runbios-sdk 0.2.1-dev.105 → 0.2.1-dev.106

package/dist/index.d.ts

@@ -35,7 +35,7 @@ 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.105";
+export declare const VERSION = "0.2.1-dev.106";
 export declare class RunBiOS {
     /** Search models, fetch configs, check adapter compatibility. */
     readonly models: Models;

package/dist/index.js

@@ -35,7 +35,7 @@ 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.105';
+export const VERSION = '0.2.1-dev.106';
 export class RunBiOS {
     /** Search models, fetch configs, check adapter compatibility. */
     models;

package/dist/resources/loop.d.ts

@@ -1,5 +1,5 @@
 import type { HttpClient } from '../client.js';
-import type { LoopTrace, LoopTraceListResponse, LoopTraceListParams, LoopCaptureParams, LoopCaptureResult, LoopSignal, LoopSignalParams, LoopDataset, LoopDatasetListParams, LoopDatasetItem, LoopDatasetCreateParams, LoopConfig, LoopConfigParams, LoopStats } from '../types.js';
+import type { LoopTrace, LoopTraceListResponse, LoopTraceListParams, LoopCaptureParams, LoopCaptureResult, LoopSignal, LoopSignalParams, LoopCandidate, LoopCandidateParams, LoopDataset, LoopDatasetListParams, LoopDatasetItem, LoopDatasetCreateParams, LoopConfig, LoopConfigParams, LoopStats } from '../types.js';
 /**
  * The Conscious Loop -- capture what your model was asked and answered, record
  * whether it was right, and turn those judgements into training data.
@@ -18,7 +18,7 @@ import type { LoopTrace, LoopTraceListResponse, LoopTraceListParams, LoopCapture
  * await client.loop.setConfig('my-agent', { enabled: true, retention_days: 30 });
  *
  * // 2. Send what your model was asked and what it answered. Works whoever
- * //    served it -- us, OpenAI, your own vLLM.
+ * //    served it -- us, another provider, or your own servers.
  * const { trace_id } = await client.loop.capture({
  *   deployment_id: 'my-agent',
  *   model: 'gpt-4o',
@@ -46,7 +46,7 @@ export declare class Loop {
      * Send one exchange to be captured.
      *
      * Works regardless of who served the model: ours, OpenAI, Fireworks, your own
-     * vLLM, an agent framework. Pass `tool_calls` and `tools` when the turn
+     * own servers, an agent framework. Pass `tool_calls` and `tools` when the turn
      * invoked a function -- without them a tool-using exchange trains the model to
      * answer in prose where it should have called something.
      *
@@ -90,6 +90,37 @@ export declare class Loop {
     signal(traceId: string, params: LoopSignalParams): Promise<LoopSignal>;
     /** Every verdict on one conversation, oldest first. */
     listSignals(traceId: string): Promise<LoopSignal[]>;
+    /**
+     * Submit an alternative answer to a prompt already captured.
+     *
+     * This is what makes preference learning scale. A human rewrite is the best
+     * signal there is and the least available -- somebody has to sit down and
+     * write it. Sample the model several times for the same prompt, score the
+     * samples, and a DPO pair falls out automatically: best becomes chosen,
+     * worst becomes rejected. Point a stronger model at the prompt instead and
+     * the same machinery does distillation.
+     *
+     * Score them, or they cannot pair: an unscored alternative says nothing about
+     * which answer anybody prefers. `score_source` is required alongside a score,
+     * because precedence between a verifier, a judge and a person is the whole
+     * reason we record who judged.
+     *
+     * A human correction always outranks any score.
+     *
+     * @example
+     * ```ts
+     * for (const sample of await sampleMyModel(prompt, 4)) {
+     *   await client.loop.addCandidate(traceId, {
+     *     completion: sample.text,
+     *     score: await myVerifier(sample.text),
+     *     score_source: 'verifier',
+     *   });
+     * }
+     * ```
+     */
+    addCandidate(traceId: string, params: LoopCandidateParams): Promise<LoopCandidate>;
+    /** Every alternative answer recorded for one prompt, oldest first. */
+    listCandidates(traceId: string): Promise<LoopCandidate[]>;
     /**
      * Build a training set from the feedback recorded so far.
      *

