runbios-sdk 0.2.1-dev.112 → 0.2.1-dev.114

package/dist/index.d.ts

@@ -36,7 +36,7 @@ import { Loop } from './resources/loop.js';
  * SDK version. Sent as part of the User-Agent header.
  * Must match package.json "version" -- enforced by a contract test.
  */
-export declare const VERSION = "0.2.1-dev.112";
+export declare const VERSION = "0.2.1-dev.114";
 export declare class RunBiOS {
     /** Search models, fetch configs, check adapter compatibility. */
     readonly models: Models;

package/dist/index.js

@@ -36,7 +36,7 @@ import { Loop } from './resources/loop.js';
  * SDK version. Sent as part of the User-Agent header.
  * Must match package.json "version" -- enforced by a contract test.
  */
-export const VERSION = '0.2.1-dev.112';
+export const VERSION = '0.2.1-dev.114';
 export class RunBiOS {
     /** Search models, fetch configs, check adapter compatibility. */
     models;

package/dist/resources/loop.d.ts

@@ -171,9 +171,18 @@ export declare class Loop {
      *
      * Labels are lowercased and trimmed, so `Refunds` and `refunds` are one label.
      */
-    label(traceId: string, labels: string[], parent?: string): Promise<LoopLabel[]>;
-    unlabel(traceId: string, label: string): Promise<{
+    label(traceId: string, labels: string[], parent?: string, attributes?: Record<string, string>): Promise<LoopLabel[]>;
+    /**
+     * Set named dimensions on a conversation: `{ category: 'billing',
+     * language: 'es' }`.
+     *
+     * An upsert per dimension, so correcting `category` leaves `task` and any
+     * bare tags exactly where they were — no delete-then-add.
+     */
+    setAttributes(traceId: string, attributes: Record<string, string>): Promise<LoopLabel[]>;
+    unlabel(traceId: string, label: string, key?: string): Promise<{
         deleted: boolean;
+        key: string;
         label: string;
     }>;
     /** Every label in the workspace, with how many conversations carry it. */
@@ -187,7 +196,11 @@ export declare class Loop {
      * - `sft` -- answers you marked right, and answers you rewrote.
      * - `dpo` -- answers you REWROTE, so a better and a worse version of the same
      *   reply exist. Nothing else produces a pair.
-     * - `grpo` -- answers with a value or fact they can be checked against.
+     * - `grpo` -- answers with a value or fact they can be checked against, or
+     *   a judge rubric, which scores answers the run has not written yet.
+     * - `kto` -- anything carrying a yes or a no, INCLUDING a thumbs-down with
+     *   nothing written. That row trains nothing under the other three methods,
+     *   which is why this one exists: it is the feedback people actually give.
      *
      * `holdout_percent` holds back your most recent work rather than a random
      * slice, so the evaluation measures whether the model generalised instead of

