← Back to archive

Repo of the Day

ontogr/agentic-graphrag

Published: Nov 20, 2026

Open repository ↗

Agentic GraphRAG is an open source Python framework that connects your documents, agent activity, and decisions in one graph. Your agents can find evidence, continue work across sessions, and revisit decisions when their sources change

Summary

Agentic GraphRAG is a Python framework that connects enterprise knowledge, agent conversations, and decision records in one graph. It preserves the passages and task events behind an answer, allowing agents to cite exact evidence, resume interrupted work, and flag decisions when their source documents change.

What it is useful for

Use it when a RAG system needs durable context rather than only a chat transcript. For example, an operations team could ingest runbooks, policies, and spreadsheets, then ask an agent which approval process applies. The built-in agent searches the permitted documents, reads the matching passages, and returns numbered sources with citation checks.

The framework keeps document structure instead of relying only on arbitrary text windows. Docling extracts hierarchy from formats including Markdown, HTML, Word, PowerPoint, Excel, and PDFs. Sections, chunks, tables, and figures become graph nodes, allowing an agent to navigate from a matching passage to its section, neighboring content, or related table rows.

It also supports schema-constrained entity and relationship extraction. Results outside the declared schema remain review candidates and are excluded from search until a schema patch is approved. Entity resolution links likely duplicates without immediately merging them, while Leiden-based community detection supports broad, thematic questions over a corpus.

History features are useful for long-running work. A task stores messages, tool calls, tool results, and named steps. A later process can continue from the recorded state and rerun only unfinished steps. Decisions record their conclusion, rationale, assumptions, and supporting passages, events, or earlier decisions. When a cited document changes, the decision is flagged for review; decisions needing review are omitted from earlier-decision search and cannot be cited as support.

The framework runs from Python or through MCP, supports Neo4j or FalkorDB, and can store vectors in the graph or in Qdrant, Milvus, or Weaviate.

How engineers can use it

The documented prerequisites are Python 3.11 or newer, a running Neo4j instance, and an OpenAI-compatible LLM endpoint. Neo4j can be started locally with:

docker run -d --name agrag-neo4j -p 7687:7687 \
  -e NEO4J_AUTH=neo4j/your-neo4j-password neo4j:5.26.0-community

The supplied README cuts off during the installation section, so its exact package-install command is not available here; follow its Quickstart for installation into a fresh virtual environment.

A practical workflow is to define a schema, ingest approved documents, and restrict retrieval to the documents an agent may use. Then run the built-in agent or expose the same evidence tools through MCP. Review extraction candidates, apply approved schema changes, run community detection when needed, and review decisions flagged by document updates. Remember that tracing is opt-in, and OpenTelemetry spans may contain full question, evidence, and model text, so destination and retention policies should match your data controls.