Adjust system prompt

Author: edkaya

genai/controller/generate_controller.py

@@ -78,15 +78,15 @@ def upload_file():
 
 @generate_bp.route('/genai/generate', methods=['POST'])
 def generate():
-    """API Endpoint for generating recipe responses based on document retrieval.
-    
-    This endpoint processes user queries against a vector database of recipes and returns
-    AI-generated responses using retrieved context.
-    
+    """API Endpoint for generating recipe response based on document retrieval
+
+    This endpoint processes user queries against a vector database of recipes 
+    and returns AI-generated responses using retrieved context.
+
     Request Body:
         query (str): The user's recipe-related query
         conversation_id (str): Unique identifier for the conversation thread
-    
+
     Returns:
         JSON response containing the generated recipe response or error message
     """
@@ -96,7 +96,7 @@ def generate():
         return jsonify({"error": "Missing 'query' or 'conversation_id'"}), 400
 
     query = data["query"]
-    #conversation_id = data["conversation_id"] # will be used in the future
+    # conversation_id = data["conversation_id"] # will be used in the future
 
     try:
         collection_name = "recipes"
@@ -106,10 +106,15 @@ def generate():
             vector_store = qdrant.create_and_get_vector_storage(
                 collection_name
             )
-            #todo: retrieve messages from chat history as BaseMessage
+            # todo: retrieve messages from chat history as BaseMessage
             messages = []
             retrieved_docs = retrieve_similar_docs(vector_store, query)
-            prompt = prepare_prompt(query, retrieved_docs, messages)
+            prompt = prepare_prompt(
+                llm.get_system_prompt(),
+                query,
+                retrieved_docs,
+                messages
+                )
 
             response = llm.invoke(prompt)
 

genai/rag/llm/chat_model.py

@@ -10,24 +10,39 @@
 class ChatModel(BaseChatModel):
     model_name: str = Field(default="llama3.3:latest")
 
-    def _generate(self, messages: List[BaseMessage], stop=None, **kwargs) -> ChatResult:
-        prompt = "\n".join([msg.content for msg in messages if isinstance(msg, HumanMessage)])
+    def _generate(self, messages: List[BaseMessage],
+                  stop=None, 
+                  **kwargs) -> ChatResult:
+        prompt = "\n".join([
+            msg.content for msg in messages if isinstance(msg, HumanMessage)
+            ])
         response_text = generate_response(self.model_name, prompt)
 
         return ChatResult(
-            generations=[ChatGeneration(message=AIMessage(content=response_text))]
+            generations=[
+                ChatGeneration(message=AIMessage(content=response_text))
+                ]
         )
 
     @property
     def _llm_type(self) -> str:
         return "recipai-custom-model"
 
-# For Testing purposes 
-# if __name__ == "__main__":
-#     llm = ChatModel(model_name="llama3.3:latest")
-
-#     message = HumanMessage(content="What is langchain, explain very briefly?")
-
-#     response = llm.invoke([message])
-
-#     print("LLM response:\n", response.content)
+    def get_system_prompt(self) -> str:
+        """System prompt for the LLM"""
+        return """
+            You are RecipAI, an intelligent assistant that helps users discover and generate 
+            recipes based on the ingredients they provide. 
+            
+            Use the contextual information provided below to tailor your responses. 
+            
+            If relevant recipes or suggestions are found in the context, prioritize 
+            those. If no relevant context is available, use your own knowledge to help the user.
+
+            Context:
+            {context}
+
+            Be clear, creative, and helpful. If the user also asks follow-up 
+            questions (e.g., dietary adjustments, name references, meal timing), 
+            answer them precisely based on the context and query.
+            """

genai/service/openwebui_service.py

@@ -4,8 +4,9 @@
 
 BASE_URL = "https://gpu.aet.cit.tum.de/"
 
+
 def generate_response(model_name: str, prompt: str):
-    """Making a POST request to the respective endpoint for 
+    """Making a POST request to the respective endpoint for
     response generation by an LLM"""
     url = f"{BASE_URL}/api/chat/completions"
 
@@ -18,7 +19,7 @@ def generate_response(model_name: str, prompt: str):
         "model": model_name,
         "messages": [
             {
-                "role": "user", 
+                "role": "user",
                 "content": prompt
                 }
             ]
@@ -35,7 +36,9 @@ def generate_response(model_name: str, prompt: str):
         return response.json()["choices"][0]["message"]["content"]
 
     except requests.exceptions.HTTPError as e:
-        raise RuntimeError(f"HTTP error from LLM server: {e} (status {response.status_code})") from e
+        raise RuntimeError(
+            f"HTTP error from LLM server: {e}, {response.status_code})"
+            ) from e
     except requests.exceptions.Timeout as e:
         raise RuntimeError(f"Request to LLM timed out: {e}") from e
     except requests.exceptions.RequestException as e:

genai/service/rag_service.py

@@ -3,6 +3,8 @@
 from langchain_qdrant import QdrantVectorStore
 from langchain_core.messages import BaseMessage, HumanMessage
 from langchain_core.prompts import ChatPromptTemplate, MessagesPlaceholder
+from genai.rag.llm.chat_model import ChatModel
+
 
 def retrieve_similar_docs(vector_store: QdrantVectorStore, user_query: str):
     """Retrieve similar documents based on the user query"""
@@ -11,10 +13,14 @@ def retrieve_similar_docs(vector_store: QdrantVectorStore, user_query: str):
     docs_content = "\n\n".join(doc.page_content for doc in retrieved_docs)
     return docs_content
 
-def prepare_prompt(user_query: str, docs_content: str, messages: List[BaseMessage]):
+
+def prepare_prompt(system_prompt: str,
+                   user_query: str,
+                   docs_content: str,
+                   messages: List[BaseMessage]):
     """Prepare the prompt with prompt templates to give to LLM"""
     prompt_template = ChatPromptTemplate([
-        ("system", "You are a helpful assistant for recipe generation based on the given ingredients and the following context:\n\n{context}"),
+        "system", system_prompt,
         MessagesPlaceholder("msgs")
     ])
 
@@ -26,3 +32,11 @@ def prepare_prompt(user_query: str, docs_content: str, messages: List[BaseMessag
     })
 
     return prompt
+
+# For testing purposes
+# if __name__ == "__main__":
+#     msg = HumanMessage(content="My name is John Doe.")
+#     llm = ChatModel()
+#     prompt = prepare_prompt(llm.get_system_prompt(), "Suggest me a basic breakfast. Btw, what is my name?", "", [msg])
+#     response = llm.invoke(prompt)
+#     print(response.content)