web-search

Web Search

Use web_search for research and page extraction. Powered by OpenAI web search (Bing-backed): search + extraction + summarization in one call.

Key Insight

The query is a prompt to an LLM with web access. Write instructions, not bare keywords.

✗ "Snowflake pricing"
✓ "List all pricing tiers from https://snowflake.com/pricing in a table with features"

Parameters

• search_context_size: low — quick facts
• search_context_size: medium — general research (default)
• search_context_size: high — URL extraction, deep analysis

Use high whenever a specific URL is in the query.

Query Patterns

URL extraction — read a specific page and structure the output:

"Extract all items from https://example.com/page and list each with details in a table"

Structured research — request comparison tables, bullet summaries:

"Compare {A} vs {B} vs {C} covering pricing, features, and target audience"

Search operators — "exact phrase", site:domain.com, AND/OR, -exclude:

site:reddit.com "nextjs" vs "remix" experience
"{company}" AND ("funding" OR "revenue") 2025 2026

Translation — read non-English pages:

"Translate and summarize: https://example.co.jp/news/"

API endpoints — read JSON responses:

"Describe the JSON structure of https://api.github.com/repos/{owner}/{repo}"

Strengths

• Bypasses bot protection (Reddit, G2, Gartner) via Bing's index
• Extracts structured data (pricing, jobs, reviews, changelogs)
• Reads and translates any language
• Supports search operators (site:, AND/OR, "quoted")

Limitations

• No verbatim full-text — summarizes copyrighted content instead of reproducing it
• 24-48h recency lag — not suitable for real-time monitoring
• No PDF internals — filetype:pdf is unreliable
• No auth pages — cannot access login-required content
• Paywalls — reads public portions only

Workarounds for Limitations

When web_search falls short, suggest these alternatives to the user:

• Real-time monitoring → RSS feeds (many sites expose /feed or /rss)
• Raw HTML / interactive pages → Playwright (headless browser; click, scroll, screenshot)
• Full-text extraction → Playwright to fetch raw HTML, then parse locally

Examples

Example: Extract page content

Input: "What jobs are open at Company X?" Action: "List all job openings from https://company-x.com/careers/ with title and location" (high)

Example: Research with multiple angles

Input: "Research Company X's market position" Action: Run in parallel:

1. "Company X market position analyst reports 2026" (medium)
2. "Extract key metrics from https://company-x.com/about" (high)
3. site:reddit.com "Company X" opinions experience (medium)

Example: Translate foreign page

Input: "What does this Japanese press release say?" Action: "Translate to English and summarize: https://example.co.jp/press/" (high)