Update app.py

app.py

@@ -1,58 +1,437 @@
-from transformers import AutoTokenizer, AutoModelForCausalLM
-import gradio as gr
-import torch
 
-model_name = "microsoft/phi-3-mini-4k-instruct"
+  import json
+  import re
+  import os
+  import gradio as gr
+  from huggingface_hub import hf_hub_download
+  from llama_cpp import Llama
 
-tokenizer = AutoTokenizer.from_pretrained(model_name)
+  # ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
+  # 1. LOAD QUANTIZED MODEL (Q4_K_M = ~2.4GB instead of ~15GB float32)
+  # ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
 
-model = AutoModelForCausalLM.from_pretrained(
-    model_name,
-    torch_dtype=torch.float32,
-    low_cpu_mem_usage=True
-)
+  MODEL_PATH = hf_hub_download(
+      repo_id="microsoft/Phi-3-mini-4k-instruct-gguf",
+      filename="Phi-3-mini-4k-instruct-q4.gguf",
+  )
 
-def generate_question(topic):
+  llm = Llama(
+      model_path=MODEL_PATH,
+      n_ctx=2048,       # context window (keep small for speed)
+      n_threads=2,      # HF free tier has 2 vCPUs
+      n_batch=64,       # batch size for prompt processing
+      verbose=False,
+  )
 
-    prompt = f"""
-You are a professional psychometric assessment AI.
+  # ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
+  # 2. JSON EXTRACTION HELPER
+  # ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
 
-Generate ONE psychology-based multiple choice question to understand a person's personality.
+  def extract_json(text: str) -> dict | None:
+      """Try multiple strategies to extract valid JSON from model output."""
+      # Strategy 1: Direct parse
+      try:
+          return json.loads(text.strip())
+      except json.JSONDecodeError:
+          pass
 
-Topic: {topic}
+      # Strategy 2: Find JSON block in markdown code fence
+      match = re.search(r"```(?:json)?\s*(\{.*?\})\s*```", text, re.DOTALL)
+      if match:
+          try:
+              return json.loads(match.group(1))
+          except json.JSONDecodeError:
+              pass
 
-Format:
-Question:
-A)
-B)
-C)
-D)
-"""
+      # Strategy 3: Find first { ... } block
+      match = re.search(r"\{.*\}", text, re.DOTALL)
+      if match:
+          try:
+              return json.loads(match.group(0))
+          except json.JSONDecodeError:
+              pass
 
-    inputs = tokenizer(prompt, return_tensors="pt")
+      return None
 
-    outputs = model.generate(
-        **inputs,
-        max_new_tokens=150,
-        temperature=0.6,
-        do_sample=True,
-        top_p=0.9
-    )
 
-    result = tokenizer.decode(outputs[0], skip_special_tokens=True)
+  # ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
+  # 3. PROMPT BUILDER (Phi-3 chat template)
+  # ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
 
-    # remove prompt
-    result = result.replace(prompt, "").strip()
+  def build_prompt(system: str, user: str) -> str:
+      """Build prompt using Phi-3's official ChatML template."""
+      return (
+          f"<|system|>\n{system}<|end|>\n"
+          f"<|user|>\n{user}<|end|>\n"
+          f"<|assistant|>\n"
+      )
 
-    return result
 
+  # ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
+  # 4. ENDPOINT 1: GENERATE PSYCHOMETRIC QUESTION
+  #    (Called by your Flutter AIController.generateNextPsychometricQuestion)
+  # ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
 
-demo = gr.Interface(
-    fn=generate_question,
-    inputs=gr.Textbox(label="Enter Topic"),
-    outputs=gr.Textbox(label="Generated Question"),
-    title="Psychometric AI",
-    description="AI generated psychology questions"
-)
+  DIMENSIONS = [
+      "PERSONALITY", "EMOTIONAL_INTELLIGENCE", "COGNITIVE_STYLE", "MOTIVATION",
+      "VALUES", "FUTURE_GOALS", "STRESS_COPING", "RELATIONSHIPS",
+      "DECISION_MAKING", "SELF_CONCEPT", "LIFESTYLE", "CREATIVITY",
+      "LEADERSHIP", "LEARNING_STYLE", "MEANING_PURPOSE", "HIDDEN_POTENTIAL",
+      "INTELLIGENCE_STYLE", "PASSION_MAPPING", "STRENGTH_WEAKNESS",
+      "LIFE_PHILOSOPHY", "CAREER_ALIGNMENT", "CROSS_DOMAIN",
+      "PEAK_EXPERIENCES", "NATURE",
+  ]
 
