feat: switch llm to local ollama qwen3

src/audio/guild-voice-session.ts

@@ -24,7 +24,7 @@ import { float32ToPcm16Buffer, int16ArrayToFloat32, Stereo48kToMono16kDownsample
 import { ConversationMemory, type UserUtterance } from "../services/conversation.js";
 import { ElevenLabsSttService } from "../services/elevenlabs-stt.js";
 import { ElevenLabsTtsService, type PreparedSpeechAudio } from "../services/elevenlabs-tts.js";
-import { OpenAiLlmService } from "../services/openai-llm.js";
+import type { LlmService } from "../services/llm.js";
 
 interface GuildVoiceSessionOptions {
   client: Client;
@@ -35,7 +35,7 @@ interface GuildVoiceSessionOptions {
   textChannelId?: string;
   stt: ElevenLabsSttService;
   tts: ElevenLabsTtsService;
-  llm: OpenAiLlmService;
+  llm: LlmService;
 }
 
 interface SpeechJob {

src/audio/local-voice-session.ts

@@ -14,14 +14,14 @@ import { takeFrame, int16ArrayToFloat32, float32ToPcm16Buffer } from "./pcm.js";
 import { ConversationMemory, type UserUtterance } from "../services/conversation.js";
 import { ElevenLabsSttService } from "../services/elevenlabs-stt.js";
 import { ElevenLabsTtsService, type PreparedSpeechAudio } from "../services/elevenlabs-tts.js";
-import { OpenAiLlmService } from "../services/openai-llm.js";
+import type { LlmService } from "../services/llm.js";
 
 interface LocalVoiceSessionOptions {
   config: AssistantRuntimeConfig;
   logger: Logger;
   stt: ElevenLabsSttService;
   tts: ElevenLabsTtsService;
-  llm: OpenAiLlmService;
+  llm: LlmService;
 }
 
 interface SpeechJob {

src/config.ts

@@ -15,12 +15,14 @@ const envSchema = z.object({
   DISCORD_BOT_TOKEN: emptyToUndefined,
   DISCORD_APPLICATION_ID: emptyToUndefined,
   DISCORD_COMMAND_GUILD_ID: emptyToUndefined,
-  OPENAI_API_KEY: emptyToUndefined,
-  OPENAI_MODEL: z.string().min(1).default("gpt-5.4-mini"),
   ELEVENLABS_API_KEY: emptyToUndefined,
   ELEVENLABS_VOICE_ID: emptyToUndefined,
   ELEVENLABS_STT_MODEL: z.string().min(1).default("scribe_v2_realtime"),
   ELEVENLABS_TTS_MODEL: z.string().min(1).default("eleven_flash_v2_5"),
+  OLLAMA_BASE_URL: z.string().min(1).default("http://localhost:11434"),
+  OLLAMA_MODEL: z.string().min(1).default("qwen3:0.6b"),
+  OLLAMA_KEEP_ALIVE: z.string().min(1).default("5m"),
+  OLLAMA_NUM_CTX: z.coerce.number().int().min(512).max(32768).default(4096),
   BOT_DEFAULT_LANGUAGE: z.string().min(2).default("ko"),
   MAX_CONVERSATION_TURNS: z.coerce.number().int().min(4).max(30).default(12),
   LOCAL_AUDIO_SOURCE: emptyToUndefined,
@@ -35,7 +37,6 @@ const envSchema = z.object({
 
 export type AppConfig = z.infer<typeof envSchema>;
 export type AssistantRuntimeConfig = AppConfig & {
-  OPENAI_API_KEY: string;
   ELEVENLABS_API_KEY: string;
   ELEVENLABS_VOICE_ID: string;
 };
@@ -58,7 +59,6 @@ function requirePresent(value: string | undefined, name: string): string {
 export function requireAssistantRuntimeConfig(config: AppConfig): AssistantRuntimeConfig {
   return {
     ...config,
-    OPENAI_API_KEY: requirePresent(config.OPENAI_API_KEY, "OPENAI_API_KEY"),
     ELEVENLABS_API_KEY: requirePresent(config.ELEVENLABS_API_KEY, "ELEVENLABS_API_KEY"),
     ELEVENLABS_VOICE_ID: requirePresent(config.ELEVENLABS_VOICE_ID, "ELEVENLABS_VOICE_ID"),
   };

