TinyResearch + TinySet - Delegate the deep web work: cited reports and real datasets

Hi Product Hunt, Lily from TinyFish. Today we're launching two solutions that came out of watching our own team do the same two jobs by hand for customers, over and over: a proper research report on a market or a company, and a dataset that didn't exist anywhere in one piece.

TinyResearch takes one hard question and runs the whole investigation. It breaks the question into the smaller questions inside it, searches wide, ranks sources by how credible they are for that specific question, reads the best ones live, pulls the specific claims off each page, and goes back for another round wherever there's a gap. The report is written only from those claims, so there's no filler. And it's shaped for a decision: a summary up top, comparison tables, clear headlines, numbered citations. You can take it into a meeting without rewriting it. A Deep report usually takes 5 to 20 minutes, you can run five at once, and you can keep it to the sources you trust, block the ones you don't, or limit it to news, research papers or a date window.

TinySet is for when you already know what you need: the same facts across a lot of companies, places, projects or products. Describe the dataset in one sentence, review the columns it proposes, and it finds the matching entities and fills the rows from public web pages while you watch. Every row gets its own agent, so row 200 gets the same care as row 2. A cell is only filled from a page the agent actually opened, and that's checked in code, not just asked for in a prompt. If nothing backs a value, the cell stays empty. Up to 250 rows per run, export to CSV or Excel, or share a read-only link.

We ran TinyResearch on three public research benchmarks with their own tasks and scoring: 51.74 on DRACO, 77.10 compliance on ResearchRubrics, 46.42 on DeepResearch Bench II. The full setup is in the technical blog.

Both run on the same TinySearch and TinyFetch underneath, and both work in the TinyFish playground, in ChatGPT through our plugin, in Claude through our connector, and through the API, MCP, CLI and SDKs.

The best way to see them is https://www.tinyfish.ai/deep-work, where we put 30+ real reports and datasets with the exact prompts: where the AI data-center buildout is heading, which humanoid robotics companies are really ahead, why AI startups are shutting down, what YC is betting on right now.

What would help us most: the question you'd never trust an AI to research, and the dataset you've built by hand more than once. Drop them below and we'll run them in the thread!

Manav, who built TinyResearch, and Simantak, who built TinySet, are here all day.

Shuhao from TinyFish here. I talk to customers every day and the thing that made these 2 feel inevitable is how often I watched our own team hand-build them. A prospect would ask “can you get us every X with Y and Z” and someone would spend a weekend stitching it together from TinySearch and TinyFetch. Or they’d ask for a proper read on a market before a decision and we’d write it up by hand. TinySet and TinyResearch are those two jobs, productized. If you’ve got a dataset you’ve been building in a spreadsheet for weeks, drop the prompt here and I’ll run it for you!

Sudheesh from TinyFish. Here's what's underneath. Both products run on the same TinySearch and TinyFetch that agents already call tens of millions of times a month, so when TinyResearch reads a page, it reads the live version, fetched the moment you asked. Driving all of it is Mako, our own model, trained on billions of page trajectories from agents working the real web. That training buys accuracy, and accuracy is what brings cost down, because an agent that gets the page right the first time skips the retries and re-fetches that make web work expensive.

Neither product can write anything it didn’t pull from a real page. TinyResearch extracts the claims first and writes only from those, and TinySet rejects any cell unless the agent actually opened the page it cites. We enforce that rule in code, where no prompt can talk its way around it. Happy to answer anything about how they fit with the rest of the stack.

Thanks Lily! I’m Manav, and I’ve been building TinyResearch at TinyFish over the past few months.

The idea came from trying to get research done with general-purpose models and coding agents. The answers were often too shallow. Pushing for more depth ate through my context window and usage limits, and I still wasn’t happy with the results.

I now use TinyResearch for my own research, both directly and through my agent. I’m here today and happy to talk about how it works or help you try it. If you have a question you’ve struggled to get a good answer to, drop it here. I’d love to run it and hear whether the report was useful.