Agentic GraphRAG

Chapter 8. Optimization

A note for Early Release readers

With Early Release ebooks, you get books in their earliest form—the author’s raw and unedited content as they write—so you can take advantage of these technologies long before the official release of these titles.

This will be the 8th chapter of the final book. The GitHub repo will be made active later on.

If you’d like to be actively involved in reviewing and commenting on this draft, please reach out to the editor at @oreilly.com.

Your agent works. It reasons through multi-step problems, queries live systems, and learns from its own mistakes. In a notebook, it looks like the future of enterprise AI. Then someone asks what it costs to run.

The invoice from a single week of production traffic, every workflow node powered by a frontier model, every query routed through the most expensive option, is enough to shelve the project. Cost is only the first problem. An agent with unrestricted access to your knowledge graph can traverse from a public service catalog to internal cost data to employee records in a single query. A graph traversal that completes in milliseconds during development takes seconds at production scale, and a federation of specialist models that reasons beautifully offline becomes a latency bottleneck when an SRE is waiting for an incident diagnosis.

Production is where three forces collide: cost, governance, and performance. This chapter addresses all three.

We will start with selective intelligence, the ...