Update app.py

app.py

@@ -61,13 +61,8 @@ def hash_key(kb, question):
 # ============================================================
 
 def generate_schema_with_llm(kb, question):
-    """
-    Uses a FREE HuggingFace LLM (Flan-T5) to generate grading schema.
-    HF-safe, CPU-safe, deterministic.
-    """
-
     prompt = f"""
-Extract the expected answer from the knowledge base.
+From the knowledge base below, answer the question using short factual points.
 
 Knowledge Base:
 {kb}
@@ -75,40 +70,35 @@ Knowledge Base:
 Question:
 {question}
 
-Return ONLY valid JSON in this format:
-
-{{
-  "question_type": "FACT",
-  "required_concepts": ["fact1", "fact2"],
-  "forbidden_concepts": [],
-  "allow_extra_info": true
-}}
+Write 1–3 short factual bullet points. Do NOT explain.
 """
 
     inputs = llm_tokenizer(prompt, return_tensors="pt", truncation=True).to(DEVICE)
 
     outputs = llm_model.generate(
         **inputs,
-        max_new_tokens=256,
+        max_new_tokens=128,
         temperature=0.0,
         do_sample=False
     )
 
     text = llm_tokenizer.decode(outputs[0], skip_special_tokens=True)
 
-    try:
-        json_text = text[text.find("{"):text.rfind("}") + 1]
-        return json.loads(json_text)
-    except Exception:
-        # HARD FAIL SAFE
-        return {
-            "question_type": "FACT",
-            "required_concepts": [],
-            "forbidden_concepts": [],
-            "allow_extra_info": True,
-            "error": "LLM schema parse failed",
-            "raw_output": text
-        }
+    # Extract bullet-like facts
+    facts = [
+        line.strip("-• ").strip()
+        for line in text.split("\n")
+        if len(line.strip()) > 3
+    ]
+
+    return {
+        "question_type": "FACT",
+        "required_concepts": facts,
+        "forbidden_concepts": [],
+        "allow_extra_info": True,
+        "raw_llm_output": text
+    }
+
 
 # ============================================================
 # ANSWER DECOMPOSITION
@@ -160,20 +150,34 @@ def evaluate_answer(answer, question, kb):
     logs["coverage"] = coverage
 
     # ---------------- CONTRADICTION CHECK ----------------
-    contradictions = []
-    kb_sents = split_sentences(kb)
-
-    for claim in claims:
-        for sent in kb_sents:
-            probs = softmax_logits(nli_model.predict([(sent, claim)]))
-            if probs[0] > CONTRADICTION_THRESHOLD:
-                contradictions.append({
-                    "claim": claim,
-                    "sentence": sent,
-                    "confidence": round(probs[0] * 100, 1)
-                })
-
-    logs["contradictions"] = contradictions
+    # contradictions = []
+    # kb_sents = split_sentences(kb)
+
+    # for claim in claims:
+    #     for sent in kb_sents:
+    #         probs = softmax_logits(nli_model.predict([(sent, claim)]))
+    #         if probs[0] > CONTRADICTION_THRESHOLD:
+    #             contradictions.append({
+    #                 "claim": claim,
+    #                 "sentence": sent,
+    #                 "confidence": round(probs[0] * 100, 1)
+    #             })
+
+    # logs["contradictions"] = contradictions
+
+
+relevant_kb = schema.get("required_concepts", [])
+
+for claim in claims:
+    for sent in relevant_kb:
+        probs = softmax_logits(nli_model.predict([(sent, claim)]))
+        if probs[0] > CONTRADICTION_THRESHOLD:
+            contradictions.append({
+                "claim": claim,
+                "sentence": sent,
+                "confidence": round(probs[0] * 100, 1)
+            })
+
 
     # ---------------- FINAL DECISION ----------------
     if contradictions: