The new Firecrawl /search

The excerpt approach is a clean tradeoff for factual lookups where the signal is usually in one paragraph. Curious how the model handles research queries where the evidence is spread across 3-4 short paragraphs on the same page, or where two sources directly contradict, does it surface multiple excerpts per result, or one best-match per page? That gap is where we've seen agents get confidently wrong answers from snippet-based search.

Benchmark scrore is really good. What are the scenarios where it failed to answer correctly?

94.7% on SimpleQA is impressive. How does the relevance model handle ambiguous queries?

The 10x fewer tokens claim is the part that caught my attention. In agent workflows, the expensive failure is often not retrieval itself, but letting the model chew through irrelevant page structure. I’d be curious to see where the relevance model struggles most: tables, long docs, or queries with several equally plausible answers?

Congrats on your launch! 🚀

Returning the relevant excerpt instead of making every agent carry the whole page is the practical win. Long agent runs fail in boring ways when context fills with page furniture, so keeping retrieval lean is a reliability feature as much as a token-cost feature.

Returning the relevant excerpt instead of making every agent carry the whole page is the practical win. Long agent runs fail in boring ways when context fills with page furniture, so keeping retrieval lean is a reliability feature as much as a token-cost feature.

firecrawl has become the default answer every time someone in the agent space asks "how do i give my agent reliable web access." that kind of organic word of mouth in developer communities is hard to manufacture. the "interact with" part beyond just scraping is where it gets interesting for agentic use cases. filling forms, clicking through flows, handling dynamic content that only renders after javascript executes. curious how the rate limiting and retry logic works when a target site starts throttling requests mid-crawl.

That's a solid benchmark score. Does the relevance model retrain continuously or is it a fixed release?

94.7% on SimpleQA is a strong number, but that benchmark is specifically short factual questions with a fairly clean right answer, which is close to the best case for a relevance model scoring paragraphs against a query. Curious how much that accuracy holds up on queries where the right answer isn't a single fact but requires synthesizing across multiple excerpts, since scoring individual paragraphs and lists against a query is a different task than judging whether the combination of several partial excerpts actually adds up to a complete answer.

Also, returning excerpts instead of full pages means the model never sees the surrounding context that a paragraph originally sat in. If a relevant looking paragraph is actually a counterexample, an outdated caveat, or a quote the source page was explicitly refuting, does the excerpt carry any signal that it needs to be read skeptically, or does stripping the surrounding page context risk agents treating pulled quotes as more authoritative than the original author intended.

The excerpt model is a smart place to spend the effort. One thing I'd poke: it works on pages you already retrieved. For grounding, the failure I hit most isn't a found page being too long, it's the one page that actually holds the answer never making it into the result set at all.SimpleQA rewards well-covered facts. The queries that hurt are the long-tail ones where the answer sits on a single thin, badly-linked page. Excerpt selection can't rescue a page that never surfaced.Two things I'd want to know: does the new model touch recall, or only what you do with what's already retrieved? And on those thin pages, does aggressive excerpting ever drop the one passage that mattered, to save tokens?

That side-by-side cost/accuracy comparison thing is exactly the stuff that ends up getting litigated in G2/Capterra reviews later, when prospects are comparing tools and can't just ask you directly. Curious if you've found review platforms actually surface that kind of signal for you, or if it's still mostly direct sales conversations at this stage?

Really like that you explained why search was slow/expensive before showing the fix, and you backed it up with a real benchmark score (94.7%), not just 'trust us it's better.

Hey Product Hunt 👋 Eric, Caleb, and Nick from Firecrawl here.

AI agents rely on search tools to find answers on the web, but those answers are often buried in noise. Navigation, boilerplate, and tangents wrap around the excerpt that matters, so an agent processes far more content than it needs to. That makes search slower, more expensive, and more likely to return the wrong answer.

Firecrawl /search now runs every result through a new custom relevance model. It scores every paragraph, list, and table against your query, then returns the excerpts that best answer it.

We evaluated the improved /search on SimpleQA, OpenAI's factuality benchmark for short, fact-seeking questions. AI agents using Firecrawl /search achieve state-of-the-art accuracy at 94.7%, the highest score of any search provider we tested.

By returning relevant excerpts instead of full pages, each response uses 10x fewer tokens. That leaves more of your agent's context window free and makes the results cheaper and faster to reason over.

AI agents with web search keep their context lean across long chains of searches. RAG pipelines get source-linked excerpts ready for a prompt without separate chunking or reranking. Research and enrichment workflows can pull the specific fact, figure, or quote they need.

The upgrade is live across the API, SDKs, CLI, and MCP, and the API shape hasn't changed, so existing /search calls start returning better results automatically!

If you're building products with agentic search, we'd love to hear how it improves your results.