Search, Fetch & Summarize

By the end of this page, your agent will be able to search the web, fetch and read any URL, and give you structured research summaries — on demand or on a schedule.

Time: ~20 minutes

What research mode looks like

You send:

"What are the current best practices for RAG architecture in production LLM applications? Check a few recent sources."

Your agent searches the web, fetches the top results, reads them, and gives you a synthesized summary with the key points and sources — not just a list of links.

That's the goal.

Enable the research tools

Open ~/.openclaw/openclaw.json:

{
  "tools": {
    "web_search": {
      "enabled": true,
      "maxResults": 5
    },
    "web_fetch": {
      "enabled": true,
      "maxContentLength": 50000
    }
  }
}

web_search lets your agent search the web and get a list of results with titles, URLs, and excerpts.

web_fetch lets your agent fetch the full content of a URL — articles, documentation, GitHub READMEs, PDFs, etc.

Both together give your agent the ability to do real research.

Restart

openclaw gateway restart

Test with a research prompt

In your channel (Telegram, WhatsApp, or WebChat), try:

"Research the latest approaches to vector database selection for RAG. Find 2-3 recent sources and summarize the key trade-offs."

Your agent will:

1. Run a web search
2. Fetch the top results
3. Read and synthesize the content
4. Return a structured summary

Prompt design for research

Generic prompts get generic results. Specific prompts get useful research:

Too generic:

"Research machine learning"

Better:

"Find 3 recent articles (2024 or later) comparing transformer architectures for time series forecasting. Summarize: what approaches are compared, what the key findings are, and which one performs best on standard benchmarks."

The pattern: topic + source criteria + output format.

Format research output usefully

By default your agent will prose-summarize. For research work, you often want structure. Tell it:

Format your research as:
1. Summary (2-3 sentences)
2. Key findings (bullet points)
3. Sources (title + URL)
4. My take (your assessment of reliability/relevance)

Add this to SOUL.md under a "Research" section to make it the default for research requests.

Scheduled research digests

You can also set up automated research checks. Example — a daily industry news briefing:

{
  "automations": [
    {
      "id": "industry-news",
      "schedule": "0 8 * * 1-5",
      "prompt": "Search for the latest news about [your industry] from the last 24 hours. Summarize the 3 most relevant items. Include source name and link for each. Skip press releases and sponsored content.",
      "channel": "telegram"
    }
  ]
}

Cost discipline for research

Web search + fetch can add up. Each research task might use 2,000-8,000 tokens depending on how many sources are read.

To stay lean:

• Set "maxResults": 3 instead of 5
• Set "maxContentLength": 20000 to limit how much of each page is read
• Use a cheaper model for research tasks: "model": "claude-haiku-4-5-20251001"

Most research summaries don't require the most capable model. Haiku handles synthesis well and costs 10x less.