AI Capture Brief with provenance
A structured two-page brief generated from the tender pack — scope, evaluation weights, eligibility traps, submission mechanics — with every claim anchored to its source paragraph, page number, and confidence score. Built for capture managers who cannot afford hallucinated facts on 8-figure bids.
Why provenance is non-negotiable
In 2025, the US Government Accountability Office dismissed more than 20 AI-drafted bid protests. The common thread: hallucinated citations — AI systems that cited clauses that did not exist in the source document. The consequences were disqualification and in some cases formal sanctions.
Capture managers at engineering and consulting firms bidding into aid-funded infrastructure work face the same risk. A brief that says "the evaluation allocates 40% to Technical Approach" is meaningless unless you can point to the exact paragraph on page 34 that says so.
TenderTism's AI Capture Brief never emits a claim without anchoring it. Every fact is retrievable, auditable, and reproducible. The brief is a tool your legal team can review — not a chatbot summary you paste into a Go/No-Go decision and hope for the best.
Sample: AI Capture Brief excerpt
Hover or tap any underlined claim to see its source. In the app, click to open the source paragraph verbatim.
AI CAPTURE BRIEF
Consultancy Services — Power Transmission Upgrade, Honiara
ADB-PNG-2026-047 · Solomon Power · ADB-financed
Brief confidence
89% — high
Scope
The assignment covers feasibility study, detailed engineering design, and construction supervision for a 33 kV substation and 45 km of transmission line in the Honiara–Lungga corridor. Construction supervision is expected to last 36 months following the 8-month design phase.
Evaluation weights
Eligibility traps
- Firms with ongoing ADB-financed contracts exceeding USD 50M are ineligible to bid. Verify your current portfolio before proceeding.
- JV lead must be from an ADB member country; nationality of sub-consultants is unrestricted.
How the brief is generated
- 1
Document ingest
Every file in the tender pack — PDFs, Word documents, spreadsheets — is ingested by LlamaParse. Tables, footnotes, and appendices are preserved. Each paragraph gets a document reference (doc name, page, paragraph, line).
- 2
Semantic chunking and embedding
The ingested text is split into overlapping semantic chunks and embedded with Cohere embed-v4 (multilingual). Embeddings are stored in Postgres + pgvector alongside the verbatim paragraph text and its document reference.
- 3
Targeted extraction with strict citation schema
Claude Sonnet runs a structured extraction prompt against the top-k retrieved chunks for each section of the brief (scope, evaluation weights, eligibility, etc.). The system prompt mandates that every extracted claim include a verbatim source paragraph reference. Hallucinations are structurally impossible — the model cannot emit a claim without the reference field being populated from the retrieval result.
- 4
Fidelity scoring
Confidence is computed from three signals: retrieval margin (gap between top-1 and top-2 similarity scores), self-consistency (claim is stable across 3 sampled extractions), and citation density (high-confidence claims have multiple corroborating passages). Claims below a confidence threshold show a "verify before acting" flag.
- 5
Provenance UI
The brief renders with every claim underlined in copper. Click any claim to open the source paragraph verbatim — the exact sentence highlighted, page reference shown, and a link to that location in the pack viewer. Nothing is paraphrased in the source display.