api_server analyze에서 query로 history 함께 입력가능하도록 수정 #31

api_server.py

@@ -3,7 +3,7 @@ import sys
 import threading
 from queue import Empty, Queue
 from threading import Thread
-from typing import Optional
+from typing import List, Optional, Union
 
 from fastapi import FastAPI
 from fastapi.encoders import jsonable_encoder
@@ -189,11 +189,52 @@ async def create_persona(request: PersonaRequest):
     return StreamingResponse(event_stream(), media_type="text/event-stream", headers=headers)
 
 class QueryRequest(BaseModel):
-    query: str
+    query: Union[str, List["ChatMessage"]]
     stream: bool = True
     persona_name: Optional[str] = None
 
 
+class ChatMessage(BaseModel):
+    role: str
+    content: str
+
+
+def _normalize_chat_role(role: str) -> str:
+    role = (role or "").strip().lower()
+    return role
+
+
+def _normalize_query_input(query_input):
+    if isinstance(query_input, str):
+        return query_input.strip(), []
+
+    if not isinstance(query_input, list):
+        return "", []
+
+    conversation = []
+    for message in query_input:
+        if isinstance(message, ChatMessage):
+            role = _normalize_chat_role(message.role)
+            content = (message.content or "").strip()
+        elif isinstance(message, dict):
+            role = _normalize_chat_role(message.get("role", ""))
+            content = (message.get("content", "") or "").strip()
+        else:
+            continue
+
+        if not role or not content:
+            continue
+        conversation.append({"role": role, "content": content})
+
+    current_user_query = ""
+    for message in reversed(conversation):
+        if message["role"] == "user":
+            current_user_query = message["content"]
+            break
+
+    return current_user_query, conversation
+
+
 def _sse(payload: dict) -> str:
     return f"data: {json.dumps(payload, ensure_ascii=False)}\n\n"
 
@@ -215,11 +256,14 @@ def _build_result_payload(result, stdout: str = "") -> dict:
 
 @app.post("/analyze/")
 async def analyze(request: QueryRequest):
-    query = (request.query or "").strip()
+    query, conversation = _normalize_query_input(request.query)
     stream = request.stream
 
     persona_name = (request.persona_name or "").strip() or None
 
+    if not query:
+        return JSONResponse(status_code=400, content={"error": "query 필드가 비어 있습니다."})
+
     if not stream:
         stdout_messages = []
 
@@ -241,6 +285,7 @@ async def analyze(request: QueryRequest):
         try:
             result = run_pipeline(
                 query,
+                conversation=conversation,
                 persona_name=persona_name,
                 status_callback=None,
                 stream_callback=None,
@@ -268,6 +313,7 @@ async def analyze(request: QueryRequest):
             try:
                 result = run_pipeline(
                     query,
+                    conversation=conversation,
                     persona_name=persona_name,
                     status_callback=on_status,
                     stream_callback=on_delta if stream else None,

llm/generator.py

@@ -24,6 +24,36 @@ def extract_response_text(resp):
     return "\n".join(texts)
 
 
+def _normalize_history_role(role):
+    role = (role or "").strip().lower()
+    if role in {"user", "assistant", "system"}:
+        return role
+    return None
+
+
+def _split_conversation_history(conversation, current_user_query):
+    if not conversation:
+        return []
+
+    last_user_index = -1
+    for i in range(len(conversation) - 1, -1, -1):
+        role = (conversation[i].get("role") or "").strip().lower()
+        content = (conversation[i].get("content") or "").strip()
+        if role == "user" and content == current_user_query:
+            last_user_index = i
+            break
+
+    history = conversation[:last_user_index] if last_user_index >= 0 else conversation
+    normalized_history = []
+    for message in history:
+        role = _normalize_history_role(message.get("role"))
+        content = (message.get("content") or "").strip()
+        if not role or not content:
+            continue
+        normalized_history.append({"role": role, "content": content})
+    return normalized_history
+
+
 def generate_search_keywords(client, user_query, intent):
     """LLM을 통해 구글 뉴스 검색어 리스트 생성"""
     language = intent.get("language", "ko")
@@ -110,7 +140,7 @@ def generate_persona(client, user_query):
     return None
 
 
-def build_full_prompt(user_query, context, intent, persona=None):
+def build_system_prompt(intent, persona=None):
     analysis_type = intent.get("analysis_type", "general")
     language      = intent.get("language", "ko")
     system_prompt = SYSTEM_PROMPTS.get(analysis_type, SYSTEM_PROMPTS["general"])
@@ -119,9 +149,6 @@ def build_full_prompt(user_query, context, intent, persona=None):
 
     if persona:
         system_prompt += f"""
-[사용자 질의]
-{user_query}
-
 [선택된 페르소나]
 이름: {persona.name}
 요약: {persona.summary}
@@ -133,26 +160,41 @@ def build_full_prompt(user_query, context, intent, persona=None):
         if persona.famous_quotes:
             system_prompt += f"\n대표 어록: {' / '.join(persona.famous_quotes)}"
 
