Frequently asked questions

Datool helps teams inspect AI traces, build evaluation datasets, and measure quality with scorers. Follow a request from input to output and see where it needs work.

Use recorded traces or dataset cases to evaluate your AI workflows. Combine code-based and LLM scorers with human reviews, then compare results across evaluation runs.

On WhatsApp, in a chat widget on your website and inside your product. Your team works the same conversations from one inbox and can take over whenever a person should answer.

Add your catalog, help articles and policies as knowledge, and give the agent actions such as checking an order or opening a ticket. Guardrails and tests keep its answers on track before and after launch.

Sign in with your approved Google account, select a workspace and project, and follow the project setup instructions to connect your application.

Yes. Connect any AI application with the Datool SDK or OpenTelemetry to trace its requests, evaluate real examples and measure quality. You don't need to build your agents in Datool.

Use the Datool SDK to record workflows, model calls, and tool calls, or add Datool's span processor to your existing OpenTelemetry setup. You can also connect an application for dataset evaluations without the tracing SDK. Instrumenting the steps inside it gives you the detailed trace.

Follow a request through its recorded steps, with inputs, outputs, timing, errors, and scores in context. For instrumented model calls, Datool also shows reported token usage and cost estimates when usage and pricing information are available. Related traces can be grouped into sessions.

Yes. Save useful examples from recorded traces into a dataset, add reference answers where needed, and use those cases in evaluation runs. That unexpected answer becomes a repeatable check for the next version of your workflow.

Choose a built-in evaluator, write a JavaScript or Python scorer for a precise rule, or use an LLM scorer with your own rubric. Test the scorer on representative examples before running it across a dataset. Human reviews can add ratings and comments alongside automated judgments.

Yes. Run the same cases against different application versions or prompt and model settings supported by your app. Compare scores and individual outputs to see which cases improved or regressed. You can also re-score saved outputs with a different scorer without calling the application again.

Yes. Create a review session from selected traces, define the ratings that matter to your team, and inspect each answer in context. Reviewers can save scores, leave comments, and annotate specific parts of the input or output. Those judgments complement your automated evaluations.

Your business's own WhatsApp number, connected in Datool. Customers keep messaging the number they already know.

Add the chat widget to your site or link to the hosted help page. You choose its position and appearance.

You connect your catalog as a data source. It recommends only from what you've connected, so it never suggests something you don't sell.

It sends checkout links from your store, and it can call your own APIs for things like order lookups.

From the help articles, policies and data sources you connect. It can also call your own APIs to look up orders or accounts.

It hands the conversation to your team in the inbox, with a summary and everything it already collected.

No. It takes the repetitive conversations so your team spends its time on the ones that need a person.

Yes. You write its instructions, choose its data sources and set guardrails for topics it should avoid. You can test it on real conversations before it goes live.

Request a demo on WhatsApp. We set it up with your own content so you can see real answers before you decide.

You add the chat widget to your app, then define the actions it can run there and the APIs it can call.

It can run client-side actions you define, such as opening a page or highlighting an element, and call your APIs.

Yes, through the APIs you connect. It looks up the account it's helping, so answers fit that user instead of a generic article.