Live demo Firecrawl API Workers AI RAG

Firecrawl vs raw fetch

A side-by-side in the Firecrawl problem space: the same URL pulled two ways, a bare edge fetch versus Firecrawl clean markdown, then the grounded answer each one produces. Clean input, better answer.

Live demo

Clean input is the whole game

Same URL, two extractors. A bare edge fetch keeps all the nav, footer, and boilerplate; Firecrawl returns just the main content as clean markdown. Then ask a question against each and watch the answer quality follow the input quality.

Native always runs at the edge. The Firecrawl side needs a FIRECRAWL_API_KEY; without one it shows the gap you would close by using the real product.

How it works

One URL, two extractors, then a grounded answer from each. The page is fetched two ways in parallel, both outputs are measured, and Workers AI answers the same question against each — so you can watch input quality drive output quality.

🔗

URL

any page

submit

→

⚡

Edge fetch + Firecrawl

in parallel

two markdowns

→

🧠

Workers AI

grounds an answer

→

Raw → noisy markdown

Firecrawl → clean markdown

Key decisions

Both in parallel

Promise.allSettled runs both extractors at once, so one failing never sinks the comparison.

SSRF-guarded

Blocks localhost, private, and metadata IPs; every fetch is size- and time-capped.

Honest delta

Character counts are taken on full length before the 12k display cap, so the boilerplate-reduction badge stays true on large pages.

Grounded, not freeform

Each answer is generated only from its own extraction, so answer quality tracks input quality.

Built with:Firecrawl /v1/scrapeWorkers AI (Llama 3.3 70B)Pages Functions

Reference

Firecrawl docs

Why I built it

The hard part of any "chat with a website" feature is the boring part: getting clean, structured content out of an arbitrary page. So I run both extractors on the same URL and put them side by side. The bare edge fetch keeps every nav link, footer, and cookie banner; Firecrawl returns just the main content. Ask a question against each and the answer quality tracks the input quality, which is the case for paying for Firecrawl. Both extractors are SSRF-guarded and size and time capped.