An agent-friendly knowledge graph design that skips Neo4j and cuts GraphRAG costs
TheRedfather · reddit · 2026-07-21
A practical knowledge graph for agents, built 1,000x cheaper than GraphRAG
The author describes a production system for querying company knowledge with an agent, built without Neo4j or a dedicated graph database.
Why not vanilla RAG
- Top-k retrieval works for narrow fact lookup, but fails on questions like “tell me everything about X” or exact corpus-wide counts/lists.
Why not full GraphRAG
- Indexing is expensive.
- Entity resolution is hard and mistakes cascade.
- A separate graph DB adds operational overhead, especially in multi-tenant SaaS.
What they built instead
- A fixed ontology shared across customers: organization, person, product, project, event, location, etc.
- Waterfall entity resolution: alias table first, then vector similarity, then a cheap LLM only for ambiguous cases.
- Lazy, reversible merges with daily cleanup.
- “Edges” are just co-occurrence counts stored as top-N lists, not typed graph edges.
- Entity summaries are generated lazily on first request and cached.
The pitch is that this gives a graph-like experience using standard DBs and search infrastructure, with lower cost and simpler ops.
More from coding & agent
- A tutorial shows how to rebuild Claude Code inside Pi with harness engineering — eptwts · 2026-07-21
- For agents and chatbots, the RAG vs. tuning choice depends on the problem — Roker_51 · 2026-07-21
- Reddit user asks for a practical local Ollama-and-Hermes desktop agent stack — Tonka-Jahari-Pizza · 2026-07-21
- Using Codex to set up Claude Code because opening a terminal takes one extra click — emollick · 2026-07-21
- An interactive Zarr explainer shows how AI is changing technical education — MaxLenormand · 2026-07-21
- Most people still use Claude like a chatbot, not an agent — heypearlai · 2026-07-21