Sprint record — LangGraph live validation (2026-05-18)

Archived. Active design: LANGGRAPH.md. Roadmap: PRODUCT_ROADMAP.md.

Goal

Prove LangGraph KB path on live AWS (Bedrock + Knowledge Base) with real sources, then add opt-in web research.

Outcome

Delivered	Notes
RAG_ENGINE=langgraph on live AWS	KB answers with real corpus sources (LMS / policy articles)
Linear graph	condense → retrieve → generate → format (rerank/multi-query added in Phase 5 (retrieval))
Web branch	research_mode=web; mock + optional Tavily
Vue toggle	VITE_WEB_RESEARCH_ENABLED + disclaimer UI (Phase 6b)
CI	tox -e lint,backend green with mock RAG (RAG_ENGINE=chain in conftest)

Deferred from this sprint: LangGraph-native SSE (Phase 6 (agentic)a); strict RAGAS ±0.02 gate (Phase 3 lite completed separately).

Acceptance (manual, live AWS)

1. Set RAG_ENGINE=langgraph, AWS credentials, RAG_FORCE_MOCK=false.
2. Ask a Canvas LMS or policy question in Vue; confirm metadata.sources with KB URLs.
3. Optional: compare same prompt on RAG_ENGINE=chain.
4. LangSmith: run name chat-session-<id> with condense / retrieve / generate spans.

Related

• LANGGRAPH.md — current graph design
• WEB_RESEARCH.md
• EVALUATION.md