Web extractor.
The web extractor pulls information out of messy websites and into a clean table, for example every product recall a regulator has published. It checks each value against the page it came from, so a made-up figure is caught rather than trusted.
- 01You point it at a site
- 02It looks around first
- 03It writes down a recipe
- 04Readers work in parallel
- 05Every value is checked
Step by step
What happens when you ask.
You point it at a site
And say what you want: “Every recall from this site: date, product, hazard, units and remedy.”
It looks around first
Before extracting anything it explores how the site is laid out: where the list is, how pages link, where the detail sits.
It writes down a recipe
Plain instructions for that site, so the next run is faster and gives consistent results.
Readers work in parallel
Helper agents read the detail pages at the same time, which keeps long sites quick.
Every value is checked
Each figure in the table is matched against the text of its page. Anything that is not there is flagged instead of kept.
Judgment
Where it has to think, not just look things up.
These are the moments a simple search tool gets wrong. The demo data is built to include them.
Where the information lives
Some sites use tables, some bury the facts in paragraphs, some spread them across several pages. It adapts to each.
Making it consistent
One site writes “3 March 2025”, another “2025-03-03”. It records them all the same way, so the table can be sorted and compared.
When a page is unclear
If a detail is missing or ambiguous, it leaves the cell empty and says why, rather than guessing.
Who does the work
Two agents, one answer.
The lead
A large, capable model that explores the site, writes the recipe and assembles the table.
The page readers
Smaller, faster models that each read one page and report back what it says.
Every agent here is built the same way: a leading AI model that plans the work and a smaller one that checks it, a set of tools each is allowed to use, and rules on what it must prove before it answers. They run on leading models from Anthropic or OpenAI, inside this website.
Safeguards
What keeps it honest.
- It only visits the demo websites.
- Values that cannot be found on their page are rejected.
- The recipe is shown in full, so you can see exactly how the site was read.
- Each conversation has a time limit and a spending limit.
Three invented recall-notice websites, each built differently: a table split over several pages, detail pages written as prose, and a plain-text bulletin.