KKissan Ki Pehchan
Overview

Design Principles

Non-negotiable rules for reasoning, evidence, safety, simplicity and government control.

BlueprintVersion 0.25 Aug 2026

Reasoning is the product

The frontier model is the principal analyst, not a decorative classifier. It frames the case, forms hypotheses, selects the next question, chooses tools, conducts research, challenges its leading explanation and communicates the result.

Evidence outranks familiarity

Vivid internal knowledge is useful for stable, low-risk explanations, but does not establish freshness, local validity or legal authorization.

Ask before guessing

When a missing fact can change the outcome, ask the farmer, request better media or call a context tool. Choose the highest-information-gain action rather than a static questionnaire.

Local authority is paramount

Departmental rules, approved documents, verified field history and current registration data govern operational action. External research can expand scientific understanding but cannot silently override local authority.

Research depth is adaptive

Use the least expensive reliable path:

  1. memory for stable low-risk knowledge;
  2. targeted question;
  3. local retrieval;
  4. live context tool;
  5. focused web research;
  6. deep multi-source research;
  7. human escalation.

Summaries navigate; sources prove

LLM-generated chapter summaries and graphs help locate evidence. Original passages, tables, rules and verified datasets remain the evidence of record.

Determinism is narrow

Use deterministic code for arithmetic, unit conversion, product registration, doses, intervals, expiry, prohibited combinations and emergency routing. Do not recreate all agricultural reasoning as a brittle expert system.

Provenance is first-class

Every material item is labelled farmer-reported, media-observed, API-observed, database-recorded, locally retrieved, externally researched, model-inferred or human-confirmed.

Consequence controls assurance

A general explanation and a pesticide instruction cannot share the same verification threshold.

Rural-first

Assume low-cost Android devices, intermittent uplink, noise, code-switching, low literacy and interrupted sessions.

Ponytail/YAGNI

Reuse before rebuilding; native platform and PostgreSQL features before new dependencies; one service before microservices; measured demand before clusters. Security, accessibility, data-loss prevention and trust-boundary validation remain mandatory.

Version and evaluate everything

Version models, prompts, tools, schemas, documents, rules, retrieval settings, applications and evaluation sets.

Continuous transport, selective reasoning

The farmer experiences a continuous call. The reasoning model receives only diagnostically relevant, quality-approved frames and explicit captures. This preserves conversational continuity without paying the cost, privacy burden or distraction of analyzing every frame.