Update app.py

app.py

@@ -409,7 +409,7 @@ class RAGIndex:
         return answer
 
     def answer(self, question: str) -> str:
-        """Answer a question using RAG with a generative seq2seq model (Flan-T5, BART, etc.)."""
+        """Answer a question using RAG with a simple extractive approach."""
         if not self.initialized:
             return "❌ Assistant not properly initialized. Please check the logs."
 
@@ -451,24 +451,45 @@ class RAGIndex:
         # 2) Combine contexts into a single evidence block
         combined_context = "\n\n".join(context_texts)
 
-        # Keep context at a reasonable size for the model
+        # Keep context at a reasonable size
         max_context_chars = 3000
         if len(combined_context) > max_context_chars:
             combined_context = combined_context[:max_context_chars]
 
-        # 3) Generate grounded answer from context
-        try:
-            answer_text = self._generate_from_context(
-                question=question,
-                context=combined_context,
-                max_new_tokens=180,
-            ).strip()
-        except Exception as e:
-            print(f"Generation error: {e}")
-            return (
-                "There was an error while generating the answer. "
-                "Please try again with a shorter question or different wording."
-            )
+        # 3) Sentence-level relevance scoring
+        #    We pick the sentences that best match the question terms
+        raw_sentences = re.split(r'(?<=[.!?])\s+', combined_context)
+        question_words = {
+            w.lower()
+            for w in re.findall(r"\w+", question)
+            if len(w) > 3  # ignore very short/common words
+        }
+
+        scored_sentences = []
+        for s in raw_sentences:
+            s_clean = s.strip()
+            if len(s_clean) < 20:
+                continue
+            words = {w.lower() for w in re.findall(r"\w+", s_clean)}
+            overlap = question_words & words
+            score = len(overlap)
+            scored_sentences.append((score, s_clean))
+
+        if scored_sentences:
+            # Sort by overlap score (descending)
+            scored_sentences.sort(key=lambda x: x[0], reverse=True)
+
+            # Take the best 2–3 sentences that have some overlap
+            top = [s for score, s in scored_sentences if score > 0][:3]
+
+            # If none have overlap (e.g., very vague question), just take the top 2 sentences overall
+            if not top:
+                top = [s for _, s in scored_sentences[:2]]
+
+            answer_text = " ".join(top)
+        else:
+            # Fallback: just take a slice of the combined context
+            answer_text = combined_context[:400].strip()
 
         if not answer_text:
             answer_text = NO_ANSWER_MSG
@@ -481,7 +502,6 @@ class RAGIndex:
         )
 
 
-
 # Initialize RAG system
 print("=" * 50)
 rag_index = RAGIndex()