Three layers with a strict division of labour. The test for whether the line is drawn correctly: if the LLM provider went down and you swapped in a template renderer, would the answers still be factually correct? They would — you would lose fluency and translation, not truth.
Decides every fact
Services, scenarios, eligibility rules, required documents, procedure steps, fees and offices are rows. Conditions are stored as JSON expression trees and evaluated by code with three-valued logic, so "we do not know yet" is a distinct outcome from "no". Nothing here is summarised by a model.
Finds supporting text
Official prose is chunked, embedded and retrieved by hybrid search: Postgres full-text plus vector cosine, fused with reciprocal rank fusion. Every chunk carries its source title and last-verified date. Retrieval finds text; it does not decide truth, and "nothing documented" is a valid result that routes the citizen to the office.
Only phrases and translates
Four jobs: detect intent, translate, route context, and render already-decided content. It is never asked what documents a service needs — it is handed the list and asked to express it in Urdu. Every call has a required deterministic fallback, enforced by the type signature of the client.
3
Services
10
Scenarios
21
Requirements
10
Eligibility rules
15
Procedure steps
7
Exception routes
11
Offices
10
Sources
Retrieval corpus: 11 documents, 20 chunks, 20 embedded with Xenova/multilingual-e5-small at 1024 dimensions.
Every citizen-facing fact carries a verification tier and a source. The seeded knowledge base is deliberately unverified: it is structurally complete and attributed to real official pages, but no one has yet confirmed each value against the live page. Fees are stored as NULL rather than as a plausible guess.
0
verified · 0%
51
unverified · 82%
11
synthetic · 18%
Every requirement, step, rule and fee has a source. Zero orphaned facts.
Language model
groq → mock
Live. Keys: dashscope 0, groq 1.
Embeddings
local / Xenova/multilingual-e5-small @ 1024d
Multilingual semantic retrieval active.
Grounding
Strict mode on. Sources go stale after 180 days. Evidence floor 0.75 cosine.
Database
pg — PostgreSQL with pgvector.
No evaluation run recorded yet. Run npm run eval.