-    full_prompt = f"""{system_prompt}
+    return system_prompt
+
+
+def build_analysis_input(user_query, context, intent, persona=None, conversation=None):
+    system_prompt = build_system_prompt(intent, persona=persona)
+    history = _split_conversation_history(conversation, user_query)
+
+    input_messages = [{"role": "system", "content": system_prompt}]
+    input_messages.extend(history)
+    input_messages.append(
+        {
+            "role": "user",
+            "content": f"""[현재 사용자 질의]
+{user_query}
 
 [수집된 시장 데이터]
-{context}"""
-    return full_prompt
+{context}""",
+        }
+    )
+    return input_messages
 
-def generate_analysis(client, user_query, context, intent, persona=None):
-    full_prompt = build_full_prompt(user_query, context, intent, persona)
+def generate_analysis(client, user_query, context, intent, persona=None, conversation=None):
+    analysis_input = build_analysis_input(user_query, context, intent, persona, conversation=conversation)
     LLM_MODEL_NAME = os.environ.get('LLM_MODEL_NAME')
     resp = client.responses.create(
         model=LLM_MODEL_NAME,
-        input=full_prompt,
+        input=analysis_input,
     )
 
     result = extract_response_text(resp)
     return result or "(분석 결과를 가져오지 못했습니다)"
 
-def generate_analysis_stream(client, user_query, context, intent, persona=None):
+def generate_analysis_stream(client, user_query, context, intent, persona=None, conversation=None):
    
-    full_prompt = build_full_prompt(user_query, context, intent, persona)
+    analysis_input = build_analysis_input(user_query, context, intent, persona, conversation=conversation)
     llm_model_name = os.environ.get("LLM_MODEL_NAME")
     print(f"[⑤] LLM 분석 스트리밍 생성 중 (Responses API, 모델: {llm_model_name})...")
 
@@ -166,7 +208,7 @@ def generate_analysis_stream(client, user_query, context, intent, persona=None):
     # SDK에 따라 stream API 형태가 다를 수 있어 create(stream=True) 기준으로 처리
     stream = client.responses.create(
         model=llm_model_name,
-        input=full_prompt,
+        input=analysis_input,
         stream=True,
     )
 

pipeline.py

@@ -69,7 +69,7 @@ def save_result_jsonl(result):
     with open(file_name, "a", encoding="utf-8") as f:
         f.write(json.dumps(ordered_data, ensure_ascii=False) + "\n")
 
-def pipeline(query, persona_name=None, status_callback=None, stream_callback=None, stream=True):
+def pipeline(query, conversation=None, persona_name=None, status_callback=None, stream_callback=None, stream=True):
     """
     파이프라인:
         ① 인텐트 파싱  (Chat Completions + Function Calling)
@@ -120,7 +120,14 @@ def pipeline(query, persona_name=None, status_callback=None, stream_callback=Non
         if stream:
             chunks = []
             print("[⑤] 스트리밍 응답 수신 중...")
-            for delta in generate_analysis_stream(client, query, context, intent, persona=persona):
+            for delta in generate_analysis_stream(
+                client,
+                query,
+                context,
+                intent,
+                persona=persona,
+                conversation=conversation,
+            ):
                 if not delta:
                     continue
                 chunks.append(delta)
@@ -130,7 +137,14 @@ def pipeline(query, persona_name=None, status_callback=None, stream_callback=Non
             response = "".join(chunks).strip() or "(분석 결과를 가져오지 못했습니다)"
         else:
             print("[⑤] 단일 응답 생성 중...")
-            response = generate_analysis(client, query, context, intent, persona=persona)
+            response = generate_analysis(
+                client,
+                query,
+                context,
+                intent,
+                persona=persona,
+                conversation=conversation,
+            )
             if response:
                 emit_delta(response)
 
@@ -201,13 +215,24 @@ def main():
         except ValueError:
             print("잘못된 입력입니다. 기본 모드로 진행합니다.")
 
+    conversation = []
+    print("\n멀티턴 대화 모드입니다. 이전 질문/답변이 다음 분석에 함께 반영됩니다.")
+    print("대화 초기화: reset 또는 clear | 종료: exit, quit, 종료")
+
     while True:
         text = input("\n질문 > ").strip()
         if text.lower() in ("exit", "quit", "종료"):
             break
+        if text.lower() in ("reset", "clear"):
+            conversation = []
+            print("대화 히스토리를 초기화했습니다.")
+            continue
         if not text:
             continue
-        result = pipeline(text, persona_name=persona_name)
+
+        current_conversation = conversation + [{"role": "user", "content": text}]
+        result = pipeline(text, conversation=current_conversation, persona_name=persona_name)
+        conversation = current_conversation + [{"role": "assistant", "content": result.llm_response}]
         print_result(result)