Feedback

Record user feedback against a generation or session — the substrate for evals and fine-tuning datasets. Send it from your server (@bryel/feedback, secret key) or straight from the browser (@bryel/browser, publishable key). Both are zero-dependency.

Kinds and targets

Field	Values
`kind`	`thumb` · `score` · `comment` · `correction` · `label`
`target.type`	`message` (a single turn) · `session` (the whole conversation)
`source`	`end_user` · `labeler` · `model_judge` · `code`

A 👎 with a written note is a thumb (score: 0) plus a comment. Free-text feedback the user typed about the whole conversation is a comment on a session target.

Server (secret key)

Use a secret key (bk_…) from your backend. It has no origin requirement.

import { recordFeedback, thumbsUp } from "@bryel/feedback";

await recordFeedback({
  apiKey: process.env.BRYEL_KEY!,
  target: { type: "message", id: messageId },
  kind: "thumb",
  score: 0,
  comment: "wrong colors",
  userId,
  source: "end_user",
});

await thumbsUp({ apiKey, target: { type: "message", id: messageId } });

Session-level typed feedback uses a comment on a session target:

await recordFeedback({
  apiKey: process.env.BRYEL_KEY!,
  target: { type: "session", id: sessionId },
  kind: "comment",
  comment: "Visual design: good\nTask completion: yes\nEfficiency: no",
  source: "end_user",
});

Browser (publishable key)

@bryel/browser re-exports the same recordFeedback. In the browser, use a publishable key (bkp_…): it is write-only and origin-locked, so it is safe to ship in your client bundle — no backend proxy required. The /v1/feedback endpoint is CORS-enabled and accepts a bkp_ key sent from an allowed origin, the same model as browser traces.

import { recordFeedback } from "@bryel/browser";

await recordFeedback({
  apiKey: "bkp_…",
  target: { type: "message", id: messageId },
  kind: "thumb",
  score: 0,
  comment,
  sessionId,
  source: "end_user",
});

Only publishable bkp_ keys belong in client code. Never ship a secret bk_ key to the browser — proxy those through your backend.

Joining feedback to a trace

Feedback arrives after the answer, so target.id must be an id that also exists on the trace. Two options:

Reuse the id your tracing already produced. If each turn already has a stable id on its trace (the id you set as metadata.messageId, or a per-turn event id), pass that as target.id. No new id to mint — this is what the Framer integration does.
Mint one if you don’t have it. Otherwise create a messageId per turn, stamp it into telemetry metadata, and return it to the client.

Mint, stamp, and return on the server

const messageId = crypto.randomUUID();
streamText({ experimental_telemetry: { isEnabled: true, metadata: { sessionId, userId, messageId } } });
return result.toTextStreamResponse({ headers: { "x-bryel-message-id": messageId } });

Record feedback against that id

await recordFeedback({ apiKey, target: { type: "message", id: messageId }, kind: "thumb", score: 1, source: "end_user" });

To capture the messageId upstream in the first place, see the Vercel instrumentation guide.

Getting started

SDK

Concepts

Guides

Kinds and targets

Server (secret key)

Browser (publishable key)

Joining feedback to a trace

​Kinds and targets

​Server (secret key)

​Browser (publishable key)

​Joining feedback to a trace

Kinds and targets

Server (secret key)

Browser (publishable key)

Joining feedback to a trace