Where AI appears in the product
A single feature: the /chat question-answering
assistant, reachable from the REST API and MCP server (it is
not surfaced as a chat widget in the main browser UI). It
exists to explain field-guide concepts (survey numbers,
Khata, PPB, eKYC, land types, Acres.Guntas) and to summarize
records you've already fetched — for example, finding an
exact survey, adding up combined extent correctly, or listing
the top landholders in a village.
What is deliberately not AI
Everything else is ordinary, deterministic code: the location hierarchy browser, the record fetch and streaming, search, filtering, sorting, pagination, the highest-landholder calculation, and CSV/JSON export. None of that involves a model, and none of it can hallucinate a number — the acres/guntas math and aggregation logic is fixed arithmetic, covered by tests, not a model guess.
What the assistant is allowed to touch
Even in AI-backed mode, the assistant does not get free rein over the dataset. It can call a small, fixed set of read-only tools: list districts/mandals/villages, list survey or Khata numbers, and inspect a cached village's records, which returns aggregate figures plus at most 20 matching rows per call. The full multi-thousand-row village payload and any write-capable operation are excluded from what the model can invoke.
Automated decisions
The assistant does not make decisions about you or on your behalf — it answers questions and summarizes data you requested. It does not determine eligibility, approve or deny anything, or take any action against the government portal (all tools are read-only). It is explicitly designed to not claim that a record establishes clear title or that land can legally be bought or sold; it explains what a source field reports and recommends official verification for consequential decisions.
Human review
Individual chat responses are not reviewed by a human before you see them — the assistant is designed to answer instantly. The system prompt, the tool boundaries, and the deterministic fallback logic are, however, written and reviewed by the people maintaining this project. If you spot a pattern of bad answers, see Reporting a bad answer.
Accuracy and limits
The assistant states whether it used live fetched records, the portal hierarchy, or its built-in field guide, so you can judge how grounded an answer is. It can still be wrong, especially about anything outside the bounded tool set or outside the current session's fetched data. Treat its answers as a convenience for understanding and summarizing public records, not as legal or financial advice.
Reporting a bad answer
If the assistant gives an answer that's wrong, misleading, or claims something it shouldn't (like asserting clear title), report it through the project repository or the API documentation at /docs, with the question and response if possible.