SIEVE

New: the SIEVE API waitlist is open

Search your LLM can afford to use

SIEVE reads the web, throws out the junk and checks the sources against each other. Your model gets one exact quote and its evidence, not ten pages. Fewer tokens, better answers, and the cheapest AI search we can build.

10 free searches · no card

Building with LLMs? Join the API waitlist
~150
tokens in your context, median
~8 s
median time to a verdict
0
words written by a model
4
clear verdicts, incl. “Nothing found”

How it works

An investigation, not a list of links

Every search runs the same four moves, and you can watch each one.

  1. 1

    Scout

    Reads your question, works out what kind it is, and searches the right places: the web, Wikipedia, PubMed, Stack Overflow or the news.

  2. 2

    Read and filter

    Opens the pages and throws out ads, clickbait, off-topic and outdated ones before anything reaches your model.

  3. 3

    Cross-examine

    A skeptic checks whether the good sources actually agree, so a disagreement never gets hidden.

  4. 4

    Judge and quote

    Picks the winner and copies the exact sentence that answers you. Nothing is rewritten, so nothing can be invented.

Answers that fit the question

The right shape for every answer

A number for a number. A tally for a yes or no. A timeline for a history. Pick one and watch it build.

you askhow tall is mount everest

0.00 m

= 29,031.7 ft · units converted by code, not by a model

3 sources agree on this figure

1 older page said 8,848 m and was marked as older

Illustrative examples of the shapes SIEVE can return. Every piece is exact text from a source.

Less to read, less to pay for

Your model reads 150 tokens, not 14,000

Drag the sliders to match your own setup.

Set your own numbers

Yours are assumptions you can change. The 150 is measured: the median size of SIEVE's answer plus its supporting quote across 22 recorded searches, at about 4 characters per token. Real searches vary. This compares what enters your model's context. SIEVE does the reading and filtering on its own.

Tokens your model reads, per search

Reading the pages14,000
Reading a SIEVE verdict150

93×

fewer tokens in your context

At 1,000 searches a day

14M → 150K tokens

Built for LLM apps

Drop it in wherever your model needs the web

Tap a card to see what comes back.

  • Research agents

    “Who founded this company, and what did it launch this year?”

    Without it: Ten pages stuffed into the prompt, and the agent has to decide which to believe.

  • Support and FAQ bots

    “Does this phone support that charging speed?”

    Without it: The bot answers from memory and is confidently wrong.

  • RAG fallback

    “A question your own documents cannot answer.”

    Without it: Your index comes up empty and the model guesses.

  • Coding assistants

    “How do I read a file line by line in Go?”

    Without it: A pile of blog posts with half-working snippets.

  • Market and sales research

    “What is this company known for, and what is the latest news?”

    Without it: A day of tabs, or an agent that reads them all and bills you for it.

  • Fact-checking

    “Is this claim true?”

    Without it: A search list that leaves the reading to you.

Compared

Most search APIs hand you the homework

Here is how SIEVE differs from a typical search API.

What your LLM receivesLinks and snippets, or whole pagesThe answer, the exact quote and ranked sources
Tokens in your contextThousands per searchAbout 150 (measured median)
Who decides what to trustYour model, on every single callSIEVE, before your model sees anything
Text a model made upPossible whenever it summarisesNone. Quotes are copied, never written
When sources disagreeYou find out by reading them allMarked “Disputed”, with both sides
When there is no answerFiller results that look like answersSays “Nothing found” plainly

Nothing made up

Quotes are copied from the page, so your agent cannot pass on an invention.

Honest about gaps

“Nothing found” and “Disputed” are real answers your app can act on.

Fast enough to use

A verdict in about 8 seconds at the median, with every step visible.

Questions

Good to know before you try it

Is this a chatbot?

No. SIEVE is a search tool. It investigates the web and returns a verdict with sources. Nothing it shows you was written by a model.

How does it save my LLM tokens?

Your model reads a short answer and its quote instead of ten pages. SIEVE does the reading and filtering itself. The API is not open yet, so the savings shown here compare what would enter your model's context.

What does it cost?

The website gives you 10 free searches per Google account. API pricing is not announced. Our aim is to make it the cheapest way to give an LLM web search, and waitlist members hear first.

Why do I need to sign in?

Search costs us real money each time, so the free searches are counted per Google account. We only read your name, email and photo. See the privacy policy.

Can it be wrong?

Yes, like any search. What SIEVE does differently is show its evidence: the quote, the sources and how they agree, and it says when it is not sure.

Watch it work on your question

Sign in with Google and get 10 free investigations. Then join the waitlist to put it inside your own LLM app.

Join the API waitlist

We only read your name, email and photo. Privacy policy