> ## Documentation Index
> Fetch the complete documentation index at: https://docs.ai-stats.phaseo.app/llms.txt
> Use this file to discover all available pages before exploring further.

# Ground follow-up extraction with web fetch

> Use `gateway:web_fetch` to pull bounded page text into the tool loop for deterministic follow-up extraction and review.

Use this recipe when a workflow needs one page fetched and turned into bounded text before the model performs extraction, summarization, or classification.

## 1. Use `gateway:web_fetch` for one-page grounding

`gateway:web_fetch` is a gateway-managed server tool.

Use it when you need:

* one URL fetched directly by the gateway
* bounded text returned to the tool loop
* request-level usage tracking
* fetched-page visibility in request logs

This is a better fit than broad search when you already know the URL.

## 2. Keep the request explicit

Typical pattern:

1. ask the model to call `gateway:web_fetch`
2. let the gateway fetch and normalize the page
3. let the model continue with extraction against the returned text

Keep the prompt explicit about the final task after fetch:

* extract fields
* summarize one section
* classify the document
* validate one claim

## 3. Bound fetch size intentionally

Use `max_chars` when the default bound is too large or too small for the task.

Recommended pattern:

* smaller bounds for title/summary extraction
* larger bounds for structured field extraction

If the page is very large, prefer a narrower extraction task instead of unbounded fetch size.

## 4. Validate logs, not just the final answer

The request details should show:

* fetched URL
* final URL after redirects
* page title when available
* HTTP status
* content type
* returned character count
* truncation flag

Use that view to confirm the model grounded on the page you expected.

## 5. Know the tool limits

`gateway:web_fetch` is intentionally bounded.

It is good for:

* HTML pages
* plain text
* light JSON or XML responses

It is not the right tool for:

* binary files
* large crawls
* JavaScript-only rendering checks

For browser-rendered verification, use a browser workflow instead of relying on fetched text alone.

## 6. Recommended rollout pattern

1. start with one preset or one endpoint path
2. inspect request logs for fetch metadata and truncation
3. tighten prompt scope if the fetched content is too broad
4. widen only after the grounded extraction path is stable

## Related

* [Debug web search requests](./web-search-debugging.mdx)
* [Tool calling](../guides/tool-calling.mdx)
* [Use response healing for structured JSON](./response-healing-for-structured-json.mdx)