Upload 8 files

app.py

@@ -1,5 +1,6 @@
 import os
 import httpx
+import json  # <-- 新增导入
 from fastapi import FastAPI, Request, HTTPException
 from fastapi.responses import HTMLResponse, JSONResponse
 from fastapi.staticfiles import StaticFiles
@@ -69,13 +70,18 @@ async def call_deepseek_api(messages: list, model: str = "deepseek-chat", temper
             detail=error_msg
         )
 
-# --- New AI Agent Endpoint ---
+# --- NEW: AI Agent Endpoint (Extractor + Expert Model) ---
 @app.post("/generate")
 async def generate_content(request: Request):
     """
-    This endpoint uses a two-step LLM process:
-    1. Generate a high-quality prompt based on user data.
-    2. Use that prompt to generate the final content.
+    This endpoint uses a robust "Extractor + Template + Expert" pattern.
+    
+    1. (Optional) LLM 1 (Extractor): For tasks with messy user input (like 'resume'),
+       this step cleans and structures the data into a reliable JSON.
+    2. (Required) Human Template: We use a precise, human-written f-string template
+       to build the perfect prompt for the expert.
+    3. (Required) LLM 2 (Expert): This model receives the clean prompt and generates
+       the final, high-quality content.
     """
     try:
         body = await request.json()
@@ -85,41 +91,70 @@ async def generate_content(request: Request):
         if not task or not data:
             raise HTTPException(status_code=400, detail="Missing 'task' or 'data' in request body")
 
-        # Step 1: Generate a high-quality prompt for the main task
-        meta_system_prompt = "You are an expert prompt engineer. Your task is to create a detailed and effective 'user' prompt for another AI model, which is an expert career consultant. The generated prompt must guide the second AI to produce a comprehensive, high-quality response in the required format based on the user's raw data and task."
-        
-        meta_user_prompt = f"""
-        I have the following task and user data. Create the perfect prompt for a career consultant AI to handle this.
-
-        **Task:** '{task}'
-
-        **User's Raw Data:**
-        ```json
-        {data}
-        ```
-
-        **Instructions for the prompt you will generate:**
-        - The prompt must be self-contained and include all necessary user data.
-        - It must clearly state the desired output format. 
-        - For the 'resume' task, the format MUST be a single JSON object with two keys: "resume" and "analysis", both containing well-formed HTML.
-        - For all other tasks ('interview', 'learning_path', 'cover_letter', 'linkedin', 'salary'), the format MUST be well-formed HTML content directly.
-        - The tone of the prompt should be as if a user is asking an expert for help.
-        - Incorporate all the details from the user's data into the prompt naturally.
-        """
-
-        print(f"[INFO] Generating prompt for task: {task}")
-        generated_prompt = await call_deepseek_api(
-            messages=[
-                {"role": "system", "content": meta_system_prompt},
-                {"role": "user", "content": meta_user_prompt}
-            ],
-            temperature=0.3 # Lower temperature for more deterministic prompt generation
-        )
-        print(f"[DEBUG] Generated Prompt for 2nd LLM call:\n{generated_prompt}")
+        structured_data = data  # Default to original data
 
