| name | web-research |
| description | Search and extraction strategy. Teaches the researcher which of the 8 web tools to use at each stage of a research cycle. |
| agents | researcher, cognitive |
Web research tool selection
You have 8 web tools arranged in tiers. Pick the right tool for the task:
Tier 1 — Brave Search (requires BRAVE_SEARCH_API_KEY or BRAVE_ANSWERS_API_KEY)
| Tool | Best for | Notes |
|---|
| brave_answer | Self-contained factual questions ("what is X", "when did Y") | Fastest, cheapest for simple facts |
| brave_llm_context | Default hot path — need raw context to reason over | Machine-optimised search results |
| brave_web_search | Broad discovery with titles, URLs, and snippets | Good for finding multiple sources |
| brave_news_search | Time-sensitive queries, current events | Returns recent news articles |
| brave_summarizer | Need citations + follow-up threads | Search + summary chain |
Tier 2 — Jina Reader (requires JINA_READER_API_KEY)
- jina_reader: Extract clean markdown from a URL. Better than fetch_url for content extraction.
Tier 3 — Fallback (no API keys needed)
- web_search: DuckDuckGo keyword search. Use when Brave keys are unavailable.
- fetch_url: Raw HTTP GET. Use when Jina key is unavailable, or for non-HTML content.
Decision tree
Factual, self-contained question? → brave_answer
Need raw context to reason over? → brave_llm_context (default)
Broad discovery with snippets? → brave_web_search
Time-sensitive / news? → brave_news_search
Need citations + follow-up threads? → brave_summarizer
Have a URL, need full content? → jina_reader (primary), fetch_url (fallback)
No API keys available? → web_search (DuckDuckGo fallback)
Quality signals
After extraction, check:
- Does the content have a publication date? Note it.
- Is the source primary (official, authored) or secondary (aggregator)?
- Does the content contradict earlier search results? Flag it explicitly.
Prefer primary sources. When a snippet and a full extraction conflict, trust the full extraction.
