fix: route conversational queries early in stream_answer to bypass RAG retrieval and ensure correct greeting/identity responses

app/services/engine.py

@@ -514,6 +514,45 @@ async def stream_answer(
             yield {"type": "done"}
             return
 
+        # ── Step 1.2: Early Conversational Query Check ──────────────────────────────
+        if is_conversational_query(query):
+            logger.info("Conversational query detected: '%s'", query[:80])
+            engine_mode = "hybrid" if is_online else "fallback"
+            yield {"type": "status", "engine_mode": engine_mode}
+            yield {"type": "citations", "citations": [], "is_grounded": True}
+            if is_online:
+                conv_prompt = [
+                    {
+                        "role": "system",
+                        "content": (
+                            "You are ByteAstra, a helpful AI tutoring assistant for BAMS (Bachelor of "
+                            "Ayurvedic Medicine and Surgery) students, developed by the engineering team at Risu Solutions.\n\n"
+                            "Your purpose is to guide students through the BAMS syllabus and classical Ayurvedic texts "
+                            "(like Charaka Samhita, Sushruta Samhita, and Ashtanga Hridaya) by providing detailed, exam-quality answers with verified citations.\n\n"
+                            "Introduce yourself as ByteAstra, created by Risu Solutions, and explain your purpose. "
+                            "Keep your response brief, friendly, and polite."
+                        )
+                    }
+                ]
+                conv_prompt.extend(history[-6:])
+                conv_prompt.append({"role": "user", "content": query})
+
+                async for token in llm.stream_completion(
+                    conv_prompt,
+                    model=llm_cfg.get("model_name"),
+                    temperature=0.7,
+                    max_tokens=128,   # conversational — keep short
+                ):
+                    yield {"type": "delta", "content": token}
+            else:
+                fallback_msg = get_conversational_response_fallback(query)
+                for word in fallback_msg.split(" "):
+                    yield {"type": "delta", "content": word + " "}
+                    await asyncio.sleep(0.01)
+
+            yield {"type": "done"}
+            return
+
         # Compute query embeddings for cache and search
         embedding_fn = rag._get_embedding_fn()
         query_embeddings = await asyncio.to_thread(embedding_fn, [query])