Offer Compare — commercial offer comparison (AI-First Product Builder take-home)

Live demo: https://test-codebridge-offert-compare.vercel.app
Repo: https://github.com/untxia/test-codebridge-offert-compare

Upload an original commercial offer and its revision (two text documents — PDF, Word, Excel, OpenDocument, CSV, HTML, JSON, text/Markdown; formats can be mixed). The app lists the substantive changes (scope, quantities, unit prices, totals, delivery dates), each with a clickable reference to both locations (PDF: page + highlighted area; Excel: cell reference such as Offre!A7:F7; Word/HTML: table + row; text/CSV: line; JSON: item) and a side-by-side viewer. No line item is retyped.

• Matches lines despite renaming and reordering (renames/reorders are shown, never reported as commercial changes).
• Separates confirmed changes from uncertain matches (split/merge, rename-or-replacement) that it asks you to settle; your answer recomputes the result.
• Recomputes totals deterministically (Decimal): qty × unit price, sum of lines, VAT, total incl. VAT. A stated total that disagrees with its own lines is flagged, never silently corrected.
• Declines to conclude when it should: scanned/no-text PDF or image (no OCR), no recognisable line table, mixed or different currencies.
• Interface in French and English. Nothing is stored server-side (stateless API, temp files deleted).

Run locally

python3 -m venv .venv && . .venv/bin/activate
pip install -r requirements.txt
python3 -m uvicorn server:app --port 8000      # http://localhost:8000

Use the four example links on the page (normal case, formatting only, ambiguous case, scanned document) or upload your own files (.pdf .docx .xlsx .ods .odt .csv .tsv .html .json .txt .md). "Excel vs Word" compares a spreadsheet with a Word document. Legacy .xls/.doc/.rtf, images and scans are refused with an explicit message. Dev extras (regenerate the test PDFs): pip install -r requirements-dev.txt.

CLI: python3 -m offercompare compare original.pdf revised.pdf (JSON report).

Tests and evaluation

python3 -m unittest discover -s tests -t .      # 23 unit/integration tests
python3 testset/verify_testset.py               # independent check of the test PDFs against their answer key
python3 evaluation/evaluate.py                  # missed / false changes, source references, timing (writes evaluation/results.json)
python3 testset/generate_formats.py             # same offers as xlsx/docx/ods/csv/html/md/txt/json (cross-format tests)
python3 testset/generate_testset.py             # regenerate the 5 sample PDFs + expected_differences.json
python3 evaluation/holdout/make_holdout.py      # regenerate the hold-out pairs (other vendor, EN, USD, other column order)

How it works (no LLM in the loop)

1. Extraction (offercompare/extract.py, pdfplumber; other formats are read by offercompare/readers.py into cell grids — openpyxl for Excel, stdlib XML/HTML/CSV/JSON parsers for the rest — then go through the same header-keyword rules, so a PDF can be compared with an Excel file; Excel formulas without a saved value are ignored and flagged, never recomputed): columns are recognised by header keywords (FR/EN), never by position; two strategies (ruled tables, then header-anchored word alignment for tables without borders); wrapped labels are re-attached; every row, cell and total keeps its page and bounding box. Numeric dates: day-first/month-first decided by the document itself, otherwise by currency, with a warning.
2. Matching (match.py): score = 0.55·label similarity + 0.35·terms (price/qty/date) + 0.10·position. A pair is confirmed only if the label is near-identical or renamed with a high global score, and in both cases with a clear margin over the runner-up. Plausible but unconfirmed pairs form connected groups → uncertain (one-to-one, split/merge, many-to-many); the rest are removed/added. Thresholds are constants at the top of the file.
3. Diff and report (diff.py): change list, arithmetic checks, net effect (stated vs recomputed, amount left unattributed by uncertain groups), decision: conclude, no_changes, conclude_partially_and_ask, decline.
4. UI (public/): vanilla JS; pdf.js (vendored, Apache-2.0) for the side-by-side viewer with highlights; all text inserted with textContent. Look and feel: Tailwind CSS (compiled to public/vendor/tailwind.css), inertial smooth scrolling (Lenis), a three.js hero scene (two floating sheets linked by the changed line, bundled to public/vendor/hero.js) and a floating 3D viewer window; both respect prefers-reduced-motion. Fonts (Bricolage Grotesque, IBM Plex Sans, OFL) are self-hosted. To rebuild the assets: cd frontend && npm install && npm run build (the built files are committed, no build step on Vercel).

Optional AI assistant (explains the report, never computes it)

The comparison itself stays deterministic. On top of it, an optional assistant answers questions about the already computed report ("why doesn't the total add up?", "draft an email to the supplier"). It is off unless ANTHROPIC_API_KEY is set (Vercel → Settings → Environment Variables; optional ANTHROPIC_MODEL, default claude-haiku-4-5-20251001). Without the key the panel is simply hidden.

Guard-rails (offercompare/agent.py, tests/test_agent.py): the model only receives the report (geometry stripped), never the files; the system prompt forbids recomputing or "correcting" amounts and requires source citations; document excerpts are declared as data, not instructions; every figure in the answer is searched in the report and unmatched figures are flagged to the user. Privacy note: when used, the report (not the files) is sent to the Anthropic API.

Deploy (Vercel)

vercel.json + requirements.txt + api/index.py (FastAPI) + public/ (static). vercel --prod from this folder, or import the GitHub repo in Vercel with the Other preset (the config is in vercel.json).

Reused components / limits

pdfplumber (pdfminer.six), openpyxl, FastAPI, pdf.js, reportlab/Pillow (test data only). Everything else is written for this test. Limits: text documents only (no OCR — scans and images are declined on purpose; legacy .xls/.doc refused), FR/EN column and total labels, one currency per document, layouts tested on the included samples only (see DELIVERY_NOTES.md), 2 MB per file.