Semantic Introspection
summary
Takes 15 public Singapore APIs, federates them into one GraphQL schema and adds two new introspection fields: underscore_search and underscore_definitions. Uses BM25 or embeddings + cosine vector store to semantically match natural‑language queries, then fetches only the required SDL to build valid queries (curl skill). Compared to a REST baseline (~665k tokens, 40 turns) federation cuts cost ~5x and semantic introspection runs the flow in 3–5 turns and about 59k tokens, returning precise UV index, taxi counts, car‑park occupancy and computed distances.
AI-written summary · about
chapters (6)
- 0:00Introduction and context
- 2:38API challenges for agents
- 6:11The case for semantic APIs
- 10:48Semantic introspection demo
- 20:25Performance and methodology
- 22:41Q&A and concluding thoughts
related talks
26mEvolving GraphQL: Designing APIs That Deliver Insights with Intelligent Resolv…
19mThe “Food Court Menu” of APIs: A GraphQL Introduction
19mZero-to-Federation in 60 Seconds: How a GraphQL Middleware Replaced 500+ Legac…
51mQuery Understanding With Large Language Models: Techniques and...
30mBeyond Keywords: Building an AI-Powered Search That Reads Your Mind
30mGraphQL is no excuse to stop writing docs
38mNatural Language Search-Lexical and Semantic
46mAn Introduction to Semaphore: Solving Problems with Semantics