Update app.py

app.py

@@ -173,10 +173,11 @@ def load_kb_documents(kb_dir: str) -> List[Tuple[str, str]]:
 
 def clean_context_text(text: str) -> str:
     """
-    Clean raw document context before sending to the generator:
+    Clean raw document context before sending to the answer builder:
     - Remove markdown headings (#, ##, ###)
     - Remove list markers (1., 2), -, *)
     - Remove duplicate lines
+    - Remove title-like lines (e.g. 'Knowledge Base Structure & Information Architecture Best Practices')
     """
     lines = text.splitlines()
     cleaned = []
@@ -200,6 +201,14 @@ def clean_context_text(text: str) -> str:
         if len(l) < 5:
             continue
 
+        # Heuristic: skip "title-like" lines where almost every word is capitalized
+        words = l.split()
+        if words:
+            cap_words = sum(1 for w in words if w[:1].isupper())
+            if len(words) <= 10 and cap_words >= len(words) - 1:
+                # Looks like a heading / title, skip it
+                continue
+
         # Avoid exact duplicates
         if l in seen:
             continue
@@ -409,7 +418,7 @@ class RAGIndex:
         return answer
 
     def answer(self, question: str) -> str:
-        """Answer a question using RAG with a simple extractive approach."""
+        """Answer a question using RAG with a simple extractive approach from the best chunk only."""
         if not self.initialized:
             return "❌ Assistant not properly initialized. Please check the logs."
 
@@ -432,33 +441,24 @@ class RAGIndex:
                 f"💡 Try rephrasing your question or check if relevant documents exist in the knowledge base."
             )
 
-        used_sources = set()
-        context_texts = []
+        # Use ONLY the single best scoring context (top-1)
+        best_ctx, best_source, best_score = contexts[0]
+        used_sources = {best_source}
 
-        # Clean and collect the retrieved chunks
-        for ctx, source, score in contexts:
-            used_sources.add(source)
-            cleaned_ctx = clean_context_text(ctx)
-            if cleaned_ctx:
-                context_texts.append(cleaned_ctx)
-
-        if not context_texts:
+        cleaned_ctx = clean_context_text(best_ctx)
+        if not cleaned_ctx:
             return (
                 f"{NO_ANSWER_MSG}\n\n"
                 f"💡 Try adding more detailed documents to the knowledge base."
             )
 
-        # 2) Combine contexts into a single evidence block
-        combined_context = "\n\n".join(context_texts)
-
-        # Keep context at a reasonable size
-        max_context_chars = 3000
-        if len(combined_context) > max_context_chars:
-            combined_context = combined_context[:max_context_chars]
+        # 2) Limit context size just in case
+        max_context_chars = 1500
+        if len(cleaned_ctx) > max_context_chars:
+            cleaned_ctx = cleaned_ctx[:max_context_chars]
 
-        # 3) Sentence-level relevance scoring
-        #    We pick the sentences that best match the question terms
-        raw_sentences = re.split(r'(?<=[.!?])\s+', combined_context)
+        # 3) Sentence-level relevance scoring within this single chunk
+        raw_sentences = re.split(r'(?<=[.!?])\s+|\n+', cleaned_ctx)
         question_words = {
             w.lower()
             for w in re.findall(r"\w+", question)
@@ -482,14 +482,14 @@ class RAGIndex:
             # 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 none have overlap (e.g., 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()
+            # Fallback: just take a slice of the cleaned context
+            answer_text = cleaned_ctx[:400].strip()
 
         if not answer_text:
             answer_text = NO_ANSWER_MSG