@dtelecom/agents-js 0.1.14 → 0.1.16

package/dist/index.d.mts

@@ -1,8 +1,8 @@
 import * as _dtelecom_server_sdk_node from '@dtelecom/server-sdk-node';
 import { Room, AudioSource, RemoteAudioTrack, AudioFrame } from '@dtelecom/server-sdk-node';
 import { EventEmitter } from 'events';
-import { A as AgentConfig, a as AgentStartOptions, M as Message, L as LLMPlugin, P as PipelineOptions, b as AgentState, S as STTStream, T as TranscriptionResult } from './types-BVMiP1bW.mjs';
-export { c as AgentEvents, d as AudioOutput, D as DataMessageHandler, e as LLMChunk, f as MemoryConfig, g as PipelineEvents, R as RespondMode, h as STTPlugin, i as STTStreamOptions, j as TTSPlugin } from './types-BVMiP1bW.mjs';
+import { A as AgentConfig, a as AgentStartOptions, M as Message, L as LLMPlugin, P as PipelineOptions, b as AgentState, S as STTStream, T as TranscriptionResult } from './types-MPHcuMhp.mjs';
+export { c as AgentEvents, d as AudioOutput, D as DataMessageHandler, e as LLMChunk, f as MemoryConfig, g as PipelineEvents, R as RespondMode, h as STTPlugin, i as STTStreamOptions, j as TTSPlugin } from './types-MPHcuMhp.mjs';
 
 declare class VoiceAgent extends EventEmitter {
     private readonly config;
@@ -23,6 +23,14 @@ declare class VoiceAgent extends EventEmitter {
      * Use for greetings or announcements. Supports barge-in.
      */
     say(text: string): Promise<void>;
+    /**
+     * Switch STT language on all active streams.
+     * Use for bilingual lessons — e.g. switch to 'es' with forceWhisper when
+     * expecting Spanish, back to 'auto' for Parakeet auto-detect otherwise.
+     */
+    setSTTLanguage(language: string, options?: {
+        forceWhisper?: boolean;
+    }): void;
     /** Start the agent — connect to room and begin listening. */
     start(options: AgentStartOptions): Promise<void>;
     /** Stop the agent — disconnect and clean up. */
@@ -124,6 +132,10 @@ declare class Pipeline extends EventEmitter {
     removeParticipant(identity: string): Promise<void>;
     stop(): Promise<void>;
     getContextManager(): ContextManager;
+    /** Switch STT language on all active streams (e.g. for bilingual lessons). */
+    setSTTLanguage(language: string, options?: {
+        forceWhisper?: boolean;
+    }): void;
     private lastFinalAt;
     private lastSttDuration;
     private handleTranscription;

package/dist/index.d.ts

@@ -1,8 +1,8 @@
 import * as _dtelecom_server_sdk_node from '@dtelecom/server-sdk-node';
 import { Room, AudioSource, RemoteAudioTrack, AudioFrame } from '@dtelecom/server-sdk-node';
 import { EventEmitter } from 'events';
-import { A as AgentConfig, a as AgentStartOptions, M as Message, L as LLMPlugin, P as PipelineOptions, b as AgentState, S as STTStream, T as TranscriptionResult } from './types-BVMiP1bW.js';
-export { c as AgentEvents, d as AudioOutput, D as DataMessageHandler, e as LLMChunk, f as MemoryConfig, g as PipelineEvents, R as RespondMode, h as STTPlugin, i as STTStreamOptions, j as TTSPlugin } from './types-BVMiP1bW.js';
+import { A as AgentConfig, a as AgentStartOptions, M as Message, L as LLMPlugin, P as PipelineOptions, b as AgentState, S as STTStream, T as TranscriptionResult } from './types-MPHcuMhp.js';
+export { c as AgentEvents, d as AudioOutput, D as DataMessageHandler, e as LLMChunk, f as MemoryConfig, g as PipelineEvents, R as RespondMode, h as STTPlugin, i as STTStreamOptions, j as TTSPlugin } from './types-MPHcuMhp.js';
 
