search-web

Use tvly search with --depth advanced and --chunks-per-source to get query-directed passages with URLs, or Exa web_search_advanced_exa with enableHighlights for excerpts. Keep passages with their URLs during selection so every claim can be cited.

Both provide query-directed passages: Tavily content and Exa highlights. Use Tavily for news topics, site mapping, crawling, and research runs; use Exa for category-filtered search such as PDFs and for agent_run enrichment with structured output schemas.

Download the PDF with curl and use PyMuPDF via uv run to locate pages containing a heading, then extract table rows with find_tables on the selected page. PDF page numbers differ from printed labels, so inspect rendered pages when extraction omits figures or structure.

Use tvly extract with --extract-depth advanced for complex markup, tables, or JavaScript-rendered pages. For query-directed extraction across up to 20 URLs, add --query with --chunks-per-source between 1 and 5.

Missing search results leave a claim unresolved because coverage is incomplete. Compose another index for missing evidence, read complete source passages when code or table cells decide a claim, and note that failed_results records extraction failures rather than page absence.

Use research when discovery and synthesis require repeated searches across many sources. Read selected sources directly when individual pages already answer the question, and confirm research output claims against registries, release tags, and repository APIs.