Extract links

[seanmcguire12]

why

• we want to be able to extract links from a webpage

usage

• in your zod schema, add a filed with the the type z.string().url()

example:

const extraction = await stagehand.page.extract({
  instruction: "extract the link to the 'contact us' page",
  schema: z.object({
    link: z.string().url(),
  }),
});

how it works

• if a user has defined a schema where a field type is z.string().url(), inside the extract handler, that field type is converted into z.number(), and the path to the converted field is stored
• after this type conversion, we make the inference call to extract. the LLM will return with IDs (of type number) in the converted field
• these IDs are used as keys to find the corresponding URL in the idToUrl mapping
• then, we iteratively update replace the IDs with the actual URLs, converting the types back to z.string().url() as we go

what changed

• added logic in extractHandler to recurse through the user defined schema, and transform any fields of type z.string().url() to type z.number() while recording their path within the schema
• added logic to inject the URLs into the extraction result while restoring the original types

test plan

• added evals

[changeset-bot]

🦋 Changeset detected

Latest commit: a6a1184

The changes in this PR will be included in the next version bump.

This PR includes changesets to release 1 package

Name	Type
@browserbasehq/stagehand	Minor

Not sure what this means? Click here to learn what changesets are.

Click here if you're a maintainer who wants to add another changeset to this PR

[greptile-apps]

PR Summary

This PR enhances the extraction functionality by integrating link extraction support and updating the metadata and accessibility tree processing.

• Updated /lib/inference.ts to add a new 'url_field' in the metadata schema and return it as 'urlField'.
• Modified /lib/prompt.ts to adjust extraction prompts for DOM-based link extraction and instruct returning ID-only link outputs.
• Enhanced /lib/a11y/utils.ts to remove redundant StaticText and extract URLs from node properties.
• Introduced a recursive ID-to-URL replacement in /lib/handlers/extractHandler.ts.
• Extended type definitions in /types/context.ts to support URL mapping.

_{5 file(s) reviewed, no comment(s)
Edit PR Review Bot Settings | Greptile}

[greptile-apps]

PR Summary

(updates since last review)

Added two new evaluation tasks to test URL extraction functionality and implemented the core URL extraction mechanism in the extract handler.

• Added extract_single_link.ts and extract_jfk_links.ts evaluation tasks that test extracting URLs using z.string().url() schema type
• Implemented schema transformation in extractHandler.ts that converts z.string().url() fields to z.number() during processing
• Added URL injection mechanism that replaces numeric IDs with actual URLs in the extraction results
• Modified evals.config.json to include the new evaluation tasks under the 'extract' category
• Note: URL extraction is implemented in domExtract method but not in textExtract method which is used by default

Greptile AI

_{6 file(s) reviewed, 2 comment(s)
Edit PR Review Bot Settings | Greptile}

[greptile-apps]

style: If mappedUrl is undefined, this sets the field to an empty string. Consider returning null or undefined instead, or adding a warning log when a URL mapping is missing.

[darasus]

@seanmcguire12 which models have been used to test this? I've tried o3-mini and gemini-flash but no luck