-demo.launch()
+  CATEGORY_ICONS = {
+      "PERSONALITY": "🎭", "EMOTIONAL_INTELLIGENCE": "❤️",
+      "COGNITIVE_STYLE": "🧠", "MOTIVATION": "🔥", "VALUES": "⚖️",
+      "FUTURE_GOALS": "🎯", "STRESS_COPING": "🛡️", "RELATIONSHIPS": "🤝",
+      "DECISION_MAKING": "⚡", "SELF_CONCEPT": "🪞", "LIFESTYLE": "🌿",
+      "CREATIVITY": "🎨", "LEADERSHIP": "👑", "LEARNING_STYLE": "📚",
+      "MEANING_PURPOSE": "🌟", "HIDDEN_POTENTIAL": "💎",
+      "INTELLIGENCE_STYLE": "💡", "PASSION_MAPPING": "🗺️",
+      "STRENGTH_WEAKNESS": "⚙️", "LIFE_PHILOSOPHY": "🔮",
+      "CAREER_ALIGNMENT": "💼", "CROSS_DOMAIN": "🔗",
+      "PEAK_EXPERIENCES": "🏔️", "NATURE": "🌱",
+  }
+
+  OPTION_ICONS = ["psychology", "lightbulb_outline", "people", "self_improvement"]
+
+
+  def get_dimension_for_question(q_num: int, explored: set) -> str:
+      """Pick the next dimension based on rotation strategy."""
+      # Rotation mapping (same as your Flutter code)
+      rotation = {
+          range(1, 4): ["PERSONALITY", "NATURE", "LIFESTYLE"],
+          range(4, 7): ["VALUES", "MOTIVATION", "FUTURE_GOALS"],
+          range(7, 10): ["EMOTIONAL_INTELLIGENCE", "RELATIONSHIPS"],
+          range(10, 13): ["DECISION_MAKING", "STRESS_COPING"],
+          range(13, 16): ["SELF_CONCEPT", "LEARNING_STYLE"],
+          range(16, 19): ["PASSION_MAPPING", "HIDDEN_POTENTIAL"],
+          range(19, 22): ["INTELLIGENCE_STYLE", "CREATIVITY"],
+          range(22, 25): ["STRENGTH_WEAKNESS", "LEADERSHIP"],
+          range(25, 28): ["CROSS_DOMAIN", "CAREER_ALIGNMENT"],
+          range(28, 31): ["LIFE_PHILOSOPHY", "MEANING_PURPOSE"],
+          range(31, 36): ["CAREER_ALIGNMENT"],
+          range(36, 41): ["PEAK_EXPERIENCES"],
+      }
+
+      for r, dims in rotation.items():
+          if q_num in r:
+              for d in dims:
+                  if d not in explored:
+                      return d
+              return dims[0]
+
+      # Fallback: pick any unexplored
+      for d in DIMENSIONS:
+          if d not in explored:
+              return d
+      return "PERSONALITY"
+
+
+  def generate_question(request_json: str) -> str:
+      """
+      Generate ONE adaptive psychometric question.
+
+      Input JSON: {
+        "question_number": 1,
+        "user_background": {...},
+        "previous_answers": [{"question_id":1,"category":"X","answer":"Y"}, ...],
+        "current_insight": "optional cumulative insight"
+      }
+
+      Output: Exact JSON matching Flutter's expected format.
+      """
+      try:
+          req = json.loads(request_json) if isinstance(request_json, str) else request_json
+      except json.JSONDecodeError:
+          req = {"question_number": 1, "user_background": {}, "previous_answers": []}
+
+      q_num = req.get("question_number", 1)
+      background = req.get("user_background", {})
+      prev_answers = req.get("previous_answers", [])
+      current_insight = req.get("current_insight", "")
+
+      # Determine which dimension to target
+      explored = {a.get("category", "").upper() for a in prev_answers}
+      target_dim = get_dimension_for_question(q_num, explored)
+      icon = CATEGORY_ICONS.get(target_dim, "⚙️")
+
+      # Build context from recent answers
+      recent = prev_answers[-3:] if len(prev_answers) > 3 else prev_answers
+      answers_text = "\n".join(
+          f"- Q{a.get('question_id','?')} [{a.get('category','?')}]: {a.get('answer','?')}"
+          for a in recent
+      ) or "None yet (first question)."
+
+      # Stage-aware context
+      stage = background.get("stageTitle", "Student")
+      age = background.get("ageRange", "unknown")
+
+      system_prompt = (
+          "You are an expert psychologist. Generate exactly ONE psychometric "
+          "multiple-choice question. Respond with ONLY a JSON object, no other text."
+      )
+
+      user_prompt = f"""Generate question #{q_num} for a psychometric assessment.
+
+  Target dimension: {target_dim}
+  User: {stage}, age {age}
+  Recent answers:
+  {answers_text}
+
+  Return ONLY this exact JSON structure:
+  {{"psychological_insight":"brief pattern summary from answers so
+  far","question":{{"id":{q_num},"category":"{target_dim}","category_icon":"{icon}","question":"A scenario-based
+  question about {target_dim.lower().replace('_',' ')}","subtitle":"brief
