Just Ask: Building a Natural Language GraphQL Agent from the Graph You Already…
summary
Takes a GraphQL supergraph schema, models it directly in a graph database so an LLM traverses types and relations instead of relying on RAG embeddings, builds an agent that composes GraphQL, validates via introspection and rejects mutations, and integrates persisted queries as MCP tools; a demo agent answers user questions (depth up to 8), uses a 5-minute cache, and instruments tool calls with Benchmark Broccoli; token use drops from ~100k for raw SDL to ~19k across ten queries in one example, and the system works end-to-end.
AI-written summary · about
chapters (15)
- 0:00Introduction
- 1:59GraphQL onboarding problems
- 3:48LLM integration attempts
- 5:58RAG limits for GraphQL
- 7:21GraphQL schema as graph DB
- 9:52Schema to graph DB conversion
- 10:34Selective context retrieval
- 12:20Intent detection strategy
- 14:05Guardrails and validation
- 15:32Persisted queries and MCP
- 16:36Architecture walkthrough
- 18:02Live demo and MCP
- 20:21Benchmarking MCP server
- 21:48Project resources
- 22:31Q&A
related talks
15mTraining and Using Generative AI Models with GraphQL
27mHow GraphQL Gives Coding Agents Eyes on What Actually Matters
40mAgent-Ready Graphs: Engineering the Foundations for Discoverability, Safety, and QoS
24mFrom Query to Conversation: GraphQL as an AI Interface Layer
27mGraphQL Meets LLMs & Agents: Building Production AI at Starbucks Scale
21mBuilt to Evolve: 13 Years of GraphQL
31mWhy AI needs GraphQL
28mTeach Yourself GraphQL in 2026: An Anti-blueprint