-        # Step 2: Use the generated prompt to get the final content
+        # --- Step 1: (Optional) LLM 1 (Extractor) ---
+        # We only run this for tasks where user input might be "messy"
+        if task == "resume":
+            print(f"[INFO] Task '{task}' requires data extraction. Running LLM 1 (Extractor)...")
+            
+            extractor_system_prompt = """
+            You are an expert data analyst. Your job is to extract and structure key information 
+            from a user's raw data for a resume. Pay close attention to the 'skills' field, 
+            which might be a messy, comma-separated list or natural language.
+            Your output MUST be a valid JSON object.
+            
+            Keep all fields from the original data, but add a new key 'skills_list' 
+            containing a clean Python-style list of skills extracted from the 'skills' field.
+            
+            Example Input Data:
+            { "name": "Alex", "skills": "i use react, js, and a bit of python. also project management", ... }
+            
+            Example Output JSON:
+            { "name": "Alex", "skills": "i use react, js, and a bit of python. also project management", "skills_list": ["React", "JavaScript", "Python (Beginner)", "Project Management"], ... }
+            """
+            
+            extractor_user_prompt = f"""
+            Please process the following raw user data and return ONLY a valid JSON object.
+            
+            Raw Data:
+            ```json
+            {json.dumps(data)}
+            ```
+            """
+            
+            try:
+                # --- LLM 1 Call ---
+                json_string_output = await call_deepseek_api(
+                    messages=[
+                        {"role": "system", "content": extractor_system_prompt},
+                        {"role": "user", "content": extractor_user_prompt}
+                    ],
+                    model="deepseek-chat", # Use a fast model
+                    temperature=0.1        # Low temp for high accuracy
+                )
+                
+                # Clean up potential markdown ```json ... ```
+                if "```json" in json_string_output:
+                    json_string_output = json_string_output.split("```json\n", 1)[1].split("```")[0]
+                
+                structured_data = json.loads(json_string_output)
+                print(f"[DEBUG] LLM 1 (Extractor) output: {structured_data}")
+                
+            except Exception as e:
+                print(f"[ERROR] LLM 1 (Extractor) failed: {e}. Falling back to raw data.")
+                # Fallback: If extraction fails, use the original data and do a simple split
+                structured_data = data
+                structured_data['skills_list'] = [skill.strip() for skill in data.get('skills', '').split(',')]
+
+        else:
+            print(f"[INFO] Task '{task}' does not require extraction. Using raw data.")
+            structured_data = data
+
+        # --- Step 2: Human-Written Templates ---
+        
         final_system_prompts = {
-            "resume": "You are a professional career consultant and resume expert. Please strictly follow the JSON format for the response, ensuring the HTML is well-formed and professional.",
+            "resume": "You are a professional career consultant and resume expert. Your task is to generate a JSON object with two keys: 'resume' (HTML content) and 'analysis' (HTML content). Please strictly follow the JSON format, ensuring all HTML is well-formed and professional.",
             "interview": "You are an experienced interviewer and career mentor. Provide practical, professional interview preparation materials in well-formed HTML.",
             "learning_path": "You are an experienced career mentor and learning planner. Create a personalized, actionable learning path in well-formed HTML.",
             "cover_letter": "You are an expert cover letter writer. Write a professional, persuasive, and personalized cover letter in well-formed HTML.",
@@ -128,12 +163,106 @@ async def generate_content(request: Request):
         }
         final_system_prompt = final_system_prompts.get(task, "You are a helpful AI career assistant.")
 
