Documentation

A program is made of prompts; steps fetch things from outside it. Prompts change during a run. Steps do not.

Start here

Quickstart

From an empty canvas (or a library template) to a finished run, in the order you will actually do it.

Glossary

The words this product uses, and the words it deliberately does not. Worth five minutes: two of them used to mean four things each.

Building

Programs and prompts

Prompts, placeholders, output fields, edges and what they carry, routers, the entry and terminal prompts, and the limit on loops.

Datasets

What a row looks like, choosing the label column, the hold-out, and the pilot that prices a run before you start it.

Evaluate

Scorers, the metric vector, and how the objective weighs accuracy against prompt length.

Running and shipping

The optimizer

The search loop, the minibatch gate, rounds and budget, Explore versus Guided, and how to read a run without fooling yourself.

External steps

Calling a service the optimizer never rewrites: manifests, signing in, freezing for training, and the eight validation checks.

Versions and serving

Publishing an immutable version, its endpoint and API key, traces, outcomes, and promoting corrections into a dataset.

Accounts

Keys and connections

Five different secrets with five different jobs, where each one lives, and the rule that decides.

The engine

The search itself is bpto, open source, with its experiments and findings in the repo. The method and the evidence live on quantecarlo.com.