매뉴얼 등록 분리

manual_offline.py

@@ -0,0 +1,126 @@
+import os
+
+import torch
+from sentence_transformers import SentenceTransformer
+import chromadb
+from chromadb.utils import embedding_functions
+from transformers import pipeline, AutoTokenizer, AutoModelForCausalLM
+from fastapi import FastAPI
+
+# # === 경로 설정 (모두 로컬) ===
+QWEN_MODEL_PATH = "./models/Qwen3-0.6B"
+EMBEDDING_MODEL_PATH = "./models/all-MiniLM-L6-v2"
+
+# 2. 벡터 DB 설정
+persist_directory = "./chroma_db"
+chroma_client = chromadb.PersistentClient(path=persist_directory)
+
+collection = chroma_client.get_or_create_collection(
+    name="manuals",
+)
+
+_model = None
+_tokenizer = None
+
+def get_qwen_model() :
+    global _model, _tokenizer
+    if _model is None:
+        model_name = "Qwen/Qwen3-0.6B"
+        _tokenizer = AutoTokenizer.from_pretrained(
+            QWEN_MODEL_PATH,
+            trust_remote_code=True,
+            local_files_only=True       # 🔒 오프라인 강제
+        )
+        _model = AutoModelForCausalLM.from_pretrained(
+            QWEN_MODEL_PATH,
+            torch_dtype=torch.float32,  # CPU 안정성
+            device_map="auto",
+            trust_remote_code=True,
+            local_files_only=True       # 🔒 오프라인 강제
+        )
+    return _model, _tokenizer
+
+# 3. 질의 처리
+def query_and_summarize(job: str, query: str, top_k: int = 3):
+
+    # 관련 문서 검색
+    results = collection.query(
+        query_texts=[query],
+        n_results=5,
+        where={"dept": job}
+    )
+    # results = collection.query(query_texts=[query], n_results=top_k)
+    cosine_similarities = [1 - d for d in results['distances'][0]]
+    print("유사도:", cosine_similarities)
+    # 출력 예: [0.610, 0.473, 0.154, 0.142]
+
+    context_with_score = ""
+    for i, (doc, dist) in enumerate(zip(results['documents'][0], results['distances'][0])):
+        sim = 1 - dist
+        context_with_score += f"[문서 {i+1} | 유사도: {sim:.3f}]\n{doc}\n\n"
+
+    print(context_with_score)
+    print("\n\n\n\n\n")
+    top_doc = results['documents'][0][0]
+
+    # ✅ 명시적으로 모델과 토크나이저 로드
+    model, tokenizer = get_qwen_model()
+
+    messages = [
+        {"role": "system", "content": "당신은 회사 재무/회계 업무 전문 어시스턴트입니다. 문서 내용은 그대로 사용자에게 보여 줘야 하며 이를 기반으로 부가설명을 정확하고 상세하게 답변하세요."},
+        {"role": "user", "content": f"다음 문서를 참고하세요:\n{top_doc}\n\n질문: {query}"}
+    ]
+
+    text = tokenizer.apply_chat_template(
+        messages,
+        tokenize=False,
+        add_generation_prompt=True,
+        enable_thinking=True # Switches between thinking and non-thinking modes. Default is True.
+    )
+    model_inputs = tokenizer([text], return_tensors="pt").to(model.device)
+
+    # conduct text completion
+    generated_ids = model.generate(
+        **model_inputs,
+        max_new_tokens=500
+    )
+    output_ids = generated_ids[0][len(model_inputs.input_ids[0]):].tolist()
+
+    # parsing thinking content
+    try:
+        # rindex finding 151668 (</think>)
+        index = len(output_ids) - output_ids[::-1].index(151668)
+    except ValueError:
+        index = 0
+
+    thinking_content = tokenizer.decode(output_ids[:index], skip_special_tokens=True).strip("\n")
+    end_think_id = tokenizer.convert_tokens_to_ids("</think>")
+    if end_think_id in output_ids:
+        idx = len(output_ids) - output_ids[::-1].index(end_think_id)
+    else:
+        idx = 0
+    content = tokenizer.decode(output_ids[idx:], skip_special_tokens=True).strip()
+
+    print(top_doc)
+    print("\n\n\n\n\n")
+    print("thinking content:", thinking_content)
+    print("\n\n\n\n\n")
+    return content
+
+app = FastAPI()
+
+@app.get("/")
+def question(query: str) :
+    print(1)
+    answer = query_and_summarize(job="FI", query=query)
+    return {"answer": answer}
+
+# 예시 사용
+if __name__ == "__main__":
+    # init(job="FI")
+    # FI : 재무 HR : 인사
+    print(1)
+    # user_query = "외화 송금 방법?"
+    # answer = query_and_summarize(job="FI", query=user_query)
+    # print(answer)
+    # 실행방법 uvicorn manual:app --reload