Phase 4 구현: Ollama LLM 연동 (다듬기, 번역, 스트리밍)

- LocalLLMService: Ollama REST API, 스트리밍 NDJSON 파싱, 5초 가용성 폴링
- 시스템 프롬프트: refine/translate/summarize/grammar/expand/custom 6개 액션
- VoiceModeService: 전사→LLM 후처리→텍스트 삽입, LLM 실패 시 원본 폴백
- LLM IPC 핸들러: status/models/process/cancel/serverUrl 8개
- StatusBar: Ollama 연결 상태 + 활성 모델 표시
- Preload: llm API 섹션 추가
- Bootstrap: llm-polling 초기화 단계 추가
This commit is contained in:
Yun Chan 2026-04-05 02:18:26 +09:00
parent 517210af2f
commit 4a5cf6c819
10 changed files with 663 additions and 9 deletions

View file

@ -0,0 +1,433 @@
// src/main/services/LocalLLMService.ts
// Ollama REST API를 통해 로컬 LLM과 상호작용한다.
// 설계서 01의 ILocalLLMService 구현.
import { EventEmitter } from 'events'
import { getLogger } from './LoggerService'
import { configGet } from './ConfigService'
import { D3ROError, ErrorCode } from '@shared/errors'
import type { LLMStatus, LLMModel, LLMAction, LLMConnectionState } from '@shared/types'
const logger = getLogger('LocalLLMService')
// ============================================================
// 내부 타입
// ============================================================
const enum LLMState {
Unavailable = 'unavailable',
Available = 'available',
Generating = 'generating',
Error = 'error'
}
interface GenerateOptions {
model?: string
temperature?: number
maxTokens?: number
systemPrompt?: string
stream?: boolean
}
interface GenerateResult {
text: string
model: string
promptTokens: number
completionTokens: number
totalDuration: number
}
interface OllamaGenerateResponse {
model: string
response: string
done: boolean
total_duration?: number
prompt_eval_count?: number
eval_count?: number
}
interface OllamaTagsResponse {
models: Array<{
name: string
size: number
parameter_size: string
quantization_level: string
modified_at: string
}>
}
interface LocalLLMEvents {
token: (payload: { token: string; done: boolean }) => void
complete: (payload: { result: GenerateResult }) => void
'availability-changed': (payload: { available: boolean }) => void
error: (payload: { error: D3ROError }) => void
}
// ============================================================
// 시스템 프롬프트 (설계서 Phase 4 참조)
// ============================================================
const SYSTEM_PROMPTS: Record<string, string> = {
refine: `다음 음성 전사 텍스트를 자연스럽고 격식 있는 문어체로 다듬어주세요.
, .
. .`,
translate: `다음 텍스트를 {{targetLanguage}}로 번역해주세요.
. .`,
summarize: `다음 텍스트의 핵심 내용을 3줄 이내로 요약해주세요.
.`,
grammar: `다음 텍스트의 문법 오류만 수정해주세요.
.
.`,
expand: `다음 텍스트를 더 자세하고 풍부하게 확장해주세요.
.`
}
// ============================================================
// LocalLLMService
// ============================================================
class LocalLLMService extends EventEmitter {
private _state = LLMState.Unavailable
private _pollInterval: ReturnType<typeof setInterval> | null = null
private _available = false
private _abortController: AbortController | null = null
private _disposed = false
get state(): LLMState {
return this._state
}
/**
* Ollama (5 ).
*/
startPolling(): void {
this._checkAvailability()
this._pollInterval = setInterval(() => {
if (this._state !== LLMState.Generating) {
this._checkAvailability()
}
}, 5000)
logger.info('Ollama availability polling started')
}
stopPolling(): void {
if (this._pollInterval) {
clearInterval(this._pollInterval)
this._pollInterval = null
}
}
/**
* .