src/discord-main.ts

@@ -17,7 +17,7 @@ import { type DiscordRuntimeConfig } from "./config.js";
 import { Logger } from "./logger.js";
 import { ElevenLabsSttService } from "./services/elevenlabs-stt.js";
 import { ElevenLabsTtsService } from "./services/elevenlabs-tts.js";
-import { OpenAiLlmService } from "./services/openai-llm.js";
+import { OllamaLlmService } from "./services/ollama-llm.js";
 
 export async function runDiscordBot(config: DiscordRuntimeConfig, logger: Logger): Promise<void> {
   const commands = [
@@ -39,7 +39,7 @@ export async function runDiscordBot(config: DiscordRuntimeConfig, logger: Logger
 
   const stt = new ElevenLabsSttService(config);
   const tts = new ElevenLabsTtsService(config);
-  const llm = new OpenAiLlmService(config);
+  const llm = new OllamaLlmService(config);
   const sessions = new Map<string, GuildVoiceSession>();
 
   function getVoiceChannel(interaction: ChatInputCommandInteraction): VoiceBasedChannel | null {

src/local-main.ts

@@ -8,7 +8,7 @@ import { Logger } from "./logger.js";
 import { LocalVoiceSession } from "./audio/local-voice-session.js";
 import { ElevenLabsSttService } from "./services/elevenlabs-stt.js";
 import { ElevenLabsTtsService } from "./services/elevenlabs-tts.js";
-import { OpenAiLlmService } from "./services/openai-llm.js";
+import { OllamaLlmService } from "./services/ollama-llm.js";
 
 function resolveFfmpegPath(): string {
   const ffmpegPath = ffmpegStatic as unknown as string | null;
@@ -74,7 +74,7 @@ export async function printLocalAudioDevices(): Promise<void> {
 export async function runLocalAssistant(config: AssistantRuntimeConfig, logger: Logger): Promise<void> {
   const stt = new ElevenLabsSttService(config);
   const tts = new ElevenLabsTtsService(config);
-  const llm = new OpenAiLlmService(config);
+  const llm = new OllamaLlmService(config);
   const session = new LocalVoiceSession({
     config,
     logger,

src/services/llm.ts

@@ -0,0 +1,5 @@
+import type { ConversationMemory, UserUtterance } from "./conversation.js";
+
+export interface LlmService {
+  generateReply(memory: ConversationMemory, utterance: UserUtterance): Promise<string>;
+}

