Architecture lab / 04
RAGOpsUnder the hood.
A small retrieval service surrounded by explicit quality gates, experiment tracking, and runtime instrumentation.
Evaluation-driven retrievalSelect a component to inspect
Deterministic code
Sentence-aware chunker
Python · offsets · stable IDs
Splits documents while retaining source offsets and stable chunk identities.
- Receives
- Document text.
- Produces
- Chunks with document IDs and character offsets.
Design boundary
Gold-span matching depends on preserving the original source coordinates.
Connections
Source documents Sentence-aware chunkeringest
Sentence-aware chunker Persistent indexembed & store
AI Services & code Data & output Controls- - Verification path
The core decision
A green retrieval gate is not proof of answer quality.
Scope & limits
The benchmark contains 42 questions over 10 documents. The generation gate requires a reachable LLM endpoint; Terraform infrastructure is validated scaffolding, not a production deployment.
RAGOps design & gate boundaries