An AI citizen-services assistant grounded on public datasets, with governance built in

Citizen asks. A citizen asks something like “What documents do I need to renew my licence?” in their own words, in Nepali or English.

• Least-privilege accessThe assistant can only reach the exact datasets a given service needs, nothing more.
• Approved sources onlyAnswers are built from vetted official records, never from the open internet or guesswork.
• Personal data protectedNames, IDs, and other private details are masked or redacted before the model ever sees them.
• Every answer citedEach response links back to the rule or dataset it came from, so a citizen or officer can verify it.
• Full audit trailEvery query and data access is logged in a tamper-evident record for later review.
• Human escalationAnything sensitive or uncertain is handed to a named official instead of being answered automatically.

Together these controls turn a helpful chatbot into something a government office can actually stand behind.

Governance is wired into the path: an access engine and PII redaction gate every retrieval, and an audit log records it all. Tap any node for its role and tech.

1. Citizen client (Client, Web · mobile · IVR): A citizen asks a plain-language question — 'which documents renew my licence?' — in Nepali or English, over web, app or voice. Low-bandwidth channels matter for rural reach.
2. API gateway (API, Route handlers · rate limit): Single entry point. Authenticates the session, rate-limits, and hands every request to the access engine before anything is retrieved — governance runs first, not last.
3. Access / policy engine (Auth, OPA · least privilege): Decides what data this request may reach on a least-privilege basis. The assistant can only touch the datasets a given service needs — private records stay private by policy, not by hope.
4. PII redaction (Service, NER · masking): Detects and masks names, IDs and other personal details before the model ever sees them, so private data is never sent to the LLM. Reversible only for authorised, logged operations.
5. Retriever (Service, Hybrid search): Fetches only from vetted sources, filtered by the access engine. It cannot invent rules or reach the open internet — every fact is traceable to an approved dataset.
6. Vector store (Data, pgvector): Embeddings of the approved policy corpus for semantic recall — finds the passage that answers the question even when the wording differs.
7. Approved gov datasets (Data, Policies · registries): The vetted source of truth: government policies, fee schedules, process rules and registries. The only place answers may come from — no guessing.
8. LLM answerer (Model / AI, Grounded prompt): Composes an answer strictly from the retrieved, redacted context. Constrained to say 'I don't know' and escalate rather than fabricate a rule.
9. Cited answer (Client, Citation layer): The reply links back to the exact policy or dataset it came from, so a citizen or officer can verify the source rather than take it on faith.
10. Audit log (Data, Tamper-evident): Writes every query, each data access and every answer to an append-only, tamper-evident log — so who asked what, which data was touched, and what was answered is always reviewable.
11. Human escalation (External, Named official): Anything sensitive, low-confidence or out of scope is routed to a named official instead of being answered automatically. The human keeps the judgement calls.