Skip the Server: ChromaDB Gives You a Full Vector Database in Five Lines of Python

ChromaDB is the fastest path to a working vector database: pip install chromadb, five lines of Python, and you're doing semantic search — no Docker, no server, no cloud account.

What you'll be able to do after this:

• Add vector search to any Python project without provisioning infrastructure
• Query a collection by meaning — and filter by metadata at the same time
• Persist collections to disk and reload them across sessions with a single constructor change

Resource: Mastering ChromaDB: Step-by-Step Python Guide (YouTube, July 2026)

---

Why ChromaDB first?

pgvector (which you may have seen here) requires PostgreSQL. FAISS requires manual index management and serialization. ChromaDB runs as an embedded Python library — it creates an SQLite-backed store on disk, handles embedding generation, and exposes a collections API that abstracts the rest. It's the right first vector database for prototypes and small-to-medium production workloads before you need horizontal scale.

Install

pip install chromadb sentence-transformers

Create a persistent client and collection

import chromadb
from chromadb.utils import embedding_functions

# PersistentClient writes to disk; survives restarts
client = chromadb.PersistentClient(path="./my_vectorstore")

ef = embedding_functions.SentenceTransformerEmbeddingFunction(
    model_name="all-MiniLM-L6-v2"   # free, runs locally, 384-dim
)

collection = client.get_or_create_collection(
    name="docs",
    embedding_function=ef,
    metadata={"hnsw:space": "cosine"}   # cosine similarity
)

Switch to in-memory with chromadb.EphemeralClient() for tests; switch to chromadb.HttpClient(host=..., port=...) to point at a deployed server — no other code changes needed.

Add documents with metadata

collection.add(
    ids=["doc1", "doc2", "doc3"],
    documents=[
        "LLMs can be fine-tuned on domain-specific datasets.",
        "Prompt caching reduces API costs by 90% on repeated context.",
        "pgvector adds vector search to any PostgreSQL database.",
    ],
    metadatas=[
        {"topic": "fine-tuning", "level": "intermediate"},
        {"topic": "api-cost",    "level": "beginner"},
        {"topic": "database",    "level": "intermediate"},
    ]
)

ChromaDB calls your embedding_function automatically — you pass raw text, not vectors.

Query by meaning and filter simultaneously

results = collection.query(
    query_texts=["how do I reduce my model API costs"],
    n_results=2,
    where={"level": "beginner"},     # metadata filter applied before ranking
    include=["documents", "distances", "metadatas"]
)

print(results["documents"])
# → [["Prompt caching reduces API costs by 90% on repeated context."]]
print(results["distances"])
# → [[0.12]]   (lower = more similar under cosine)

The where clause narrows the ANN search to the matching metadata subset — you get semantic ranking within the filtered population, not a post-filter on top of all results.

Swap the embedding function

The embedding_function= argument accepts any callable. Built-in options include OpenAIEmbeddingFunction, CohereEmbeddingFunction, and HuggingFaceEmbeddingFunction. Replace the constructor argument and the rest of the code stays identical — the API is model-agnostic.

When to graduate from ChromaDB

ChromaDB handles millions of vectors comfortably on a single node. When you need horizontal sharding, multi-region replication, or tight PostgreSQL integration, move to Qdrant, Weaviate, or pgvector. But for most learning projects and production prototypes, ChromaDB gets you there with zero ops overhead.

Sources: Mastering ChromaDB: Step-by-Step Python Guide (YouTube, July 2026) · ChromaDB Docs — Getting Started · ChromaDB Tutorial: From pip install to RAG Pipeline — TopicTrick