langchain-amu

LangChain integration for amu-pgvector: a VectorStore and Retriever backed by lineage-gated Analytical Memory Units on PostgreSQL + pgvector, part of a reference implementation of Lineage-Aware Memory Governance (Sangaraju & Vissa, IEEE Access, 10.1109/ACCESS.2026.3730363).

Every read runs under the store's own Postgres role, so row-level security gates LangChain retrieval exactly the same way it gates raw SQL or the MCP server — there is no separate access-control path to keep in sync.

Full project docs, the SQL schema, and benchmarks live in the main repository: https://github.com/sangaraju1988/amu-pgvector

Install

pip install langchain-amu

This depends on and reuses amu-pgvector's client. You'll also need the schema installed in Postgres — see the 60-second quickstart in the main README.

Quickstart

from amu_pgvector import AMUStore
from langchain_amu import AMUVectorStore, AMURetriever
from langchain_core.embeddings import DeterministicFakeEmbedding

store = AMUStore("postgresql://finance_agent:finance_pw@localhost:5433/amu_dev")
embedding = DeterministicFakeEmbedding(size=1536)  # swap for a real embeddings model

vectorstore = AMUVectorStore(store, embedding)
vectorstore.similarity_search("average customer income", k=5)

retriever = AMURetriever(vectorstore=vectorstore, k=5)
retriever.invoke("average customer income")

A restricted role (one never granted the sensitive column a cached result was derived from) gets nothing back from similarity_search, max_marginal_relevance_search, or the retriever — enforced by Postgres, not by this library.

What's in this package

• AMUVectorStore(VectorStore) -- similarity_search, similarity_search_by_vector, max_marginal_relevance_search, add_texts/add_documents (with ids=[...] upsert), delete, get_by_ids. Document.metadata carries metric_name, owner_department, definition_hash, and value alongside whatever metadata the caller supplied.
• AMURetriever(BaseRetriever) -- wraps an AMUVectorStore.

Tested against LangChain's own standard VectorStoreIntegrationTests suite; see the main repo's docs/design.md for the handful of tests that don't apply here and why (every AMU necessarily exposes governance fields in Document.metadata, which conflicts with a few of the suite's generic exact-metadata-echo assertions — content, ids, add/delete/mutate/search all genuinely work).

Links

• Main repo (SQL schema, quickstart, benchmarks, threat model): https://github.com/sangaraju1988/amu-pgvector
• Paper: Sangaraju & Vissa, IEEE Access, DOI 10.1109/ACCESS.2026.3730363
• Python client: amu-pgvector
• License: MIT