Update app.py

app.py

@@ -7,18 +7,14 @@ import google.generativeai as genai
 from sentence_transformers import SentenceTransformer, util
 
 # ============================================================
-# CONFIG
+# CONFIG - DO NOT LEAK YOUR KEY!
 # ============================================================
-GEMINI_API_KEY = "AIzaSyBrbLGXkSdXReb0lUucYqcNCNBkvS-RBFw"
-if not GEMINI_API_KEY:
-    raise RuntimeError("Set GEMINI_API_KEY environment variable")
-
+# Best practice: use os.environ.get("GEMINI_API_KEY") 
+GEMINI_API_KEY = "AIzaSyBrbLGXkSdXReb0lUucYqcNCNBkvS-RBFw" 
 genai.configure(api_key=GEMINI_API_KEY)
 
-
-
-
-MODEL = genai.GenerativeModel("gemini-2.0-flash")
+# Use 1.5-flash for maximum stability on Free Tier
+MODEL = genai.GenerativeModel("gemini-1.5-flash") 
 
 DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
 EMBED_MODEL = "sentence-transformers/all-MiniLM-L6-v2"
@@ -29,130 +25,72 @@ embedder = SentenceTransformer(EMBED_MODEL, device=DEVICE)
 print("✅ Ready")
 
 # ============================================================
-# UTILS
+# OPTIMIZED PIPELINE (ONE CALL ONLY)
 # ============================================================
-def split_sentences(text):
-    return [s.strip() for s in re.split(r'(?<=[.!?])\s+', text) if len(s.strip()) > 5]
-
-def gemini(prompt, max_tokens=256):
-    response = MODEL.generate_content(
-        prompt,
-        generation_config=genai.types.GenerationConfig(
-            temperature=0.0,
-            max_output_tokens=max_tokens
-        )
-    )
-    return response.text.strip()
-
-def safe_json(text):
+def get_rubric_and_intent(kb, question):
+    """Combines intent detection and rubric generation to save API quota."""
+    prompt = f"""
+    You are an expert examiner. Analyze the provided Knowledge Base and Question.
+    
+    1. Classify the intent: FACTUAL, EXPLANATORY, CHARACTER_ARC, PROCESS, or COMPARISON.
+    2. Create a grading rubric of 3-6 atomic criteria based ONLY on the Knowledge Base.
+
+    Knowledge Base: {kb}
+    Question: {question}
+
+    STRICT JSON OUTPUT ONLY:
+    {{
+      "intent": "YOUR_LABEL",
+      "criteria": ["criterion 1", "criterion 2", ...]
+    }}
+    """
     try:
+        response = MODEL.generate_content(prompt)
+        # Handle potential safety blocks or empty responses
+        if not response.candidates or not response.candidates[0].content.parts:
+            return {"intent": "ERROR", "criteria": []}
+            
+        text = response.text.strip()
+        # Clean JSON if model adds markdown backticks
+        text = re.sub(r'^```json\s*|\s*```$', '', text, flags=re.MULTILINE)
         return json.loads(text)
-    except:
-        start, end = text.find("{"), text.rfind("}") + 1
-        if start != -1 and end != -1:
-            try:
-                return json.loads(text[start:end])
-            except:
-                return None
-    return None
-
-# ============================================================
-# STEP 1: INTENT
-# ============================================================
-def detect_intent(question):
-    prompt = f"""
-Classify the question intent. Choose ONE:
-FACTUAL, EXPLANATORY, CHARACTER_ARC, PROCESS, COMPARISON
-
-Question:
-{question}
-
-Output ONLY the label.
-"""
-    out = gemini(prompt, 20)
-    return out if out in {
-        "FACTUAL","EXPLANATORY","CHARACTER_ARC","PROCESS","COMPARISON"
-    } else "EXPLANATORY"
-
-# ============================================================
-# STEP 2: RUBRIC GENERATION
-# ============================================================
-def generate_rubric(kb, question, intent):
-    prompt = f"""
-You are an examiner.
-
-Using ONLY the knowledge base, create a grading rubric for the question.
-Each item must be an atomic idea a student must mention.
+    except Exception as e:
+        print(f"API Error: {e}")
+        return {"intent": "EXPLANATORY", "criteria": []}
 
-Rules:
-- 3 to 6 criteria
-- No paraphrasing the question
-- No explanations
-- Capture progression if relevant
-- STRICT JSON ONLY
-
-Format:
-{{ "criteria": ["criterion 1", "criterion 2"] }}
-
-Knowledge Base:
-{kb}
-
-Question:
-{question}
-
-Intent:
-{intent}
-"""
-    raw = gemini(prompt, 300)
-    parsed = safe_json(raw)
-    return parsed["criteria"] if parsed and "criteria" in parsed else []
+def evaluate(answer, question, kb):
+    # STEP 1: Get logic from Gemini (Single Call)
+    data = get_rubric_and_intent(kb, question)
+    intent = data.get("intent", "EXPLANATORY")
+    rubric = data.get("criteria", [])
+    
+    if not rubric:
+        return {"final_verdict": "⚠️ API ERROR: No rubric generated."}
+
+    # STEP 2: Semantic Matching (Local - No API cost)
+    sents = [s.strip() for s in re.split(r'(?<=[.!?])\s+', answer) if len(s.strip()) > 5]
+    if not sents:
+        return {"final_verdict": "❌ ANSWER TOO SHORT"}
 
-# ============================================================
-# STEP 3: SEMANTIC MATCHING
-# ============================================================
-def score(answer, criteria):
-    sents = split_sentences(answer)
     ans_emb = embedder.encode(sents, convert_to_tensor=True)
-
-    results = []
-    for crit in criteria:
+    scored = []
+    
+    for crit in rubric:
         crit_emb = embedder.encode(crit, convert_to_tensor=True)
         sims = util.cos_sim(crit_emb, ans_emb)[0]
         best = float(torch.max(sims)) if sims.numel() else 0.0
+        scored.append({"criterion": crit, "satisfied": best >= SIM_THRESHOLD})
 
-        results.append({
-            "criterion": crit,
-            "score": round(best, 3),
-            "satisfied": best >= SIM_THRESHOLD
-        })
-    return results
-
-# ============================================================
-# FINAL VERDICT
-# ============================================================
-def verdict(scored):
+    # STEP 3: Verdict
     hit = sum(c["satisfied"] for c in scored)
-    total = len(scored)
-
-    if hit == total:
-        return "✅ CORRECT"
-    if hit >= max(1, total // 2):
-        return "⚠️ PARTIALLY CORRECT"
-    return "❌ INCORRECT"
-
-# ============================================================
-# PIPELINE
-# ============================================================
-def evaluate(answer, question, kb):
-    intent = detect_intent(question)
-    rubric = generate_rubric(kb, question, intent)
-    scored = score(answer, rubric) if rubric else []
+    if hit == len(scored): verdict_text = "✅ CORRECT"
+    elif hit >= max(1, len(scored) // 2): verdict_text = "⚠️ PARTIALLY CORRECT"
+    else: verdict_text = "❌ INCORRECT"
 
     return {
         "intent": intent,
         "rubric": rubric,
-        "scoring": scored,
-        "final_verdict": verdict(scored) if rubric else "⚠️ NO RUBRIC"
+        "final_verdict": verdict_text
     }
 
 # ============================================================