When to use it
==============

``langchain-agensgraph`` is for teams building LLM applications **in LangChain 1.x** that want
AgensGraph to be the single backing store for their graphs, embeddings, conversations, and
agent state.

**Good fits**

- **GraphRAG over your documents.** Extract a knowledge graph from text with
  ``LLMGraphTransformer``, load it with ``add_graph_documents`` (one transaction, with
  provenance edges), and answer questions with ``AgensCypherQAChain``, which generates Cypher
  against the introspected schema and runs it read-only.
- **Vector and hybrid RAG.** Store chunk embeddings in ``AgensgraphVector`` and retrieve with
  semantic search, reciprocal-rank-fusion hybrid search, or MongoDB-style metadata filters —
  with ``effective_search_ratio`` to trade latency for recall. ``AgensGraphContextRetriever``
  adds each hit's graph neighbourhood in the same statement.
- **Stateful chat agents.** Persist a conversation with ``AgensChatMessageHistory`` (optionally
  windowed) and an agent's full state with the ``AgensSaver`` LangGraph checkpointer, so a
  thread resumes after a restart or in another process. Keep memories that outlive a thread
  in ``AgensStore``.
- **Consolidating datastores.** Keep the knowledge graph, the vector index, chat history,
  checkpoints, and long-term memory in one AgensGraph instance instead of running a separate
  graph DB, vector DB, and cache.
- **Concurrent services.** Share one ``AgensEngine`` connection pool across a graph store and
  several vector stores so requests don't serialize on a single connection.

**How it compares**

- Pick **langchain-agensgraph** when your application is already built on LangChain and you
  want direct control over the stores, retrievers, and chains.
- Pick **LlamaIndex** if you build on LlamaIndex instead.
- Pick **LightRAG** or **cognee** for a turnkey graph-RAG or memory pipeline (ingestion,
  extraction, and retrieval handled for you) rather than wiring the pieces together yourself.

**Consider an alternative when**

- You are not building a Python application — to let an agent such as Claude Desktop use
  AgensGraph through tools, use the **MCP servers**.
- You only need plain Cypher access with no LLM features — the
  `agensgraph-python <https://github.com/skaiworldwide-oss/agensgraph-python>`_ driver, which
  this integration is built on, is enough on its own.