src/services/ollama-llm.ts

@@ -0,0 +1,85 @@
+import type { AssistantRuntimeConfig } from "../config.js";
+import type { ConversationMemory, UserUtterance } from "./conversation.js";
+import type { LlmService } from "./llm.js";
+
+const ASSISTANT_INSTRUCTIONS = [
+  "너는 디스코드 음성 채널 또는 로컬 마이크 테스트에서 동작하는 한국어 음성 비서다.",
+  "답변은 짧고 실용적으로 한다.",
+  "기본은 한 문장, 길어도 두 문장을 넘기지 않는다.",
+  "말투는 자연스러운 한국어로 유지한다.",
+  "speaker_id와 speaker_name은 화자 구분용이므로 필요할 때만 자연스럽게 반영한다.",
+  "잘 못 들었거나 의미가 불명확하면 짧게 다시 물어본다.",
+  "목록, 마크다운, 코드블록은 쓰지 않는다.",
+  "생각 과정을 드러내지 말고 최종 답변만 말한다.",
+].join(" ");
+
+interface OllamaChatResponse {
+  message?: {
+    content?: string;
+    thinking?: string;
+  };
+  error?: string;
+}
+
+function normalizeReply(text: string): string {
+  const strippedThink = text.replace(/<think>[\s\S]*?<\/think>/gi, " ");
+  const compact = strippedThink.replace(/\s+/g, " ").trim();
+  if (compact.length <= 180) {
+    return compact;
+  }
+
+  const sentences = compact.match(/[^.!?]+[.!?]?/g);
+  if (!sentences || sentences.length === 0) {
+    return compact.slice(0, 180).trim();
+  }
+
+  return sentences.slice(0, 2).join(" ").trim().slice(0, 180).trim();
+}
+
+export class OllamaLlmService implements LlmService {
+  constructor(private readonly config: AssistantRuntimeConfig) {}
+
+  async generateReply(memory: ConversationMemory, utterance: UserUtterance): Promise<string> {
+    const url = new URL("/api/chat", this.config.OLLAMA_BASE_URL);
+    const response = await fetch(url, {
+      method: "POST",
+      headers: {
+        "Content-Type": "application/json",
+      },
+      body: JSON.stringify({
+        model: this.config.OLLAMA_MODEL,
+        messages: [
+          {
+            role: "system",
+            content: ASSISTANT_INSTRUCTIONS,
+          },
+          {
+            role: "user",
+            content: memory.buildPrompt(utterance),
+          },
+        ],
+        think: false,
+        stream: false,
+        keep_alive: this.config.OLLAMA_KEEP_ALIVE,
+        options: {
+          num_ctx: this.config.OLLAMA_NUM_CTX,
+          temperature: 0.4,
+          num_predict: 120,
+        },
+      }),
+    });
+
+    const body = (await response.json().catch(() => ({}))) as OllamaChatResponse;
+
+    if (!response.ok) {
+      throw new Error(body.error ?? `Ollama request failed with status ${response.status}`);
+    }
+
+    const output = body.message?.content?.trim();
+    if (!output) {
+      return "잘 못 들었습니다. 한 번만 다시 말씀해 주세요.";
+    }
+
+    return normalizeReply(output);
+  }
+}

src/services/openai-llm.ts

@@ -1,64 +0,0 @@
-import OpenAI from "openai";
-
-import type { AssistantRuntimeConfig } from "../config.js";
-import type { ConversationMemory, UserUtterance } from "./conversation.js";
-
-const ASSISTANT_INSTRUCTIONS = [
-  "너는 디스코드 음성 채널에서 동작하는 한국어 음성 비서다.",
-  "답변은 짧고 실용적으로 한다.",
-  "기본은 한 문장, 길어도 두 문장을 넘기지 않는다.",
-  "말투는 자연스러운 한국어로 유지한다.",
-  "speaker_id와 speaker_name은 화자 구분용이므로 필요할 때만 자연스럽게 반영한다.",
-  "잘 못 들었거나 의미가 불명확하면 짧게 다시 물어본다.",
-  "목록, 마크다운, 코드블록은 쓰지 않는다.",
-].join(" ");
-
-function normalizeReply(text: string): string {
-  const compact = text.replace(/\s+/g, " ").trim();
-  if (compact.length <= 180) {
-    return compact;
-  }
-
-  const sentences = compact.match(/[^.!?]+[.!?]?/g);
-  if (!sentences || sentences.length === 0) {
-    return compact.slice(0, 180).trim();
-  }
-
-  return sentences.slice(0, 2).join(" ").trim().slice(0, 180).trim();
-}
-
-export class OpenAiLlmService {
-  private readonly client: OpenAI;
-
-  constructor(private readonly config: AssistantRuntimeConfig) {
-    this.client = new OpenAI({
-      apiKey: this.config.OPENAI_API_KEY,
-    });
-  }
-
-  async generateReply(memory: ConversationMemory, utterance: UserUtterance): Promise<string> {
-    const response = await this.client.responses.create({
-      model: this.config.OPENAI_MODEL,
-      instructions: ASSISTANT_INSTRUCTIONS,
-      input: [
-        {
-          role: "user",
-          content: [
-            {
-              type: "input_text",
-              text: memory.buildPrompt(utterance),
-            },
-          ],
-        },
-      ],
-      max_output_tokens: 120,
-    });
-
-    const output = response.output_text?.trim();
-    if (!output) {
-      return "잘 못 들었습니다. 한 번만 다시 말씀해 주세요.";
-    }
-
-    return normalizeReply(output);
-  }
-}