AI Architecture Bootcamp Archives

Day 25: GraphRAG for Call Graphs and Service Maps

GraphRAG for call graphs and service maps — when edges beat flat chunks for microservices reasoning.

Project: train and gate an internal SDK autocomplete student (Titan/custom) on private SDK data, served behind the router with CI evals.

Query rewriting and HyDE without inventing APIs — expand symbols and aliases; never hallucinate methods.

Ship a fine-tune only when blind A/B on real tickets moves cycle time or defect rate — not when loss curves look pretty.

Metadata filters and tenant isolation — KB filters, IAM, and the failure mode of leaking another customer’s snippets.

Rerankers after cheap recall — cross-encoders and Bedrock rerank when they buy quality, skip when they blow p95.

Know what you can fine-tune on and what you must not scrape — licenses, ToS, and customer data boundaries for training sets.

Hybrid search — BM25 + vectors + symbols — fusion ranks that beat either signal alone on code RAG.

Hide Titan vs Claude vs your distilled student behind one versioned API router with shadow traffic and instant rollback.

Project lab: build a PR review bot with diff-scoped context — only the diff plus CODEOWNERS-selected files, not the whole monorepo.