VantaDB × LangChain

LangChain VectorStore adapter for VantaDB.

Install

Not on PyPI yet. The vantadb-langchain package builds and passes twine check locally; it will go live with the first adapters-v* tag release. Until then, install from source.

# Today, from a repo checkout
cd integrations/langchain && pip install .

Install from PyPI (after first release)

pip install vantadb-langchain

Quickstart

from langchain_openai import OpenAIEmbeddings
from vantadb_langchain import VantaDBVectorStore

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

store = VantaDBVectorStore(
    embedding=embedding,
    db_path="./my_data",
    namespace="docs",
)

# Add documents
store.add_texts(
    ["VantaDB is an embedded vector database written in Rust.",
     "It supports hybrid search across vectors and text."],
    metadatas=[{"source": "docs"}, {"source": "docs"}],
)

# Search
results = store.similarity_search("vector database", k=5)
for doc in results:
    print(doc.page_content, doc.metadata)

API

• similarity_search(query, k=4) — search by text
• similarity_search_by_vector(embedding, k=4) — search by raw vector
• similarity_search_with_score(query, k=4) — search with cosine distance
• add_texts(texts, metadatas=None, ids=None) — add documents
• delete(ids=...) — delete by key
• from_texts(texts, embedding, metadatas=None, ids=None) — create + populate store

LangGraph (INTG-01)

Persistent LangGraph adapters — same embedded database, no server. Requires langgraph-checkpoint>=2,<5 (pinned major: upstream API is unstable).

from vantadb_langchain import VantaDBCheckpointer, VantaDBStore

# Short-term memory: per-thread checkpoints (put/get/list/writes/delete)
checkpointer = VantaDBCheckpointer(db_path="./my_data")
graph = builder.compile(checkpointer=checkpointer)
graph.invoke(inputs, {"configurable": {"thread_id": "user-1"}})

# Long-term memory: hierarchical KV store, namespaces are tuples
store = VantaDBStore(db_path="./my_data", embeddings=embedding)
await store.aput(("users", "123", "prefs"), "theme", {"mode": "dark"})
item = await store.aget(("users", "123", "prefs"), "theme")
results = await store.asearch(("users", "123"), query="dark mode")

Notes:

• Namespace tuples map to VantaDB namespaces joined with / (parts must be non-empty strings without /).
• search(filter=...) matches scalar value fields exactly; nested dicts/lists are payload-only.
• Semantic query needs embeddings=; without it raises ValueError.
• TTL is not supported (supports_ttl = False); index= is accepted and ignored.

Why VantaDB?

• Embedded & local-first: the storage engine is a Rust library embedded in your process — no server to deploy, no network hop; data lives in your filesystem.
• Persistent hybrid search: vectors + BM25 text search out of the box, where LangChain's InMemoryVectorStore covers only small single-process sessions.
• Zero-setup alternative to hosted stacks: unlike Zep (requires a server) or Cognee (spins up its own knowledge-graph runtime), VantaDB is a plain library you import.

Development

pip install -e .