What it does

GraphMemory provides an embedded GraphRAG engine built on DuckDB that combines vector search (HNSW), full-text (BM25), and hybrid retrieval with a fluent query builder and multi-hop traversal. It adds extraction tools (DSPy) to convert text into nodes and edges, fuzzy deduplication and merge strategies, NetworkX-based graph algorithms (PageRank, centrality), and a lightweight D3.js visualizer. The library is single-package and file-backed, intended for local or prototype deployments where a serverless, single-writer graph is desirable.

When to use it

Use GraphMemory when your retrieval needs go beyond pure vectors or pure graph databases — e.g., RAG workflows that require traversals across entities, or when you want to extract structured knowledge (entities & relations) from documents and keep it deduplicated and queryable. It is suited for prototyping, small-to-medium datasets, and applications where embedding model vector lengths are fixed. Avoid when you need distributed multi-writer graph databases at massive scale (hundreds of millions of nodes).

What's included

Scripts: none detected in the fetch, but examples are present in examples/ covering extraction and ingest patterns.
References: none bundled, but the repo includes multiple examples and tests.
Instructions: init snippets, insert/merge patterns, hybrid search examples, GraphRAG retrieval and ask() patterns, DSPy extraction usage, two-pass dedup best practices, and gotchas for production use.