Tender Checklist Workbench
Review tenders against a checklist, every answer cited
Tender Checklist Workbench is a three-pane web app for reviewing a source PDF (tender, RFP, or contract) against a structured checklist. The document renders in a PDF.js viewer on the left, a checklist of questions sits in the middle with each answer carrying page-reference chips that jump to the exact cited page, and an rc-agents AI assistant on the right answers grounded entirely on that document. Click 'Answer all' and the agent fills every question with a verdict, finding, and clickable citations back to the page. Documents are ingested through the RapidCloud document pipeline, so the agent reasons over the real parsed text, not a guess.
Read less. Verify more.
Every tender hides the answers in its fine print. Tender Checklist Workbench surfaces them. Load a PDF, attach your checklist, and let the grounded AI assistant answer each question against the document — complete with page citations you can click.
- Three panes, one screen. Document, checklist, and assistant side by side.
- Citations that move. Each answer carries page-reference chips; click one and the viewer jumps to that page.
- Templates that scale. Save your review once, reuse it on every tender.
- Grounded, not guessed. Answers come from the parsed document, with nothing made up.
EU-hosted, private, and gated behind your RapidCloud portal login. No data leaves the platform.
Under the hood▾
The app is a three-pane SPA over the RapidCloud document pipeline. Left pane: a PDF.js 4.x viewer streaming the source document, with programmatic page navigation driven by citation chips. Middle pane: a checklist whose items each hold a question, answer, found-status badge, and refs (label + page) rendered as chips that scroll and flash the cited PDF page. Right pane: an rc-agents assistant (tender-checklist) grounded per-thread on the docling content artifact from the pipeline. Grounding is page-annotated: document text is rebuilt from docling elements.json, inserting ===== Pagina N ===== markers so the agent cites real (p.N) references that a regex parses into clickable chips. Answer all returns a JSON array of verdict, finding, and refs written back to each item in one agent call. The backend is a thin FastAPI proxy to the orchestrator (files, PDF, templates), PostgREST (checklist CRUD), and rc-agents (chat). Auth is portal HS256 JWT with strict per-user isolation; the agent JWT carries the real user’s email, so callbacks re-resolve to the same scoped identity.
