feat: upgrade RAG conversational intent classification to multi-lingual LLM

RAG_FULL_APPLICATION_BACKEND/app/services/advanced_services.py

@@ -1,6 +1,7 @@
 from abc import ABC, abstractmethod
 from typing import List, Dict, Any, Tuple
 import re
+from ..services.llm_service import llm_service
 
 # ==========================================
 # 1. GRAPH RAG SERVICE (SOLID: Interface & Implementation)
@@ -70,7 +71,22 @@ class CorrectiveRAGService(ICorrectiveRAGService):
         else:
             max_similarity = max([c.get("similarity", 0.0) for c in retrieved_chunks] or [0.0])
             
-        is_low_confidence = max_similarity < 0.50
+        intent_prompt = f"""Analyze the following user query: "{query}"
+Determine if the query is EITHER:
+1. A conversational greeting (e.g., "hi", "namaste", "hello there", "namaskaram") OR
+2. A general summarization or document inquiry (e.g., "what is this document about?", "summarize this", "idi emiti")
+
+Return ONLY a JSON object:
+{{"is_conversational": true/false}}
+"""
+        try:
+            intent_res = llm_service.evaluate_json(intent_prompt)
+            is_conversational = intent_res.get("is_conversational", False)
+        except Exception:
+            # Fallback to simple heuristic if LLM fails
+            is_conversational = len(query.split()) < 3
+            
+        is_low_confidence = (max_similarity < 0.50) and not is_conversational
         
         if is_low_confidence:
             # Perform web search fallback simulation

RAG_FULL_APPLICATION_BACKEND/app/techniques/agentic_rag.py

@@ -96,6 +96,5 @@ Respond ONLY with a JSON object:
             return self.final_agent_answer
             
         await self.emit("GENERATE", "#7C3AED", "Qwen3 generating final summary...")
-        context = "\n\n".join([c["text"] for c in chunks])
-        prompt = f"Context:\n{context}\n\nQuestion: {query}\n\nAnswer based ONLY on the context:"
+        prompt = self.build_prompt(query, chunks)
         return self.llm.generate(prompt)

RAG_FULL_APPLICATION_BACKEND/app/techniques/base.py

@@ -15,6 +15,18 @@ class BaseRAGTechnique(ABC):
         self.supabase = supabase_service
         self.llm = llm_service
 
+    def build_prompt(self, query: str, chunks: List[Dict[str, Any]]) -> str:
+        context = "\n\n".join([c["text"] for c in chunks])
+        return f"""Context:
+{context}
+
+Question: {query}
+
+Instructions:
+1. If the user is just saying a general greeting (e.g., "hi", "hello"), respond politely without using the context.
+2. If the user asks a general question about the document itself (e.g., "what is this document about?", "summarize"), summarize the provided context to answer.
+3. Otherwise, answer the question based ONLY on the provided context. If the answer is not in the context, state that clearly."""
+
     @abstractmethod
     async def retrieve(self, query: str, document_id: str, top_k: int, **kwargs) -> List[Dict[str, Any]]:
         pass

RAG_FULL_APPLICATION_BACKEND/app/techniques/colbert.py

@@ -71,6 +71,5 @@ class ColBERT(BaseRAGTechnique):
 
     async def generate(self, query: str, chunks: List[Dict[str, Any]]) -> str:
         await self.emit("GENERATE", "#7C3AED", "Qwen3 generating answer...")
-        context = "\n\n".join([c["text"] for c in chunks])
-        prompt = f"Context:\n{context}\n\nQuestion: {query}\n\nAnswer based ONLY on the context:"
+        prompt = self.build_prompt(query, chunks)
         return self.llm.generate(prompt)

RAG_FULL_APPLICATION_BACKEND/app/techniques/hybrid_search.py

@@ -56,7 +56,6 @@ class HybridSearch(BaseRAGTechnique):
             "context_chunks": len(chunks),
             "temperature": 0.1
         })
-        context = "\n\n".join([c["text"] for c in chunks])
-        prompt = f"Context:\n{context}\n\nQuestion: {query}\n\nAnswer based ONLY on the context:"
+        prompt = self.build_prompt(query, chunks)
         return self.llm.generate(prompt)
 

