Update src/qa.py

src/qa.py

@@ -182,34 +182,35 @@ def cache_embeddings(file_name: str, chunks, embed_func, chunk_size: int = None,
     return embeddings
 
 # ==========================================================
-# 6️⃣ Prompt Templates (Enhanced for Structured Formatting)
+# 6️⃣ Prompt Templates (update to forbid internal metadata)
 # ==========================================================
-
 STRICT_PROMPT = (
     "You are an enterprise documentation assistant.\n"
     "Use all relevant information from the CONTEXT below.\n"
-    "When multiple causes, steps, or key points are discussed, present them as short, well-structured bullet points.\n"
-    "When the answer focuses on a single concept, definition, or explanation, write it as a clear and compact paragraph.\n"
-    "Keep the tone professional and concise. Do not invent facts outside the provided content.\n"
+    "If multiple related points appear across chunks, combine them into one clear answer.\n"
+    "Keep the answer concise but complete. Do not invent facts outside the provided content.\n"
+    "Do NOT mention any internal retrieval metadata, chunk identifiers, vector IDs, or system-level labels "
+    "(for example, do NOT say 'Chunk 7' or 'chunk 3:'). If you need to refer to the document, say "
+    "'the document' or quote the excerpt directly.\n"
     "If the answer cannot be found even after considering all chunks, say exactly:\n"
     "'I don't know based on the provided document.'\n\n"
     "Context:\n{context}\n\nQuestion: {query}\nAnswer:"
-
 )
 
-
 REASONING_PROMPT = (
     "You are an expert enterprise assistant capable of reasoning.\n"
     "Think step by step and synthesize information even if scattered across chunks.\n"
     "Base your answer primarily on the CONTEXT, but if multiple partial clues exist, combine them logically.\n"
     "You may fill reasonable gaps with general knowledge to form a complete answer.\n"
+    "Do NOT mention any internal retrieval metadata, chunk identifiers, vector IDs, or system-level labels "
+    "(for example, do NOT say 'Chunk 7' or 'chunk 3:'). If you need to reference source text, say 'the document' "
+    "or provide a short quoted excerpt.\n"
     "If absolutely nothing in the document relates, say exactly:\n"
     "'I don't know based on the provided document.'\n\n"
     "Context:\n{context}\n\nQuestion: {query}\nLet's reason step-by-step:\nAnswer:"
 )
 
 
-
 # ==========================================================
 # 7️⃣ Retrieval — FAISS + Bullet-Aware Re-rank + Neighbor Fill
 # ==========================================================