 declare class VoiceAgent extends EventEmitter {
     private readonly config;
@@ -23,6 +23,14 @@ declare class VoiceAgent extends EventEmitter {
      * Use for greetings or announcements. Supports barge-in.
      */
     say(text: string): Promise<void>;
+    /**
+     * Switch STT language on all active streams.
+     * Use for bilingual lessons — e.g. switch to 'es' with forceWhisper when
+     * expecting Spanish, back to 'auto' for Parakeet auto-detect otherwise.
+     */
+    setSTTLanguage(language: string, options?: {
+        forceWhisper?: boolean;
+    }): void;
     /** Start the agent — connect to room and begin listening. */
     start(options: AgentStartOptions): Promise<void>;
     /** Stop the agent — disconnect and clean up. */
@@ -124,6 +132,10 @@ declare class Pipeline extends EventEmitter {
     removeParticipant(identity: string): Promise<void>;
     stop(): Promise<void>;
     getContextManager(): ContextManager;
+    /** Switch STT language on all active streams (e.g. for bilingual lessons). */
+    setSTTLanguage(language: string, options?: {
+        forceWhisper?: boolean;
+    }): void;
     private lastFinalAt;
     private lastSttDuration;
     private handleTranscription;

package/dist/index.js

@@ -1274,11 +1274,20 @@ var Pipeline = class extends import_events.EventEmitter {
   getContextManager() {
     return this.context;
   }
+  /** Switch STT language on all active streams (e.g. for bilingual lessons). */
+  setSTTLanguage(language, options) {
+    for (const [identity, stream] of this.sttStreams) {
+      if (stream.setLanguage) {
+        stream.setLanguage(language, options);
+        log7.info(`STT language \u2192 ${language}${options?.forceWhisper ? " (whisper)" : ""} for "${identity}"`);
+      }
+    }
+  }
   lastFinalAt = 0;
   lastSttDuration = 0;
   async handleTranscription(speaker, result) {
     this.emit("transcription", { ...result, speaker });
-    if (!result.isFinal && result.text.trim()) {
+    if (!result.isFinal) {
       this.setAgentState("listening");
     }
     if (this.audioOutput.playing && result.text.trim().length > 0) {
@@ -1363,60 +1372,69 @@ var Pipeline = class extends import_events.EventEmitter {
         sentenceQueue.push(text2);
         wake();
       };
+      const MAX_LLM_RETRIES = 2;
       const producer = async () => {
-        let isFirstChunk = true;
         const defaultLang = this.tts?.defaultLanguage;
-        const segBuf = [];
-        const flushSegments = () => {
-          if (segBuf.length === 0) return;
-          const combined = segBuf.map(
-            (s) => s.lang !== defaultLang ? `<lang xml:lang="${s.lang}">${s.text}</lang>` : s.text
-          ).join(" ");
-          segBuf.length = 0;
-          pushSentence(combined);
-        };
-        const llmStream = this.llm.chat(messages, signal);
-        try {
-          while (!signal.aborted) {
-            const { value: chunk, done } = await llmStream.next();
-            if (done || !chunk) break;
-            if (signal.aborted) break;
-            if (chunk.type === "segment" && chunk.segment) {
-              if (isFirstChunk) {
-                tLlmFirstToken = performance.now();
-                isFirstChunk = false;
-                log7.info(`llm_first_segment: ${(tLlmFirstToken - tSpeechEnd).toFixed(0)}ms`);
-              }
-              if (fullResponse) fullResponse += " ";
-              fullResponse += chunk.segment.text;
-              segBuf.push(chunk.segment);
-              if (/[.!?]["'»)]*\s*$/.test(chunk.segment.text)) {
-                flushSegments();
-              }
-            } else if (chunk.type === "token" && chunk.token) {
-              if (isFirstChunk) {
-                tLlmFirstToken = performance.now();
-                isFirstChunk = false;
-                log7.info(`llm_first_token: ${(tLlmFirstToken - tSpeechEnd).toFixed(0)}ms`);
-              }
-              fullResponse += chunk.token;
-              const sentences = this.splitter.push(chunk.token);
-              for (const sentence of sentences) {
-                pushSentence(sentence);
+        for (let attempt = 0; attempt <= MAX_LLM_RETRIES; attempt++) {
+          if (signal.aborted) break;
+          if (attempt > 0) {
+            log7.warn(`LLM retry ${attempt}/${MAX_LLM_RETRIES}...`);
+            this.splitter.reset();
+          }
+          let isFirstChunk = true;
+          const segBuf = [];
+          const flushSegments = () => {
+            if (segBuf.length === 0) return;
+            const combined = segBuf.map(