RAG_FULL_APPLICATION_BACKEND/app/techniques/metadata_filter.py

@@ -33,6 +33,5 @@ class MetadataFilter(BaseRAGTechnique):
 
     async def generate(self, query: str, chunks: List[Dict[str, Any]]) -> str:
         await self.emit("GENERATE", "#7C3AED", "Qwen3 generating answer...")
-        context = "\n\n".join([c["text"] for c in chunks])
-        prompt = f"Context:\n{context}\n\nQuestion: {query}\n\nAnswer based ONLY on the context:"
+        prompt = self.build_prompt(query, chunks)
         return self.llm.generate(prompt)

RAG_FULL_APPLICATION_BACKEND/app/techniques/query_expansion.py

@@ -68,7 +68,6 @@ class QueryExpansion(BaseRAGTechnique):
             "llm": "Primary Qwen / Backup GLM-4.7-Flash",
             "chunks_used": len(chunks)
         })
-        context = "\n\n".join([c["text"] for c in chunks])
-        prompt = f"Context:\n{context}\n\nQuestion: {query}\n\nAnswer based ONLY on the context:"
+        prompt = self.build_prompt(query, chunks)
         return self.llm.generate(prompt)
 

RAG_FULL_APPLICATION_BACKEND/app/techniques/reranking.py

@@ -25,6 +25,5 @@ class ReRanking(BaseRAGTechnique):
 
     async def generate(self, query: str, chunks: List[Dict[str, Any]]) -> str:
         await self.emit("GENERATE", "#7C3AED", "Qwen3 generating answer...")
-        context = "\n\n".join([c["text"] for c in chunks])
-        prompt = f"Context:\n{context}\n\nQuestion: {query}\n\nAnswer based ONLY on the context:"
+        prompt = self.build_prompt(query, chunks)
         return self.llm.generate(prompt)

test_gemini.py

@@ -0,0 +1,10 @@
+import requests
+key = "YOUR_GEMINI_API_KEY"
+url = f"https://generativelanguage.googleapis.com/v1beta/models/gemini-3.1-flash-lite:generateContent?key={key}"
+payload = {
+    "contents": [{"parts": [{"text": "Hello"}]}],
+    "systemInstruction": {"parts": [{"text": ""}]}
+}
+res = requests.post(url, headers={'Content-Type': 'application/json'}, json=payload)
+print(res.status_code)
+print(res.text)

test_models.py

@@ -0,0 +1,50 @@
+from gradio_client import Client
+import time
+
+def test_hy3():
+    try:
+        print("Testing Tencent/Hy3...")
+        client = Client("tencent/Hy3")
+        result = client.predict(
+            message="Hello",
+            system_prompt="",
+            history=None,
+            think_level="high",
+            temperature=0.7,
+            max_tokens=100,
+            top_p=0.8,
+            functions_json_str="",
+            api_name="/chat"
+        )
+        print("Hy3 Response:", str(result)[:200])
+        return True
+    except Exception as e:
+        print("Hy3 Error:", e)
+        return False
+
+def test_qwen_omni():
+    try:
+        print("\nTesting Qwen3.5-Omni-Offline-Demo...")
+        client = Client("Qwen/Qwen3.5-Omni-Offline-Demo")
+        client.predict(api_name="/clear_history_offline")
+        result = client.predict(
+            text="Hello",
+            audio=None,
+            image=None,
+            video=None,
+            history=[],
+            system_prompt="",
+            temperature=0.7,
+            top_p=0.8,
+            top_k=20,
+            api_name="/chat_predict"
+        )
+        print("Qwen Omni Response:", str(result)[:200])
+        return True
+    except Exception as e:
+        print("Qwen Omni Error:", e)
+        return False
+
+if __name__ == "__main__":
+    test_hy3()
+    test_qwen_omni()

test_rag_doc.txt

@@ -0,0 +1 @@
+This is a test document for RAG ingestion. The sky is blue and water is wet.