-        print(f"[INFO] Generating final content for task: {task}")
+        # --- Build Final User Prompt from Template ---
+        final_user_prompt = ""
+        
+        if task == "resume":
+            final_user_prompt = f"""
+            Please act as a resume expert. Create an optimized resume and a matching analysis based on the following structured data.
+            
+            **User Profile:**
+            - Name: {structured_data.get('name', 'N/A')}
+            - Current Role: {structured_data.get('currentRole', 'N/A')}
+            - Years of Experience: {structured_data.get('experience', 'N/A')}
+            - Cleaned Skills List: {structured_data.get('skills_list', 'N/A')}
+            
+            **Target Opportunity:**
+            - Job Title: {structured_data.get('jobTitle', 'N/A')}
+            - Company: {structured_data.get('company', 'N/A')}
+            - Job Description:
+              ```
+              {structured_data.get('jobDescription', 'N/A')}
+              ```
+            
+            **Required Output Format:**
+            You MUST return a single, valid JSON object with two keys: "resume" and "analysis".
+            Both keys must contain well-formed HTML content.
+            """
+        
+        elif task == "interview":
+            final_user_prompt = f"""
+            Please act as an interview coach. Generate interview questions based on this data.
+            
+            - Role: {structured_data.get('role', 'N/A')}
+            - Level: {structured_data.get('level', 'N/A')}
+            - Key Skills: {structured_data.get('skills', 'N/A')}
+            
+            **Required Output Format:**
+            A single block of well-formed HTML content.
+            """
+            
+        elif task == "learning_path":
+             final_user_prompt = f"""
+             Please act as a learning planner. Create a personalized learning path.
+             
+             - Current Skills: {structured_data.get('currentSkills', 'N/A')}
+             - Target Role: {structured_data.get('targetRole', 'N/A')}
+             - Timeline: {structured_data.get('timeline', 'N/A')} months
+             
+             **Required Output Format:**
+             A single block of well-formed HTML content, detailing a roadmap.
+             """
+
+        elif task == "cover_letter":
+            final_user_prompt = f"""
+            Please act as a cover letter writer. Write a letter based on these details.
+            
+            - Company: {structured_data.get('company', 'N/A')}
+            - Role: {structured_data.get('role', 'N/A')}
+            - Key Achievement: {structured_data.get('achievement', 'N/A')}
+            - Tone: {structured_data.get('tone', 'N/A')}
+            
+            **Required Output Format:**
+            A single block of well-formed HTML content, formatted as a letter.
+            """
+
+        elif task == "linkedin":
+            final_user_prompt = f"""
+            Please act as a LinkedIn expert. Optimize a profile based on this data.
+            
+            - Current Headline: {structured_data.get('headline', 'N/A')}
+            - Current About: {structured_data.get('about', 'N/A')}
+            - Target Industry/Roles: {structured_data.get('target', 'N/A')}
+            
+            **Required Output Format:**
+            A single block of well-formed HTML content with sections for "New Headline" and "New About Section".
+            """
+
+        elif task == "salary":
+            final_user_prompt = f"""
+            Please act as a salary analyst. Provide insights for the following role.
+            
+            - Role: {structured_data.get('role', 'N/A')}
+            - Location: {structured_data.get('location', 'N/A')}
+            - Experience: {structured_data.get('experience', 'N/A')} years
+            - Company Size: {structured_data.get('companySize', 'N/A')}
+            
+            **Required Output Format:**
+            A single block of well-formed HTML content, including an estimated range and negotiation tips.
+            """
+        else:
+            final_user_prompt = f"Please perform the task '{task}' with the data: {json.dumps(structured_data)}"
+
+        print(f"[DEBUG] Final User Prompt for LLM 2:\n{final_user_prompt[:500]}...") # Log first 500 chars
+
+        # --- Step 3: LLM 2 (Expert) Call ---
+        print(f"[INFO] Generating final content for task: {task} using LLM 2 (Expert)...")
         final_content = await call_deepseek_api(
             messages=[
                 {"role": "system", "content": final_system_prompt},
-                {"role": "user", "content": generated_prompt}
-            ]
+                {"role": "user", "content": final_user_prompt}
+            ],
+            temperature=0.7 # Standard temp for creative/expert output
         )
 
         return JSONResponse(content={"content": final_content})
@@ -152,9 +281,8 @@ async def generate_content(request: Request):
 @app.post("/call-deepseek")
 async def proxy_deepseek(request: Request):
     """
-    This endpoint receives the request from our frontend (index.html),
-    adds the secret API key, and forwards it to the DeepSeek API.
-    This is the original proxy endpoint and will be replaced by the /generate endpoint logic.
+    This endpoint is kept for legacy purposes but is not used by the
+    new /generate logic.
     """
     if not DEEPSEEK_API_KEY:
         print("[ERROR] DEEPSEEK_API_KEY is not set!")
@@ -211,4 +339,4 @@ async def read_root():
         return HTMLResponse(
             content="<h1>Error: index.html not found</h1><p>Ensure index.html is in a 'static' folder.", 
             status_code=404
-        )
+        )