Blog
18 June 2026Product5 min read

Grounded answers, verifiable citations

OPSQAI refuses to guess. This post walks through how the retrieval pipeline forces every AI answer to point back at a specific paragraph in a specific SOP.

By OPSQAI Product

The default failure mode of chat AI is confident nonsense. In a warehouse or a dispatcher's office that is not a UX problem — it is a safety problem. OPSQAI is built so a wrong answer is impossible to hide.

The retrieval contract

Every user question goes through the local Worker, which embeds the question, retrieves the top-N candidate chunks from the customer's own PostgreSQL + pgvector store, and hands the model a system prompt with a single non-negotiable rule: answer only from the retrieved context; if the retrieved context is insufficient, refuse.

Citations are structural, not decorative

  • Every response comes with the source document, section, and a verbatim excerpt.
  • The UI links the citation to the exact paragraph in the ingested SOP.
  • The audit log stores the question, the retrieved chunk IDs, and the final answer — reviewable by an administrator months later.

Refusals are a feature

When retrieval returns nothing above the similarity threshold, OPSQAI does not hallucinate a plausible answer. It opens an internal request instead, which a manager can turn into a new SOP or FAQ. The knowledge base grows every time the AI does not know something — which is exactly how a knowledge base should grow.