From scattered pages to a data table with TinySet

TinyResearch and TinySet tools from TinyFish produce sourced reports or data tables from the public web.

Reciting AI conferences with their dates and prices does not require the same work as investigating the outlook for a market. In the first case, the same information must be found across many pages. In the second, sources must be cross-referenced and an analysis built. TinyFish’s new offering distinguishes these two tasks with TinySet for datasets and TinyResearch for reports.

Both products are available in the TinyFish interface and through integrations announced for ChatGPT and Claude, as well as via API and MCP for agents and applications. The launch price is $2 per completed Deep report and $3 in Max mode. Data collection is billed at $0.10 per row. An $8 credit is announced for new accounts. Choosing columns before starting collection

Preparing a dataset begins with defining what a row represents: a company, an event, a product, or a project. The system then suggests columns that the user can modify before the search. They can specify the expected fields, indicate where certain values should be found, and apply a yes/no filter.

TinySet’s operation then splits the search into subsets, for example cities or categories. Discovering the items and collecting their characteristics are separated. A row that requires additional research gets its own agent, so that information accumulated on previous rows does not fill its working context.

A control described in the technical post also rejects a source if the agent responsible for the row has not actually opened the corresponding page. This rule verifies that the source was consulted, without guaranteeing on its own that the extracted value is correct or current. Columns containing source addresses and dates can be requested to make verification easier.

Rows appear progressively, and the result can be shared read-only or retrieved in other tools. The documentation specifies, however, that collection is not a ranking: asking for ten institutions does not guarantee getting the ten best. To compare them, criteria must be available in the columns, such as a price or a score. Building a report around a question

Deep research follows a different process. A question is broken down into points to examine, then the system looks for sources and extracts the claims they support. It re-evaluates the information gathered and launches new searches when some parts remain insufficiently covered. The report is then written from the collected elements.

The planned presentation includes an initial summary, comparison tables, and sections with references. The user can guide the sources, request scientific publications or news articles, and set a time period. The investigation nevertheless keeps a budget of time and steps: it produces its result when it deems the question sufficiently covered or reaches its execution limits.

A Deep report generally takes five to twenty minutes depending on the company, with up to five reports run simultaneously. The results are saved in the dashboard. An examples gallery presents the deliverables and the instructions used, notably on data centers under construction, humanoid robotics, or startup shutdowns.

Evaluations conducted by TinyFish also provide indications about the limits of the promise. On ResearchRubrics, the team reports 77.10 with a scoring method that gives credit to partially met requirements, versus 71.04 if only fully satisfied requirements count. In its technical post, it also notes that a scientific report presented uses publications after the requested deadline. The presence of citations therefore is not enough to establish that all instructions were followed. Data limited to the public web and to their collection date

Both tools work on public pages. Content behind a login or paywall is not part of the announced scope. For TinySet, a collection is limited to 250 rows at launch and may return fewer if the requested information is not found. At the displayed rate, 250 rows represent $25.

The dataset remains a snapshot of the moment it was built. It does not update automatically when a price changes, a conference shifts its date, or a project evolves. A new collection is needed to obtain more recent data.

The dataset management documentation distinguishes adding rows from a full rebuild. In the first case, only the new rows are billed. In the second, the table is cleared, then each row is searched and billed again.