Offline evidence retrieval experiment¶
Status: local retrieval-only prototype, not a deployed RAG chatbot.
The script selects exact sections from four allowlisted Pomona operational documents: calibration provenance, temporal checks, advice boundaries and twin calibration. It returns section IDs, source paths, full-document and excerpt SHA-256 hashes, and exact excerpts. Unknown queries return insufficient evidence. Citation verification rejects unknown IDs, altered quotes, changed documents, and path substitutions. Keyword matches do not prove relevance or entailment.
No model, network, embeddings, external corpus download, dose calculation or actuator tools are involved. This is the reproducible citation mechanism inspired by AgriIR/Miranda, not a reimplementation of those systems. The experiment is deliberately not attached to the production advice endpoint or used to generate new claims. Internal docs are not a validated agronomic knowledge base.
Next gate: select licensed, crop/system-specific authoritative hydroponic references; manually review question/passage pairs, conflicts and abstentions. Then compare retrieval-only, rules-only and retrieval-assisted explanations. Keep observed sensor evidence separate from documentary guidance. An LLM must not create observations, calibrations, citations, or control commands.
Optional language/notification work is deferred pending delivery-channel and privacy choices. WhatsApp/cloud services are not required by this prototype.