+  clarification","type":"multiple_choice","options":["behavioral choice 1","behavioral choice 2","behavioral choice
+  3","behavioral choice
+  4"],"option_icons":["psychology","lightbulb_outline","people","self_improvement"]}},"rationale":"why this
+  dimension matters now"}}"""
+
+      prompt = build_prompt(system_prompt, user_prompt)
+
+      output = llm(
+          prompt,
+          max_tokens=400,
+          temperature=0.3,
+          top_p=0.85,
+          repeat_penalty=1.15,
+          stop=["<|end|>", "<|user|>", "\n\n\n"],
+      )
+
+      raw = output["choices"][0]["text"]
+      parsed = extract_json(raw)
+
+      if parsed and "question" in parsed:
+          # Validate and fix the structure
+          q = parsed["question"]
+          q["id"] = q_num
+          q["category"] = q.get("category", target_dim)
+          q["category_icon"] = q.get("category_icon", icon)
+          q["type"] = "multiple_choice"
+          if "option_icons" not in q or len(q.get("option_icons", [])) != 4:
+              q["option_icons"] = OPTION_ICONS
+          if "options" not in q or len(q.get("options", [])) != 4:
+              # Fallback options
+              q["options"] = [
+                  f"I approach it analytically and systematically",
+                  f"I follow my intuition and feelings",
+                  f"I seek input from people I trust",
+                  f"I try something creative and unconventional",
+              ]
+          parsed["question"] = q
+          return json.dumps(parsed, ensure_ascii=False)
+
+      # ── FALLBACK: Build valid response if model output was unusable ──
+      fallback = {
+          "psychological_insight": current_insight or "Gathering initial data.",
+          "question": {
+              "id": q_num,
+              "category": target_dim,
+              "category_icon": icon,
+              "question": f"When facing a challenge related to {target_dim.lower().replace('_', ' ')}, what is your
+  first instinct?",
+              "subtitle": "Choose the response closest to your natural behavior.",
+              "type": "multiple_choice",
+              "options": [
+                  "Analyze the situation carefully before acting",
+                  "Trust my gut feeling and act quickly",
+                  "Discuss it with someone I trust first",
+                  "Look for a creative or unusual solution",
+              ],
+              "option_icons": OPTION_ICONS,
+          },
+          "rationale": f"Exploring {target_dim} dimension to build a complete profile.",
+      }
+      return json.dumps(fallback, ensure_ascii=False)
+
+
+  # ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
+  # 5. ENDPOINT 2: ANALYZE ASSESSMENT RESULTS
+  #    (Called by your Flutter AnalyzingTraitsController._fetchAnalysis)
+  # ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
+
+  def analyze_assessment(request_json: str) -> str:
+      """
+      Analyze completed assessment and return full results.
+
+      Input JSON: Full userData from PsychometricUserDataStorage
+      Output: Exact JSON matching Flutter's expected analysis format.
+      """
+      try:
+          user_data = json.loads(request_json) if isinstance(request_json, str) else request_json
+      except json.JSONDecodeError:
+          return json.dumps({"error": "Invalid input JSON"})
+
+      results = user_data.get("assessment_results", [])
+      total_q = len(results)
+      is_detailed = total_q >= 40
+      stage = user_data.get("stageTitle", "Student")
+      age = user_data.get("ageRange", "")
+      preferences = user_data.get("preferences", {})
+      dynamic_ctx = user_data.get("dynamicContext", {})
+      reasons = user_data.get("assessmentReason", [])
+
+      # Summarize answers for the prompt (keep it compact for 2048 context)
+      answer_summary = []
+      for r in results[:20]:  # Limit to 20 to fit context
+          answer_summary.append(
+              f"[{r.get('category','?')}] Q: {r.get('question','?')[:60]} → A: {r.get('answer','?')}"
+          )
+      answers_text = "\n".join(answer_summary)
+
+      system_prompt = (
+          "You are a psychologist generating a psychometric assessment report. "