+              (s) => s.lang !== defaultLang ? `<lang xml:lang="${s.lang}">${s.text}</lang>` : s.text
+            ).join(" ");
+            segBuf.length = 0;
+            pushSentence(combined);
+          };
+          const llmStream = this.llm.chat(messages, signal);
+          try {
+            while (!signal.aborted) {
+              const { value: chunk, done } = await llmStream.next();
+              if (done || !chunk) break;
+              if (signal.aborted) break;
+              if (chunk.type === "segment" && chunk.segment) {
+                if (isFirstChunk) {
+                  tLlmFirstToken = performance.now();
+                  isFirstChunk = false;
+                  log7.info(`llm_first_segment: ${(tLlmFirstToken - tSpeechEnd).toFixed(0)}ms`);
+                }
+                if (fullResponse) fullResponse += " ";
+                fullResponse += chunk.segment.text;
+                segBuf.push(chunk.segment);
+                if (/[.!?]["'»)]*\s*$/.test(chunk.segment.text)) {
+                  flushSegments();
+                }
+              } else if (chunk.type === "token" && chunk.token) {
+                if (isFirstChunk) {
+                  tLlmFirstToken = performance.now();
+                  isFirstChunk = false;
+                  log7.info(`llm_first_token: ${(tLlmFirstToken - tSpeechEnd).toFixed(0)}ms`);
+                }
+                fullResponse += chunk.token;
+                const sentences = this.splitter.push(chunk.token);
+                for (const sentence of sentences) {
+                  pushSentence(sentence);
+                }
               }
             }
+          } finally {
+            await llmStream.return(void 0);
           }
-        } finally {
-          await llmStream.return(void 0);
-        }
-        if (!signal.aborted) {
-          flushSegments();
-          const remaining = this.splitter.flush();
-          if (remaining) {
-            pushSentence(remaining);
-          }
-          if (!fullResponse.trim()) {
-            log7.warn("LLM produced no output (empty response or no segments detected)");
+          if (!signal.aborted) {
+            flushSegments();
+            const remaining = this.splitter.flush();
+            if (remaining) {
+              pushSentence(remaining);
+            }
+            if (fullResponse.trim()) {
+              break;
+            }
+            log7.warn(`LLM produced no output (attempt ${attempt + 1}/${MAX_LLM_RETRIES + 1})`);
           }
         }
         producerDone = true;
@@ -1438,7 +1456,7 @@ var Pipeline = class extends import_events.EventEmitter {
                   tFirstAudioPlayed = t;
                   this.setAgentState("speaking");
                 }
-                this.emit("sentence", this.cleanText(sentence));
+                this.emit("sentence", this.cleanText(sentence), sentence);
               });
               continue;
             }
@@ -1507,7 +1525,7 @@ var Pipeline = class extends import_events.EventEmitter {
       this.setAgentState("thinking");
       await this.synthesizeAndPlay(text, signal, () => {
         this.setAgentState("speaking");
-        this.emit("sentence", this.cleanText(text));
+        this.emit("sentence", this.cleanText(text), text);
       });
       if (!signal.aborted) {
         await this.audioOutput.writeSilence(40);
@@ -1606,6 +1624,14 @@ var VoiceAgent = class extends import_events2.EventEmitter {
     }
     await this.pipeline.say(text);
   }
+  /**
+   * Switch STT language on all active streams.
+   * Use for bilingual lessons — e.g. switch to 'es' with forceWhisper when
+   * expecting Spanish, back to 'auto' for Parakeet auto-detect otherwise.
+   */
+  setSTTLanguage(language, options) {
+    this.pipeline?.setSTTLanguage(language, options);
+  }
   /** Start the agent — connect to room and begin listening. */
   async start(options) {
     if (this._running) {
@@ -1647,7 +1673,7 @@ var VoiceAgent = class extends import_events2.EventEmitter {
       maxContextTokens: this.config.maxContextTokens
     });
     this.pipeline.on("transcription", (result) => this.emit("transcription", result));
-    this.pipeline.on("sentence", (text) => this.emit("sentence", text));
+    this.pipeline.on("sentence", (text, raw) => this.emit("sentence", text, raw));
     this.pipeline.on("response", (text) => this.emit("response", text));
     this.pipeline.on("agentState", (state) => this.emit("agentState", state));
     this.pipeline.on("error", (error) => this.emit("error", error));