*/
async generate(prompt: string, options?: GenerateOptions): Promise<GenerateResult> {
if (!this._available) {
throw new D3ROError(ErrorCode.LLMServerUnreachable, 'Ollama 서버에 연결할 수 없습니다')
}
const serverUrl = configGet('ollamaServerUrl')
const model = options?.model ?? configGet('llmModelId') ?? 'qwen3:4b'
this._state = LLMState.Generating
try {
const response = await fetch(`${serverUrl}/api/generate`, {
method: 'POST',
headers: { 'Content-Type': 'application/json' },
body: JSON.stringify({
model,
prompt,
system: options?.systemPrompt,
stream: false,
options: {
temperature: options?.temperature ?? 0.3,
num_predict: options?.maxTokens ?? 2048
}
}),
signal: AbortSignal.timeout(120000)
})
if (!response.ok) {
throw new D3ROError(
ErrorCode.LLMProcessingFailed,
`Ollama responded with ${response.status}: ${response.statusText}`
)
}
const data = (await response.json()) as OllamaGenerateResponse
const result: GenerateResult = {
text: data.response,
model: data.model,
promptTokens: data.prompt_eval_count ?? 0,
completionTokens: data.eval_count ?? 0,
totalDuration: data.total_duration ? data.total_duration / 1e6 : 0
}
this._state = LLMState.Available
this.emit('complete', { result })
return result
} catch (error) {
this._state = LLMState.Available
if (error instanceof D3ROError) throw error
throw new D3ROError(
ErrorCode.LLMProcessingFailed,
`LLM generation failed: ${error instanceof Error ? error.message : String(error)}`
)
}
}
/**
* . NDJSON .
* AbortController로 .
*/
async *streamGenerate(
prompt: string,
options?: Omit<GenerateOptions, 'stream'>
): AsyncGenerator<string, GenerateResult> {
if (!this._available) {
throw new D3ROError(ErrorCode.LLMServerUnreachable, 'Ollama 서버에 연결할 수 없습니다')
}
const serverUrl = configGet('ollamaServerUrl')
const model = options?.model ?? configGet('llmModelId') ?? 'qwen3:4b'
this._state = LLMState.Generating
this._abortController = new AbortController()
try {
const response = await fetch(`${serverUrl}/api/generate`, {
method: 'POST',
headers: { 'Content-Type': 'application/json' },
body: JSON.stringify({
model,
prompt,
system: options?.systemPrompt,
stream: true,
options: {
temperature: options?.temperature ?? 0.3,
num_predict: options?.maxTokens ?? 2048
}
}),
signal: this._abortController.signal
})
if (!response.ok || !response.body) {
throw new D3ROError(
ErrorCode.LLMProcessingFailed,
`Ollama responded with ${response.status}`
)
}
const reader = response.body.getReader()
const decoder = new TextDecoder()
let buffer = ''
let fullText = ''
let lastChunk: OllamaGenerateResponse | null = null
while (true) {
const { done, value } = await reader.read()
if (done) break
buffer += decoder.decode(value, { stream: true })
const lines = buffer.split('\n')
buffer = lines.pop() ?? ''
for (const line of lines) {
if (!line.trim()) continue
try {
const chunk = JSON.parse(line) as OllamaGenerateResponse
fullText += chunk.response
this.emit('token', { token: chunk.response, done: chunk.done })
yield chunk.response
if (chunk.done) {
lastChunk = chunk
}
} catch {
logger.warn(`Failed to parse NDJSON line: ${line.substring(0, 100)}`)
}
}
}
this._state = LLMState.Available
this._abortController = null
const result: GenerateResult = {
text: fullText,
model: lastChunk?.model ?? model,
promptTokens: lastChunk?.prompt_eval_count ?? 0,
completionTokens: lastChunk?.eval_count ?? 0,
totalDuration: lastChunk?.total_duration ? lastChunk.total_duration / 1e6 : 0
}
this.emit('complete', { result })
return result
} catch (error) {
this._state = LLMState.Available
this._abortController = null
if (error instanceof D3ROError) throw error
if (error instanceof DOMException && error.name === 'AbortError') {
throw new D3ROError(ErrorCode.LLMProcessingCancelled, 'LLM generation cancelled')
}
throw new D3ROError(
ErrorCode.LLMProcessingFailed,
`LLM streaming failed: ${error instanceof Error ? error.message : String(error)}`
)
}
}
/**
* LLM .
