HT Systems sales support

Cited RAG Q&A over HT Systems' own product material for the Kera sit2sit patient-transfer hoist. Design in SPEC.md, working rules in CLAUDE.md, current state in CHECKPOINT.md.

Run it

OPENROUTER_API_KEY=sk-or-... ./run.sh

Then open http://localhost:3000.

The first question takes 2 minutes — the retriever chunks, embeds and indexes the corpus on first use ($0.01 in OpenRouter embeddings). Subsequent questions take a few seconds, most of it the local reranker on CPU.

What to try

question	expected
What is the safe working load of the Kera120?	figure quoted verbatim, one citation
How much does the Kera Home cost?	routed to a human, no model call
Will ACC fund this for my client?	routed to a human
Can someone on CPAP use the hoist?	refusal + the manual's ventilation exclusion
What is the airspeed velocity of an unladen swallow?	refusal, nearest source named

A citation rendered as a red unverified source marker means the model named a source that was not retrieved — a contract breach, shown rather than hidden. Report those.

Test

.venv/bin/python -m pytest tests/ -q     # 149
cd web && npx tsc --noEmit

Layout

ingest/   fetch, chunk, embed, index
rag/      retrieve (hybrid + rerank + filter), generate (citations, refusal, routing)
eval/     leakage, bake-off, ablation, gates
api/      FastAPI: /chat, /sources, /feedback
web/      Next.js chat + source panel