Learn¶

Start from what you have¶

A table of text : Label, sort or score every row, check it against answers you trust, make it better. Start here if unsure.

Turn notes into data : Field notes, reports, emails: several facts per note, your protocol, a score per field.

From a prompt you already have : Your OpenAI messages, sent exactly as they are; then types, tables and measurement.

Coming from another tool : The OpenAI SDK, DSPy, pandas and polars, Instructor and Pydantic AI, side by side.

Is it right? : A score with its range, every answer in a table, fair comparisons, how many rows to label.

Make it better : From free to expensive: rules, types, examples, optimizers, bigger models.

Make it cheaper : Several models on your data, tokens into money, where tokens go.

Big tables : Any data frame in, AI functions as columns, pandas or polars out.

Ship it: save, verify, load : A program and everything it depends on, in a folder, proven to run elsewhere.

Serve it over HTTP : A program as a web service; a served program used from Python like a local one.

Bake it into a small model : Train a small model to answer a function; send the unsure cases to a big one.

Every call, on record : Keep each call, mark answers right or wrong, and learn from the corrections.

How a function becomes a prompt : The mapping from Python to prompt, and what happens on each call.

Types : Lists, records, choices, maybe-missing values; as answers and as inputs.

Reasoning and several answers : Think first, return several values, get everything with predict.

Watch it being written : The answer as the model writes it; what an outside caller may see; every event, kept as it happens.

Tools : Let the model call your Python functions; a person approves the ones that change things.

Memory : Conversations: turns that remember each other, branches, stopping, helpers that remember.

Plugins : Change what programs do without changing them, every change on record.

Multi-step programs : Plain Python calling several AI functions, measured and saved as one.

Prompt formats and chat templates : How values are written into the prompt, and writing the conversation yourself.

Models and settings : Any provider, the default, and where settings come from.

Signing in : API keys and subscriptions (Claude, ChatGPT, Copilot, xAI).

Inspecting calls : Exactly what was sent and what came back.

When the model gets it wrong : Repairs, retries, provider errors, the reply cache.

Upgrading : From 1.1 to the next release, and from 0.x to 1.0.