chroma vs Pinecone vs Qdrant vs Weaviate