post-graph-rag

Why it's ranked here

Most graph retrieval tools only accumulate, so a corpus spanning years hands the model contradictions. This one closes superseded facts from the order documents arrive, with no dates required from the extractor. The authors report 94.0% on the full 500-question LongMemEval set against 71.2% published for Graphiti with gpt-4o, and list the caveats themselves: different judges, different model generations, one question excluded. Those numbers are self-reported, but the harness and failing cases are said to ship in the repository. Apache 2.0, a database most teams already run, and unusually candid engineering notes make it worth a serious look.

What's good

The code refuses to fail quietly. If the vector extension was missing when a table was built, startup raises an error instead of letting every search return nothing. Embedding failures raise by default, because a stand-in vector from another model would corrupt the index. Vague links such as "relates to" are rejected at extraction, pronouns never become graph nodes, and bare money figures are filtered out. Low-value entities can be archived by an importance score, but archiving is reversible and nothing is deleted. Comments explain the concrete failure behind each choice, such as 80 transcripts collapsing onto one document key.

Tradeoffs

Every indexed chunk costs language model calls for extraction, and more for contradiction checks and cluster summaries, so ingestion is slower and pricier than plain vector search. Answer quality rests on the extraction model; the headline row uses one specific model and a three-model judging panel. The package declares Python 3.9 and up, yet its own config notes that continuous integration exercises 3.11 only. The storage layer reaches into private methods of its companion graph package, which ties upgrades of the two together. Native clustering libraries install by default.

How to use it well

Best fit: teams already on Postgres who need agent memory, contract histories, org charts or quarterly filings where the newest statement should win and older ones stay auditable. Feed documents in the order they were written, give each a stable source and title so re-indexing replaces rather than duplicates, and enable the vector extension before first start. Turn on retrieval telemetry if you want coverage reports or archiving. It is a library, not a service: you bring the database, an OpenAI-compatible endpoint and your own agent loop. For a small, static corpus with no changing facts, plain vector search is cheaper.

Observed

License

Apache-2.0, with the OSI classifier declared.

Language

Python, requires 3.9 or later; a project comment states CI exercises Python 3.11 only.

Packaging

Published on PyPI as post-graph-rag and installed with pip; built with hatchling.

Dependencies

post-graph, openai, pydantic, igraph, leidenalg.

Interface

Async Python library; no command-line entry point is declared.

Platform

Requires PostgreSQL with the pgvector extension; schema setup raises if the embedding column is missing.

Model backend

Any OpenAI-compatible endpoint for completions and embeddings, with optional fallback completion models.

Benchmarks

LongMemEval and ECT-QA figures are self-reported by the authors with stated methodological caveats.

Data retention

Entity archiving is reversible demotion, not deletion, and requires retrieval telemetry to be enabled.