Feedback and preferred edits

People can tell a chat two useful things about a reply: a rating, whether it was good, and a preferred edit, what it should have said instead.

Recording feedback changes no model, prompt or agent by itself. It appends evidence to the session's journal, next to the reply it judges. june/train, and applications such as the Copy Agent, decide later how to use that evidence.

turn.completed (the reply, never changed)
  └ feedback.recorded  rating, comment, tags
  └ feedback.recorded  preferred: the reply a person would have wanted
        │
        ▼  later, when you choose
june/train: conversation datasets, preference pairs, eval cases

Recording feedback

Record feedback on the session, by turn id:

import type { Session } from "june/chat";

declare const session: Session;

const turn = session.send("Explain the render loop.");
await turn.result;

await session.feedback(turn.id, {
  rating: "negative",
  comment: "Too verbose.",
});
await session.feedback(turn.id, {
  preferred: "June re-renders only the steps whose inputs changed.",
  tags: ["verbosity"],
});

console.log(session.turns().at(-1)?.feedback);

• rating is "positive", "negative", or a number on your own scale. It judges the reply as it is.
• preferred is the reply a person would have wanted: a human edit. It's stored next to the original, which never changes, so a later preference dataset can use the edit as chosen and the original as rejected.
• comment and tags are free text and labels for whoever curates the data.

Each call validates the feedback, needs at least one field, and appends one feedback.recorded entry. Nothing is edited in place: a turn's feedback lists every entry oldest first, with its id and at, and a later rating doesn't erase an earlier one, so curation can decide which to trust. Feedback is only recorded on completed turns; running, failed, canceled and interrupted turns have no reply to judge, and feedback() rejects them before writing anything.

In the terminal, june chat's /feedback up and /feedback down rate the last reply; see the CLI reference. Edits and comments come from your own interface: a text box, an editor, or a command, as the Copy Agent does with /edit.

Feedback is written to the session's store like any other entry, so it's subject to the same one writer per session rule. Record feedback through the Session that's running the conversation, or after it's closed, not from a second process while it's open.

Feedback in context

By default, feedback is stored for later improvement and isn't shown to later turns: fullContext() and summaryContext() send the original replies, as they were, and never the ratings or edits. If you want later turns to see it, say so in a custom context, for example by showing the model a person's preferred edit instead of its own reply; see Custom context.

Knowing what was judged

Feedback is only useful if you know what produced the reply it judges. Every completed turn records its responder's kind, id and version, the model, usage and finish reason: see Provenance. Set a responder's version whenever its prompt, model or workflow changes, so feedback on old replies isn't mistaken for feedback on new ones. When you later compare a candidate with what's active, that's what tells you which version each piece of evidence judged.

From feedback to improvement

june/chat only records the signal. Turning it into training data and eval cases is june/train's job, which reads sessions and never the other way around:

• a turn with a preferred edit becomes a conversation example that teaches the edit, a preference pair with the edit chosen over the original, and an eval case that expects the edit;
• a positively rated turn can become a training target and a regression case;
• a negatively rated turn without an edit is eval material, never a training target.

Nothing here changes what answers your chats. Evaluating a candidate on cases from these sessions, and switching the chat to it, is a step a person takes. Feedback is written by people and can contain anything, including personal data, so review it before training on it; see Privacy and security.

Next: Learning from chat turns recorded feedback into datasets and eval cases, with its curation rules and provenance. src/examples/chat/learning.ts does it in one script.