[khushishelat]

What this adds

A new Python recipe, python-recipes/parallel-tear-sheets. A CSV of public companies goes in, and a cited two-page PDF tear sheet per company comes out, plus a third page showing the research basis behind every field.

cd python-recipes/parallel-tear-sheets
cp .env.example .env   # add PARALLEL_API_KEY
uv run tearsheet.py companies.csv --limit 3

Page 1	Page 2	Research basis

How it works

• Research uses only the Task API. One Task Group holds every run for the CSV. Each company gets five focused core runs: profile, financials, market, news and signals. As soon as those finish, a pro thesis run starts with their facts as its input.
• Each section gets its own source policy. Financials are restricted to sec.gov plus the company's own site. News is limited to the last 90 days with after_date.
• Python does the arithmetic. Growth, margins, free cash flow, P/E, the section headlines, and the cross-checks between independently researched fields are all computed from the reported figures. No other model is involved.
• Runs are resumable. Each run ID is saved as soon as it is created, so rerunning the command never pays twice.

Sample input and output

• companies.csv: 23 deliberately different companies, including a bank, an insurer, a REIT, a utility, an airline, a pre-revenue biotech, a recent IPO, and EUR, JPY, GBP, TWD and CNY reporters.
• sample/: the NVIDIA, LVMH and JPMorgan PDFs, plus NVDA.json with the raw Task results (citations, reasoning, confidence). Verbatim page excerpts are stripped from the JSON.

Testing

• Full run on all 23 companies: 138 Task runs, 0 failures, 9.2 minutes, $5.17 at list prices.
• Output: every company returned four fiscal years in its own reporting currency, and every sheet rendered to exactly 3 pages.
• Arithmetic checks flagged three real issues:
  • Celsius's oldest EPS predates its stock split.
  • CAVA's oldest EPS predates its IPO.
  • Berkshire's dual share classes break the price × shares comparison.
• Error paths: a missing or rejected API key now exits with a one-line message, and a malformed CSV is reported clearly.
• Pre-commit: passes, including gitleaks. No keys or .env files are committed.

Notes for review

• I also registered the recipe in the root README.md (Data Enrichment table), python-recipes/README.md, and website/cookbook.json. cookbook.json validates against its schema.
• PDFs are rendered with WeasyPrint. On macOS it needs brew install pango; the script finds Homebrew's copy without any extra setup.
• Sheets use the open-source Geist fonts by default. The licensed Parallel brand fonts are not committed; TEARSHEET_FONT_DIR opts into them locally.

🤖 Generated with Claude Code