+          "Respond with ONLY a JSON object matching the exact structure requested. No other text."
+      )
+
+      # For the small model, we ask for the FAST format only (simpler, fits in context)
+      user_prompt = f"""Analyze this psychometric assessment and return a JSON report.
+
+  User: {stage}, age {age}
+  Work preference: {preferences.get('workPreference','')}
+  Decision style: {preferences.get('decisionMaking','')}
+  Risk comfort: {preferences.get('riskComfort','')}
+  Interests: {dynamic_ctx.get('careerInterest','')}
+  Assessment reason: {', '.join(reasons) if reasons else 'Self discovery'}
+
+  Responses ({total_q} questions):
+  {answers_text}
+
+  Return ONLY this JSON (fill all values based on the responses above):
+  {{"summary":"2-3 sentence personality summary","overall_score":75,"overall_potential_description":"2 sentence
+  description of cognitive strengths","full_potential_analysis":"3 sentence detailed potential
+  analysis","career_objective":{{"title":"specific career goal","description":"why this fits","timeline":"2-3
+  years"}},"career_path":[{{"step":1,"title":"step name","description":"what to do","timeline":"0-6
+  months"}},{{"step":2,"title":"step name","description":"what to do","timeline":"6-12
+  months"}},{{"step":3,"title":"step name","description":"what to do","timeline":"1-2
+  years"}},{{"step":4,"title":"step name","description":"what to do","timeline":"2-3
+  years"}}],"trait_scores":{{"Leadership":70,"Creativity":75,"Logical Reasoning":80,"Emotional
+  Intelligence":72,"Analytical":78,"Social":68,"Strategic":74}},"radar":{{"ANALYTICAL":0.78,"SOCIAL":0.68,"EMOTIONAL
+  ":0.72,"STRATEGIC":0.74,"CREATIVE":0.75,"LOGICAL":0.80}},"strengths":["strength 1","strength 2","strength
+  3"],"growth_areas":[{{"title":"area","description":"how to
+  improve","icon":"schedule"}}],"career_matches":[{{"title":"career
+  option","category":"field","match_percentage":85,"reasoning":"why this matches","required_skills":["skill1","skill
+  2"],"growth_potential":80,"icon":"psychology"}}],"focus_improvements":[{{"title":"focus area","description":"what
+  to work on","icon":"pattern"}}],"historical_progress":{{"verbal_fluency":[70,72,75,78],"reaction_time":[180,175,17
+  0,165]}}}}"""
+
+      prompt = build_prompt(system_prompt, user_prompt)
+
+      output = llm(
+          prompt,
+          max_tokens=1200,
+          temperature=0.3,
+          top_p=0.85,
+          repeat_penalty=1.15,
+          stop=["<|end|>", "<|user|>", "\n\n\n"],
+      )
+
+      raw = output["choices"][0]["text"]
+      parsed = extract_json(raw)
+
+      if parsed and "overall_score" in parsed:
+          # Ensure all required fields exist with defaults
+          defaults = {
+              "summary": "Assessment complete.",
+              "overall_score": 70,
+              "overall_potential_description": "Shows balanced cognitive abilities.",
+              "full_potential_analysis": "The user demonstrates a mix of analytical and creative thinking.",
+              "career_objective": {"title": "Explore career options", "description": "Based on assessment.",
+  "timeline": "1-2 years"},
+              "career_path": [{"step": i+1, "title": f"Step {i+1}", "description": "Continue growing.", "timeline":
+  f"{i*6}-{(i+1)*6} months"} for i in range(4)],
+              "trait_scores": {"Leadership": 70, "Creativity": 70, "Logical Reasoning": 70, "Emotional
+  Intelligence": 70, "Analytical": 70, "Social": 70, "Strategic": 70},
+              "radar": {"ANALYTICAL": 0.7, "SOCIAL": 0.7, "EMOTIONAL": 0.7, "STRATEGIC": 0.7, "CREATIVE": 0.7,
+  "LOGICAL": 0.7},
+              "strengths": ["Adaptability", "Analytical thinking"],
+              "growth_areas": [{"title": "Communication", "description": "Practice public speaking.", "icon":
+  "schedule"}],
+              "career_matches": [{"title": "Analyst", "category": "General", "match_percentage": 75, "reasoning":
+  "Analytical strength.", "required_skills": ["Analysis"], "growth_potential": 75, "icon": "psychology"}],
+              "focus_improvements": [{"title": "Consistency", "description": "Build daily habits.", "icon":
+  "pattern"}],
+              "historical_progress": {"verbal_fluency": [70, 72, 75, 78], "reaction_time": [180, 175, 170, 165]},
