Tender search thatanswers with citations

Client
Halden & Roe
Sector
Professional services
Delivered
Span 02 — Halden & Roe

In one paragraph

Twelve thousand tender documents and twenty years of bid responses made answerable, with every answer pointing at the paragraph it came from.

Measured afterwards

12,000
Documents indexed
1.8s
Median answer time
5d → 1.5d
First-draft preparation
97%
Answers with verified citation

The problem

The bid team rewrote the same answers each cycle because nobody could find the previous version. Knowledge lived in the heads of four people and in a shared drive with 12,000 files and no consistent naming. Two of those four were close to retirement.

What we did

We built a retrieval system over the full corpus with paragraph-level chunking that preserves document structure, so a clause keeps its heading and its date. Answers are composed only from retrieved passages and are rendered with inline citations that open the source document at the right page. We shipped a deliberately narrow interface — one search box, one answer, the passages beside it — and ran it against 300 real questions from past bids before anyone else saw it.

What changed

Answers come back in under two seconds with citations the bid team trusts enough to paste into a submission. First-draft preparation for a mid-size tender dropped from five days to a day and a half, and institutional knowledge survived two retirements.

Built with

PythonClaude APIpgvectorPostgresReactTypeScriptRagas