Merge pull request #229 from Team-StackUp/fix/flash-model-default

fix(ai): flash 모델 기본값을 게이트웨이에 실재하는 gemini-3.5-flash-lite 로

Author: i3months

.env.example

@@ -28,7 +28,7 @@ LLM_API_KEY=
 LLM_BASE_URL=https://factchat-cloud.mindlogic.ai/v1/gateway
 LLM_PRO_MODEL=gemini-3.1-pro-preview
 LLM_PRO_TIMEOUT_SEC=30.0
-LLM_FLASH_MODEL=gemini-3.1-flash-lite
+LLM_FLASH_MODEL=gemini-3.5-flash-lite
 LLM_FLASH_TEMPERATURE=0.4
 LLM_FLASH_MAX_TOKENS=512
 LLM_FLASH_TIMEOUT_SEC=10.0

ai/.env.example

@@ -19,7 +19,7 @@ LLM_BASE_URL=https://factchat-cloud.mindlogic.ai/v1/gateway
 LLM_PRO_MODEL=gemini-3.1-pro-preview
 LLM_PRO_TEMPERATURE=0.2
 LLM_PRO_TIMEOUT_SEC=30.0
-LLM_FLASH_MODEL=gemini-3.1-flash-lite
+LLM_FLASH_MODEL=gemini-3.5-flash-lite
 LLM_FLASH_TEMPERATURE=0.4
 LLM_FLASH_MAX_TOKENS=512
 LLM_FLASH_TIMEOUT_SEC=10.0

ai/CLAUDE.md

@@ -131,7 +131,7 @@ async def consume_resume_analyze(message: AbstractIncomingMessage) -> None:
 | 시점 | 모델 (env override 가능) | 용도 |
 |------|--------------------------|------|
 | 세션 시작 | Pro (`gemini-3.1-pro-preview` 기본) | 질문 풀 (품질) |
-| 세션 중 | Flash (`gemini-3.1-flash-lite` 기본) | 꼬리질문 (저지연 < 3s) |
+| 세션 중 | Flash (`gemini-3.5-flash-lite` 기본) | 꼬리질문 (저지연 < 3s) |
 | 분석 (이력서/레포) | Pro | 마크다운 구조화 |
 
 설정은 `settings.py` + 환경변수로 모델명 주입 (코드에 하드코딩 금지).
@@ -223,7 +223,7 @@ class Settings(BaseSettings):
     openai_api_key: str = ""
     google_api_key: str = ""
     llm_pro_model: str = "gemini-3.1-pro-preview"
-    llm_flash_model: str = "gemini-3.1-flash-lite"
+    llm_flash_model: str = "gemini-3.5-flash-lite"
     embedding_model: str = "gemini-embedding-001"
     embedding_dim: int = 1536
     core_server_base_url: str = "http://core:8080"

ai/src/ai_server/config/settings.py

@@ -110,7 +110,7 @@ class Settings(BaseSettings):
     llm_pro_timeout_sec: float = 30.0
 
     # 꼬리질문용 Flash 모델 (저지연 < 3s)
-    llm_flash_model: str = "gemini-3.1-flash-lite"
+    llm_flash_model: str = "gemini-3.5-flash-lite"
     llm_flash_temperature: float = 0.4
     llm_flash_max_tokens: int = 512
     # Flash 는 저지연 요구사항이 있어 Pro 보다 짧게.

docker-compose.yml

@@ -87,7 +87,7 @@ services:
       LLM_API_KEY: ${LLM_API_KEY:-}
       LLM_BASE_URL: ${LLM_BASE_URL:-https://factchat-cloud.mindlogic.ai/v1/gateway}
       LLM_PRO_MODEL: ${LLM_PRO_MODEL:-gemini-3.1-pro-preview}
-      LLM_FLASH_MODEL: ${LLM_FLASH_MODEL:-gemini-3.1-flash-lite}
+      LLM_FLASH_MODEL: ${LLM_FLASH_MODEL:-gemini-3.5-flash-lite}
       LLM_FLASH_TEMPERATURE: ${LLM_FLASH_TEMPERATURE:-0.4}
       LLM_FLASH_MAX_TOKENS: ${LLM_FLASH_MAX_TOKENS:-512}
       STT_PROVIDER: ${STT_PROVIDER:-auto}

docs/api-conventions.md

@@ -442,7 +442,7 @@ AI의 LLM 호출별 토큰·지연시간을 `ai_request_logs`에 기록 (US-30 
   "userId": 123,
   "sessionId": 99,
   "requestType": "generate.followup",
-  "modelName": "gemini-3.1-flash",
+  "modelName": "gemini-3.5-flash-lite",
   "inputTokens": 1820,
   "outputTokens": 210,
   "latencyMs": 1830,

docs/environment.md

@@ -106,7 +106,7 @@ LLM_BASE_URL=https://factchat-cloud.mindlogic.ai/v1/gateway
 LLM_PRO_MODEL=gemini-3.1-pro-preview
 LLM_PRO_TEMPERATURE=0.2
 LLM_PRO_TIMEOUT_SEC=30.0          # 요청 타임아웃(초). 미설정 시 SDK 기본값까지 무기한 대기 위험
-LLM_FLASH_MODEL=gemini-3.1-flash-lite   # 꼬리질문(US-19) 저지연 모델
+LLM_FLASH_MODEL=gemini-3.5-flash-lite   # 꼬리질문(US-19) 저지연 모델
 LLM_FLASH_TEMPERATURE=0.4
 LLM_FLASH_MAX_TOKENS=512
 LLM_FLASH_TIMEOUT_SEC=10.0        # Pro 보다 짧게 — 꼬리질문 저지연(<3s) 요구사항

docs/observability.md

@@ -105,7 +105,7 @@ async def trace_middleware(request, call_next):
 
 ```
 request_type    예: 'session.followup', 'resume.analyze'
-model_name      예: 'gemini-3.1-pro', 'gemini-3.1-flash', 'whisper-1'
+model_name      예: 'gemini-3.1-pro-preview', 'gemini-3.5-flash-lite', 'whisper-1'
 input_tokens    토큰 카운트
 output_tokens
 latency_ms