+          }
+          for key, val in defaults.items():
+              if key not in parsed:
+                  parsed[key] = val
+
+          return json.dumps(parsed, ensure_ascii=False)
+
+      # ── FALLBACK: Return a valid but generic result ──
+      fallback_result = {
+          "summary": f"Based on {total_q} responses, this {stage.lower()} shows a balanced personality profile with
+  notable strengths in analytical and interpersonal areas.",
+          "overall_score": 72,
+          "overall_potential_description": "Demonstrates solid cognitive abilities with room for growth in
+  leadership and strategic thinking.",
+          "full_potential_analysis": f"As a {stage.lower()}, you show strong self-awareness and balanced
+  decision-making. Your responses indicate a preference for {preferences.get('decisionMaking', 'balanced')} thinking
+   with {preferences.get('riskComfort', 'moderate')} risk tolerance. Continued growth in cross-domain skills will
+  unlock significant potential.",
+          "career_objective": {
+              "title": f"Explore {dynamic_ctx.get('careerInterest', 'career')} opportunities",
+              "description": "Aligned with your interests and cognitive profile.",
+              "timeline": "Next 1-2 years",
+          },
+          "career_path": [
+              {"step": 1, "title": "Self-assessment", "description": "Identify core strengths", "timeline": "0-3
+  months"},
+              {"step": 2, "title": "Skill building", "description": "Develop key competencies", "timeline": "3-6
+  months"},
+              {"step": 3, "title": "Practical experience", "description": "Apply skills in real scenarios",
+  "timeline": "6-12 months"},
+              {"step": 4, "title": "Career positioning", "description": "Target specific opportunities", "timeline":
+   "1-2 years"},
+          ],
+          "trait_scores": {"Leadership": 68, "Creativity": 74, "Logical Reasoning": 76, "Emotional Intelligence":
+  72, "Analytical": 78, "Social": 70, "Strategic": 71},
+          "radar": {"ANALYTICAL": 0.78, "SOCIAL": 0.70, "EMOTIONAL": 0.72, "STRATEGIC": 0.71, "CREATIVE": 0.74,
+  "LOGICAL": 0.76},
+          "strengths": ["Analytical thinking", "Self-awareness", "Adaptability"],
+          "growth_areas": [{"title": "Leadership initiative", "description": "Take more lead in group settings.",
+  "icon": "schedule"}],
+          "career_matches": [
+              {"title": "Analyst", "category": "Technology", "match_percentage": 80, "reasoning": "Strong analytical
+   and logical traits.", "required_skills": ["Analysis", "Problem Solving"], "growth_potential": 78, "icon":
+  "psychology"}
+          ],
+          "focus_improvements": [{"title": "Strategic thinking", "description": "Practice long-term planning
+  exercises.", "icon": "pattern"}],
+          "historical_progress": {"verbal_fluency": [70, 72, 75, 78], "reaction_time": [180, 175, 170, 165]},
+      }
+      return json.dumps(fallback_result, ensure_ascii=False)
+
+
+  # ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
+  # 6. GRADIO APP WITH TWO API ENDPOINTS
+  # ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
+
+  with gr.Blocks(title="Psychometric AI") as demo:
+      gr.Markdown("# Psychometric Assessment AI\nGenerate questions & analyze results.")
+
+      with gr.Tab("Generate Question"):
+          q_input = gr.Textbox(
+              label="Request JSON",
+              placeholder='{"question_number": 1, "user_background": {}, "previous_answers": []}',
+              lines=5,
+          )
+          q_output = gr.Textbox(label="Generated Question (JSON)", lines=10)
+          q_btn = gr.Button("Generate")
+          q_btn.click(fn=generate_question, inputs=q_input, outputs=q_output, api_name="generate_question")
+
+      with gr.Tab("Analyze Assessment"):
+          a_input = gr.Textbox(
+              label="Full User Data JSON",
+              placeholder='{"stageTitle":"Student", "assessment_results": [...], ...}',
+              lines=10,
+          )
+          a_output = gr.Textbox(label="Analysis Result (JSON)", lines=15)
+          a_btn = gr.Button("Analyze")
+          a_btn.click(fn=analyze_assessment, inputs=a_input, outputs=a_output, api_name="analyze_assessment")
+
+  demo.launch()