Chroma

Chroma is a local-first, embedded vector database — no server to run, one pip install, data on disk. It's the default recommendation for learning and prototyping in this section.

from langchain_chroma import Chroma
from langchain_openai import OpenAIEmbeddings

embeddings = OpenAIEmbeddings(model="text-embedding-3-small")

vector_store = Chroma(
    collection_name="example_collection",
    embedding_function=embeddings,
    persist_directory="./chroma_langchain_db",  # omit for in-memory only
)

vector_store.add_documents(documents)
vector_store.similarity_search("query text", k=4, filter={"source": "intro"})

Persistence and collections

Passing persist_directory writes the index to disk so it survives process restarts; omit it for a throwaway in-memory store during experimentation. A collection_name scopes documents — multiple collections can share one persist_directory without mixing results, useful for keeping test data separate from real data during development.

Metadata filtering

The filter argument on similarity_search restricts candidates to documents whose metadata matches — set at load time in Document Loaders, since metadata can't be recovered after chunking. Filtering runs alongside the vector search, not after it, so a narrow filter doesn't cost you full-index accuracy.

tip

Chroma is the right default while you're still iterating on chunking and retrieval strategy — the zero-setup cost means every dead end costs you nothing but time.

See also

• Overview — the shared VectorStore interface.
• Comparison — when to graduate off Chroma.