extract-structured-data

# Extract Structured Data ## When to use The user has unstructured content (email, document, web page, transcript) and wants specific fields extracted as structured data. ## Steps 1. **Identify the schema** — what fields does the user want? If unclear, propose one via `ask_questions` or just `offer_choices` with common templates: - Contacts (name, email, phone, company) - Events (title, date, location, attendees) - Products (name, price, description, sku) - Tasks (title, due, assignee, priority) - Custom (ask user to list fields) 2. **Get the source content**: - For URLs: use `browser_extract` — Cloudflare Browser Rendering's `/json` endpoint runs Workers AI extraction natively, very efficient - For attachments: read the document content (already in context or via `fs_read`) - For plain text in the message: use directly 3. **Extract** — for browser-based extractions, prefer `browser_extract` because it's done server-side. For other content, extract using your own reasoning. 4. **Validate** — check the extraction: - Required fields populated? - Numbers are numbers, dates are dates? - Anything obviously wrong (e.g. email field contains a phone number)? 5. **Display** — use `show_data_table` to render the extracted records cleanly. Columns should match the schema fields. 6. **Offer next steps** via `offer_choices`: - "Save as CSV" - "Add another field" - "Re-extract with different schema" - "Looks good" ## Style - Be precise about what's missing. Use `null` or empty string consistently — pick one and document it. - Preserve original casing and formatting where possible - For dates, normalise to ISO 8601 (YYYY-MM-DD) unless told otherwise ## What not to do - Don't infer values that aren't in the source. If a field is missing, say so — don't guess. - Don't reformat numbers (e.g. "$1,234.56" — keep as-is unless asked) - Don't deduplicate without permission — the user may want all records