package/dist/resources/loop.js

@@ -76,6 +76,16 @@ export class Loop {
             q.set('to', params.to);
         if (params.signalled)
             q.set('signalled', 'true');
+        if (params.label)
+            q.set('label', params.label);
+        // Repeated rather than comma-joined: a value containing a comma would
+        // otherwise silently split into two filters that match nothing.
+        for (const [k, v] of Object.entries(params.attributes ?? {})) {
+            if (k && v)
+                q.append('attr', `${k}:${v}`);
+        }
+        if (params.unlabelled)
+            q.set('unlabelled', 'true');
         if (params.limit != null)
             q.set('limit', String(params.limit));
         if (params.offset != null)
@@ -218,12 +228,24 @@ export class Loop {
      *
      * Labels are lowercased and trimmed, so `Refunds` and `refunds` are one label.
      */
-    async label(traceId, labels, parent) {
-        const res = await this._http.fetchPost(`/api/loop/traces/${encodeURIComponent(traceId)}/labels`, { labels, parent });
+    async label(traceId, labels, parent, attributes) {
+        const res = await this._http.fetchPost(`/api/loop/traces/${encodeURIComponent(traceId)}/labels`, { labels, parent, attributes });
         return res.labels;
     }
-    async unlabel(traceId, label) {
-        return this._http.fetchDelete(`/api/loop/traces/${encodeURIComponent(traceId)}/labels/${encodeURIComponent(label)}`);
+    /**
+     * Set named dimensions on a conversation: `{ category: 'billing',
+     * language: 'es' }`.
+     *
+     * An upsert per dimension, so correcting `category` leaves `task` and any
+     * bare tags exactly where they were — no delete-then-add.
+     */
+    async setAttributes(traceId, attributes) {
+        const res = await this._http.fetchPost(`/api/loop/traces/${encodeURIComponent(traceId)}/labels`, { attributes });
+        return res.labels;
+    }
+    async unlabel(traceId, label, key) {
+        const q = key ? `?key=${encodeURIComponent(key)}` : '';
+        return this._http.fetchDelete(`/api/loop/traces/${encodeURIComponent(traceId)}/labels/${encodeURIComponent(label)}${q}`);
     }
     /** Every label in the workspace, with how many conversations carry it. */
     async listLabels() {
@@ -240,7 +262,11 @@ export class Loop {
      * - `sft` -- answers you marked right, and answers you rewrote.
      * - `dpo` -- answers you REWROTE, so a better and a worse version of the same
      *   reply exist. Nothing else produces a pair.
-     * - `grpo` -- answers with a value or fact they can be checked against.
+     * - `grpo` -- answers with a value or fact they can be checked against, or
+     *   a judge rubric, which scores answers the run has not written yet.
+     * - `kto` -- anything carrying a yes or a no, INCLUDING a thumbs-down with
+     *   nothing written. That row trains nothing under the other three methods,
+     *   which is why this one exists: it is the feedback people actually give.
      *
      * `holdout_percent` holds back your most recent work rather than a random
      * slice, so the evaluation measures whether the model generalised instead of

package/dist/types.d.ts

@@ -1959,7 +1959,7 @@ export interface ChatCompletionUsage {
     prompt_tokens_details?: PromptTokensDetails;
 }
 /** What a training set is shaped for. The three need different things. */
-export type LoopMethod = 'sft' | 'dpo' | 'grpo';
+export type LoopMethod = 'sft' | 'dpo' | 'grpo' | 'kto';
 /** Who judged an answer. */
 export type LoopSource = 'human' | 'verifier' | 'judge' | 'behavioural';
 /** The judgement itself. */
@@ -2073,6 +2073,15 @@ export interface LoopTraceListParams {
     to?: string;
     /** Only conversations that already carry a verdict. */
     signalled?: boolean;
+    /** One bare tag. A tag with children matches them too. */
+    label?: string;
+    /**
+     * Named dimensions, ANDed together: `{ category: 'billing', language: 'es' }`.
+     * Exact within their key — `source=support` does not match `team=support`.
+     */
+    attributes?: Record<string, string>;
+    /** Only the conversations nobody has described yet. */
+    unlabelled?: boolean;
     limit?: number;
     offset?: number;
 }
@@ -2095,6 +2104,14 @@ export interface LoopDatasetCreateParams {
     /** Holds back your most recent work, not a random slice. 0-50. */
     holdout_percent?: number;
     max_items?: number;
+    /**
+     * Named dimensions, ANDed with each other and with `label`:
+     * `{ category: 'billing', language: 'es' }`. This is what turns one captured
+     * corpus into a different dataset for every task somebody trains for.
+     */
+    attributes?: Record<string, string>;
+    /** Only the conversations nobody has described yet. */
+    unlabelled?: boolean;
 }
 export interface LoopDataset {
     id: string;
@@ -2358,9 +2375,12 @@ export interface LoopLabel {
     parent: string | null;
     created_by: string | null;
     created_at: string;
+    /** The dimension this value belongs to. `tag` for a bare name. */
+    key: string;
 }
 export interface LoopLabelCount {
     label: string;
     parent: string | null;
     traces: number;
+    key: string;
 }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "runbios-sdk",
-  "version": "0.2.1-dev.112",
+  "version": "0.2.1-dev.114",
   "description": "Official TypeScript SDK for the Run BiOS training and deployment platform API",
   "type": "module",
   "main": "./dist/index.js",