Build a Traceable RAG Execution Pipeline
This project implements the lifecycle as a small, inspectable Python system. It includes document metadata, access checks, retrieval scoring, reranking, context assembly, answer/refusal/escalation behavior, traces, and golden evals.
Project strengthThis is no longer a tiny demo. The reference solution includes role-based access control, approved/draft source filtering, lexical scoring, reranking, context limits, answered/refused/escalated outcomes, citation checks, trace diagnostics, and three golden evals.
1 Implement lifecycle stages
Classify intent, build retrieval query, apply filters, score candidates, rerank, assemble context, answer or refuse, and emit a trace.
2 Add production controls
Include role-aware access checks, source status, context limits, citation rules, and escalation behavior.
3 Add golden evals
Test one answered case, one refusal case, and one escalation case. Report pass/fail results.
4 Extend one retrieval strategy
Add keyword weighting, hybrid-style scoring, multi-query expansion, or reranking logic.