*/
async processText(
text: string,
action: LLMAction,
targetLanguage?: string,
customPrompt?: string
): Promise<string> {
let systemPrompt: string
if (action === 'custom' && customPrompt) {
systemPrompt = customPrompt
} else if (action === 'translate') {
systemPrompt = SYSTEM_PROMPTS.translate.replace(
'{{targetLanguage}}',
targetLanguage ?? 'English'
)
} else {
systemPrompt = SYSTEM_PROMPTS[action] ?? SYSTEM_PROMPTS.refine
}
const result = await this.generate(text, { systemPrompt })
return result.text.trim()
}
cancelGeneration(): void {
if (this._abortController) {
this._abortController.abort()
this._abortController = null
logger.info('LLM generation cancelled')
}
}
/**
* Ollama에 .
*/
async getModels(): Promise<LLMModel[]> {
const serverUrl = configGet('ollamaServerUrl')
try {
const response = await fetch(`${serverUrl}/api/tags`, {
signal: AbortSignal.timeout(5000)
})
if (!response.ok) return []
const data = (await response.json()) as OllamaTagsResponse
return data.models.map((m) => ({
id: m.name,
name: m.name,
sizeBytes: m.size,
parameterSize: m.parameter_size ?? '',
quantization: m.quantization_level ?? '',
modifiedAt: m.modified_at
}))
} catch {
return []
}
}
getStatus(): LLMStatus {
const connectionState: LLMConnectionState = this._available
? this._state === LLMState.Generating
? 'connecting'
: 'connected'
: 'disconnected'
return {
connectionState,
serverUrl: configGet('ollamaServerUrl'),
activeModel: configGet('llmModelId'),
serverVersion: null
}
}
isAvailable(): boolean {
return this._available
}
dispose(): void {
this._disposed = true
this.stopPolling()
this.cancelGeneration()
this.removeAllListeners()
logger.info('LocalLLMService disposed')
}
// ── 가용성 체크 ────────────────────────────────────────
private async _checkAvailability(): Promise<void> {
if (this._disposed) return
const serverUrl = configGet('ollamaServerUrl')
try {
const response = await fetch(`${serverUrl}/api/tags`, {
signal: AbortSignal.timeout(3000)
})
const wasAvailable = this._available
this._available = response.ok
if (!wasAvailable && this._available) {
this._state = LLMState.Available
this.emit('availability-changed', { available: true })
logger.info('Ollama server connected')
} else if (wasAvailable && !this._available) {
this._state = LLMState.Unavailable
this.emit('availability-changed', { available: false })
logger.warn('Ollama server disconnected')
}
} catch {
if (this._available) {
this._available = false
this._state = LLMState.Unavailable
this.emit('availability-changed', { available: false })
logger.warn('Ollama server unreachable')
}
}
}
// ── EventEmitter 타입 오버라이드 ───────────────────────
override on<K extends keyof LocalLLMEvents>(event: K, listener: LocalLLMEvents[K]): this {
return super.on(event, listener)
}
override off<K extends keyof LocalLLMEvents>(event: K, listener: LocalLLMEvents[K]): this {
return super.off(event, listener)
}
override emit<K extends keyof LocalLLMEvents>(
event: K,
...args: Parameters<LocalLLMEvents[K]>
): boolean {
return super.emit(event, ...args)
}
}
// ── 싱글톤 ─────────────────────────────────────────────
let instance: LocalLLMService | null = null
export function getLocalLLMService(): LocalLLMService {
if (!instance) {
instance = new LocalLLMService()
}
return instance
}