Agentic RAG meets knowledge graphs
A new implementation wires knowledge graphs to agentic RAG, OpenAI function-calling, and graph traversal so agents can query, update, and reason over structured context automatically. The pattern shows a repeatable architecture for explainable, context-rich agents that maintain structured memory—useful as a portfolio demo or interview case study. (marktechpost.com)
Michal Sutter published the step‑by‑step tutorial on March 27, 2026, implementing IWE’s “Context Bridge” with four OpenAI function‑calling tools named iwe_find, iwe_retrieve, iwe_tree, and iwe_stats. (marktechpost.com) IWE is described in its documentation as a Rust‑powered, local‑first CLI/LSP personal knowledge‑management system that represents markdown notes as a directed, navigable knowledge graph. (iwe.md) The tutorial wires OpenAI transforms (summarization, link suggestion, todo extraction) directly into the graph, includes DOT graph export and a document “squash” operation for consolidation, and demonstrates agents performing multi‑hop reasoning across linked notes. (marktechpost.com) A companion GitHub project labeled “agentic‑rag‑knowledge‑graph” shows a production‑style stack using PostgreSQL with pgvector for semantic search and Neo4j with Graphiti for temporal knowledge‑graph features. (github.com) A separate Context Bridge package published on PyPI advertises crawling technical documentation via Crawl4AI to produce RAG‑ready knowledge bases and developer tooling for doc ingestion. (pypi.org) OpenAI’s developer cookbook on Temporal Agents recommends emitting time‑stamped triplets for temporal knowledge graphs to enable precise time‑based querying and trend analysis, which aligns with the tutorial’s emphasis on stats and traversal tools. (developers.openai.com) Both the tutorial and the GitHub implementation highlight a multi‑agent Agentic RAG pattern where agents maintain isolated vector stores while sharing the same graph to support domain‑specialized retrieval and explainable multi‑hop answers. (marktechpost.com)
