Build a Dataset with AI Using TinySet (Step-by-Step Guide)
Some lists don't exist on a single page. Every independent bookstore in Portland is scattered across a city guide, a few directories and each store's own site, and putting it together means an afternoon of copy-paste or a scraper per site.
TinySet is an AI dataset builder that turns a single sentence into a structured dataset using information collected from multiple websites. Every value is backed by a source, so you can see where the information came from. Powered by TinySearch, TinyFetch, and TinyFish's web agents, TinySet is available as Datasets in the TinyFish Playground and is currently rolling out in beta at agent.tinyfish.ai.
What can you build with TinySet?
TinySet is useful when the information you need is scattered across multiple websites rather than available in one place. It brings those details together into a structured dataset, with sources you can trace back to.
- Target lists: "All YC W25 companies with founders, website and funding"
- Places in an area: every dealership in a state
- The same facts for a known group: "Ivy League universities with their city and website" gets you one row per school
When to use something else
| You want | Use |
|---|---|
| Every row from one site, like one directory | TinyFetch on the pages directly |
| One site, but behind filters or "load more" buttons | TinyAgent |
| A written answer, not a table | TinyResearch, a report with citations |
Not sure which TinyFish product fits your use case? Our guide to choosing the right TinyFish API breaks down when to use each tool, from extracting web data to automating browser workflows and conducting research.
How to build a dataset with AI using TinySet
- Open Playground → Datasets. If it's not in your sidebar, the beta hasn't reached your account yet.
- Under "What dataset would you like?", write one sentence that names the thing, the scope and the fields.
- Click Build or press Cmd/Ctrl+Enter.
💡"Coffee shops" is too vague to plan. "Coffee shops in Austin with their address and website" gives the agents something concrete to search for.
⚠️ Stick to English for now. Other languages aren't reliable yet.
Planning takes a couple of minutes. A live feed shows every search and every page the agents read. If it heads somewhere you didn't mean, press Esc or click Cancel and rephrase.
Review your dataset plan and sources
Nothing gets built until you've seen the plan and agreed to it. The whole product is built around this step.
- Columns: each has a type (text, URL, date, number or true/false), and one is marked as the column that identifies a row.
- Seen on: under each column, the pages where the agents actually found that value. If the only source is a page you don't trust, you find out now, not after the build.
- Partitions: the listing pages the rows will come from, each with the row names found there. Remove any you don't want.
- Sites that refused the agents show up struck through, and a summary line says how many searches ran and how many pages were checked.
You can rename the dataset and its columns, change types, and edit the short guide that tells agents what to look for in each column. If the plan is off, click Start over. Otherwise click Build dataset.
How TinySet builds your dataset
After Build dataset you land on the dataset's page. Agents take each row from the partitions, read pages about it and fill in the columns in parallel. How long that takes depends on how many rows and columns you asked for.
A value only goes in if it came from a page an agent read during that build. If no page backs it, the cell stays empty. We'd rather show you a gap than a guess.
Everything you build is listed under Your datasets on the Datasets page. Datasets are private to you and aren't shared with teammates.
What it doesn't do yet
- Public web only: nothing behind logins or paywalls, and a business with no web presence won't show up.
- Row cap per build during the beta (the final number isn't set), and expect daily usage limits too. It's meant for focused lists, not full-web crawls.
- Browser only: no API, SDK, MCP tool, CLI or CSV export yet.
- No new columns after a build.
- Pricing hasn't been announced. We'll post it, along with the limits, in the docs.
⚠️ Datasets are snapshots, not feeds. There's no refresh schedule, so build it again when you want newer data.
How it compares: TinySet vs. other AI dataset builders
- Exa Websets and Parallel FindAll both have APIs and paid plans, and they're further along there. If you need list building in code today, use one of them, or build on TinySearch and TinyFetch directly.
- BigSet is our open-source dataset builder. It's a separate, self-hosted project under AGPL with its own features, including scheduled refresh. TinySet is the hosted product inside TinyFish. For a closer look at the underlying approach, read our technical breakdown of how BigSet works.
Our bet with TinySet is the plan step: you see the columns, where each one was found and which listing pages the rows come from, and fix them before anything runs.
Ready to build your first dataset? Open TinyFish Playground, select Datasets, and describe the information you want to collect in one sentence. If Datasets isn't available in your account yet, it's still rolling out in beta.
AI disclosure
Content on this website may be created or refined with the assistance of AI tools and is subject to human editorial review.
