grafeo-langchain

LangChain graph store and vector store backed by GrafeoDB: an embedded graph database with native vector search.

No servers, no Docker, no configuration. Just uv add and go.

Install

uv add grafeo-langchain

# Optional: langchain-graph-retriever integration (requires >=0.8)
uv add "grafeo-langchain[retriever]"

Quick Start

Knowledge Graph (GraphStore)

Store LLM-extracted triples and query them with GQL/Cypher:

from langchain_openai import ChatOpenAI
from langchain_experimental.graph_transformers import LLMGraphTransformer
from langchain_core.documents import Document
from grafeo_langchain import GrafeoGraphStore

llm = ChatOpenAI(model="gpt-4o-mini", temperature=0)
transformer = LLMGraphTransformer(llm=llm)

documents = [
    Document(page_content="Alice works at Microsoft. Bob works at Google. Alice knows Bob."),
]
graph_documents = transformer.convert_to_graph_documents(documents)

store = GrafeoGraphStore(db_path="./knowledge.db")
store.add_graph_documents(graph_documents, include_source=True)

results = store.query("MATCH (p:Person)-[:WORKS_AT]->(c) RETURN p.node_id, c.node_id")
print(store.get_schema)

Vector + Graph Retrieval (GraphVectorStore)

Combine vector similarity search with graph traversal for Graph RAG:

from langchain_openai import OpenAIEmbeddings
from grafeo_langchain import GrafeoGraphVectorStore

embeddings = OpenAIEmbeddings(model="text-embedding-3-small")
store = GrafeoGraphVectorStore(
    embedding=embeddings,
    db_path="./doc_graph.db",
    # embedding_dimensions auto-detected from the model
)

store.add_texts(
    texts=["Python is a programming language...", "Guido van Rossum...", "ABC influenced..."],
    metadatas=[
        {"id": "python", "__graph_links__": [{"target_id": "abc", "type": "INFLUENCED_BY"}]},
        {"id": "guido"},
        {"id": "abc", "__graph_links__": [{"target_id": "python", "type": "INFLUENCED"}]},
    ],
    ids=["python", "guido", "abc"],
)

# Standard vector search
docs = store.similarity_search("What programming languages exist?", k=2)

# Vector search + graph traversal
docs = store.traversal_search("What programming languages exist?", k=4, depth=2)

# MMR-diversified graph traversal
docs = store.mmr_traversal_search("programming history", k=4, depth=2, lambda_mult=0.7)

# Filtered search (only documents with matching metadata)
docs = store.similarity_search("languages", k=4, filter={"category": "systems"})

# Delete documents
store.delete(["python", "abc"])

Persistence

All data is stored in a single .db file when you pass db_path. Close the store, reopen it later, and your documents, embeddings, and graph links are all still there:

from langchain_openai import OpenAIEmbeddings
from grafeo_langchain import GrafeoGraphVectorStore

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

# Write phase
store = GrafeoGraphVectorStore(embedding=embeddings, db_path="./my_store.db")
store.add_texts(["Python is great", "Rust is fast"], ids=["py", "rs"])
store.close()

# Later: reopen and query
store = GrafeoGraphVectorStore(embedding=embeddings, db_path="./my_store.db")
docs = store.similarity_search("programming languages", k=2)
store.close()

Omit db_path (or pass None) for a purely in-memory store that is discarded when the process exits.

Graph Retriever Integration

Note: The [retriever] extra is required for this feature. Install with uv add "grafeo-langchain[retriever]" (requires langchain-graph-retriever>=0.8).

Use GrafeoAdapter with langchain-graph-retriever for advanced traversal strategies (Eager, BFS, MMR) via metadata edges:

from grafeo_langchain import GrafeoGraphVectorStore
from grafeo_langchain.adapter import GrafeoAdapter
from langchain_graph_retriever import GraphRetriever

store = GrafeoGraphVectorStore(embedding=embeddings)
store.add_texts(
    texts=["Python is a language", "Rust is a language"],
    metadatas=[{"topic": "python"}, {"topic": "rust"}],
    ids=["py", "rs"],
)

adapter = GrafeoAdapter(vector_store=store)
retriever = GraphRetriever(store=adapter, edges=[("topic", "topic")])
docs = retriever.invoke("programming")

Filters

All filter parameters use exact-match equality. Pass a dict where each key is a metadata field name and the value is the expected value. Only documents whose metadata matches every key-value pair are returned:

docs = store.similarity_search("query", k=4, filter={"category": "science", "year": 2024})

Supported value types: str, int, float, bool. Compound types (lists, dicts) are not supported as filter values.

Graph Links Format

Graph links between documents are specified via the __graph_links__ metadata key. Each link is a dict with the following fields:

Example:

store.add_texts(
    texts=["Source document", "Target document"],
    metadatas=[
        {
            "__graph_links__": [
                {"target_id": "target", "type": "CITES"},
                {"target_id": "other", "type": "RELATES_TO", "properties": {"weight": 0.9}},
            ]
        },
        {},
    ],
    ids=["source", "target"],
)

The __graph_links__ key is consumed during ingestion and is not stored as document metadata.

Why Grafeo?

API Reference

GrafeoGraphStore

• GrafeoGraphStore(db_path=None): in-memory or persistent graph store
• .add_graph_documents(docs, include_source=False): ingest LLM-extracted graph documents
• .query(query, params=None): execute GQL/Cypher queries
• .get_schema / .get_structured_schema: inspect the graph schema
• .refresh_schema(): refresh the cached schema
• .client: access the underlying GrafeoDB instance

GrafeoGraphVectorStore

• GrafeoGraphVectorStore(embedding, db_path=None, embedding_dimensions=None): vector store with graph links (dimensions auto-detected from the model)
• .add_texts(texts, metadatas=None, ids=None): add documents with embeddings and optional graph links
• .similarity_search(query, k=4, filter=None): standard vector similarity search
• .similarity_search_by_vector(embedding, k=4, filter=None): search by pre-computed vector
• .traversal_search(query, k=4, depth=1, filter=None): vector search + graph traversal
• .mmr_traversal_search(query, k=4, depth=2, fetch_k=100, lambda_mult=0.5, filter=None): MMR-diversified traversal
• .delete(ids): remove documents by ID
• .from_texts(...) / .from_documents(...): factory methods

GrafeoAdapter

Requires uv add grafeo-langchain[retriever].

• GrafeoAdapter(vector_store): adapter for langchain-graph-retriever
• Works with GraphRetriever(store=adapter, edges=[...]) for Eager/BFS strategies

Requirements

• Python 3.12+

License

Apache-2.0