Documentation authority
This site is the revised implementation blueprint. The uploaded 49-page plan remains the audit baseline for programme scope, Urdu voice, multimodal reasoning, knowledge governance, safety and the controlled pilot.
The design discussion changes several architectural decisions: the frontier model is the primary analyst; retrieval is hierarchical and research-capable; confidence controls research depth; the pilot is intentionally lean; and the primary farmer interaction is now a live Urdu audio-video consultation.
Change-control rule: where this site differs from the original DOCX, the HTML blueprint records the revised proposed decision. The original file is preserved unchanged under
source/.Live-video revision
The source plan treated guided still images and short clips as the default and continuous video as a non-goal. The revised product decision is different:
- A farmer starts a continuous WebRTC audio-video consultation.
- LiveKit or an approved WebRTC equivalent transports the full call.
- The reasoning system does not send every frame to the model. It samples, scores and selects diagnostically useful frames and can request a full-resolution still.
- Short clips, still images, recorded voice and queued cases remain degradation and fallback modes.
- A field officer or agronomist may join the existing room when risk, uncertainty or policy requires human participation.
Retained decisions
| Original decision | Revised treatment |
|---|---|
| Cascaded STT → reasoning → TTS | Retained for transcripts, separate benchmarking, tool use and controlled spoken output. |
| Flutter Android-first application | Retained, now centred on a live-call screen and network degradation controls. |
| LiveKit realtime transport | Elevated from a target component to the primary interaction transport for the pilot. |
| Multimodal diagnosis | Expanded from uploaded media to stateful live-video evidence acquisition. |
| Approved local documentation | Retained as the highest operational authority. |
| Human escalation | Retained and expanded to live room join, handoff and asynchronous review. |
| Deterministic pesticide controls | Retained for exact, legal, numerical and safety-critical claims. |
| Hybrid deployment and sovereignty path | Retained. |
Revised decisions
| Area | Decision |
|---|---|
| Primary interaction | Continuous Urdu WebRTC audio-video consultation; clips and stills are fallback modes. |
| Video reasoning | Continuous transport with adaptive frame sampling, quality scoring, event-triggered bursts and model-directed still capture. |
| LLM role | The frontier reasoning model is the primary analyst: hypotheses, questions, tool use, research, synthesis and explanation. |
| Retrieval | Document → chapter → section → atomic evidence, combining metadata, lexical, semantic and optional graph signals. |
| Internet research | Controlled, cited and triggered by novelty, freshness, conflict, incompleteness or risk. |
| Confidence | Multidimensional evidence sufficiency and consequence, not a single self-reported probability. |
| Pilot topology | PostgreSQL-first application plus LiveKit, object storage and a worker; additional clusters only when measurements justify them. |
| Primary reasoner | GPT-5.6 Sol as the reference model behind a provider-neutral gateway. |
Original-plan map
| Source-plan area | Revised page |
|---|---|
| Executive decision and principles | Executive Blueprint, Design Principles |
| Live farmer interaction | Live Video Consultation, Mobile and Rural UX |
| Architecture and scale | Target Architecture, Lean Pilot |
| Speech and multimodal reasoning | Urdu Voice, Multimodal Diagnosis |
| RAG and research | Knowledge Ingestion, Hierarchical Retrieval, Web Research |
| Policy and confidence | Policy Verification, Confidence and Research |
| Security, audit and rollout | Security, Audit, Roadmap |