package/dist/resources/loop.js

@@ -16,7 +16,7 @@
  * await client.loop.setConfig('my-agent', { enabled: true, retention_days: 30 });
  *
  * // 2. Send what your model was asked and what it answered. Works whoever
- * //    served it -- us, OpenAI, your own vLLM.
+ * //    served it -- us, another provider, or your own servers.
  * const { trace_id } = await client.loop.capture({
  *   deployment_id: 'my-agent',
  *   model: 'gpt-4o',
@@ -47,7 +47,7 @@ export class Loop {
      * Send one exchange to be captured.
      *
      * Works regardless of who served the model: ours, OpenAI, Fireworks, your own
-     * vLLM, an agent framework. Pass `tool_calls` and `tools` when the turn
+     * own servers, an agent framework. Pass `tool_calls` and `tools` when the turn
      * invoked a function -- without them a tool-using exchange trains the model to
      * answer in prose where it should have called something.
      *
@@ -121,6 +121,44 @@ export class Loop {
         const res = await this._http.fetchGet(`/api/loop/traces/${encodeURIComponent(traceId)}/signals`);
         return res.signals;
     }
+    // ── alternative answers ───────────────────────────────────────────────
+    /**
+     * Submit an alternative answer to a prompt already captured.
+     *
+     * This is what makes preference learning scale. A human rewrite is the best
+     * signal there is and the least available -- somebody has to sit down and
+     * write it. Sample the model several times for the same prompt, score the
+     * samples, and a DPO pair falls out automatically: best becomes chosen,
+     * worst becomes rejected. Point a stronger model at the prompt instead and
+     * the same machinery does distillation.
+     *
+     * Score them, or they cannot pair: an unscored alternative says nothing about
+     * which answer anybody prefers. `score_source` is required alongside a score,
+     * because precedence between a verifier, a judge and a person is the whole
+     * reason we record who judged.
+     *
+     * A human correction always outranks any score.
+     *
+     * @example
+     * ```ts
+     * for (const sample of await sampleMyModel(prompt, 4)) {
+     *   await client.loop.addCandidate(traceId, {
+     *     completion: sample.text,
+     *     score: await myVerifier(sample.text),
+     *     score_source: 'verifier',
+     *   });
+     * }
+     * ```
+     */
+    async addCandidate(traceId, params) {
+        const res = await this._http.fetchPost(`/api/loop/traces/${encodeURIComponent(traceId)}/candidates`, params);
+        return res.candidate;
+    }
+    /** Every alternative answer recorded for one prompt, oldest first. */
+    async listCandidates(traceId) {
+        const res = await this._http.fetchGet(`/api/loop/traces/${encodeURIComponent(traceId)}/candidates`);
+        return res.candidates;
+    }
     // ── training sets ─────────────────────────────────────────────────────
     /**
      * Build a training set from the feedback recorded so far.

package/dist/types.d.ts

@@ -2140,3 +2140,30 @@ export interface LoopStats {
     /** Upper bounds: deduplication runs when a set is built. */
     ready: Record<LoopMethod, number>;
 }
+export interface LoopCandidateParams {
+    completion?: string;
+    /** An alternative that is itself a tool call. */
+    tool_calls?: LoopToolCall[];
+    /** Which model produced this alternative. The teacher, when distilling. */
+    model?: string;
+    model_version?: string;
+    /** Unscored alternatives cannot pair -- a missing score is not a low one. */
+    score?: number;
+    /** Required alongside a score: human | verifier | judge | behavioural. */
+    score_source?: LoopSource;
+    reason?: string;
+    metadata?: Record<string, unknown>;
+}
+export interface LoopCandidate {
+    id: string;
+    trace_id: string;
+    model: string | null;
+    model_version: string | null;
+    completion: string;
+    tool_calls?: LoopToolCall[];
+    score: number | null;
+    score_source: LoopSource | null;
+    reason: string | null;
+    metadata: Record<string, unknown>;
+    created_at: string;
+}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "runbios-sdk",
-  "version": "0.2.1-dev.105",
+  "version": "0.2.1-dev.106",
   "description": "Official TypeScript SDK for the Run BiOS training and deployment platform API",
   "type": "module",
   "main": "./dist/index.js",