Add all prompts to a centralized prompt service

Author: sahilds1

server/api/views/conversations/views.py

@@ -16,6 +16,7 @@
 from .models import Conversation, Message
 from .serializers import ConversationSerializer
 from ...services.tools.tools import tools, execute_tool
+from ...services.prompt_services import PromptTemplates
 
 
 @csrf_exempt
@@ -47,7 +48,7 @@ def extract_text(request: str) -> JsonResponse:
         messages=[
             {
                 "role": "system",
-                "content": "Give a brief description of this medicine: %s" % tokens,
+                "content": PromptTemplates.get_medicine_description_prompt(tokens),
             }
         ],
         max_tokens=500,
@@ -64,8 +65,10 @@ def get_tokens(string: str, encoding_name: str) -> str:
     output_string = encoding.decode(tokens)
     return output_string
 
+
 class OpenAIAPIException(APIException):
     """Custom exception for OpenAI API errors."""
+
     status_code = status.HTTP_500_INTERNAL_SERVER_ERROR
     default_detail = "An error occurred while communicating with the OpenAI API."
     default_code = "openai_api_error"
@@ -77,6 +80,7 @@ def __init__(self, detail=None, code=None):
             self.detail = {"error": self.default_detail}
         self.status_code = code or self.status_code
 
+
 class ConversationViewSet(viewsets.ModelViewSet):
     serializer_class = ConversationSerializer
     permission_classes = [IsAuthenticated]
@@ -93,26 +97,29 @@ def destroy(self, request, *args, **kwargs):
         self.perform_destroy(instance)
         return Response(status=status.HTTP_204_NO_CONTENT)
 
-    @action(detail=True, methods=['post'])
+    @action(detail=True, methods=["post"])
     def continue_conversation(self, request, pk=None):
         conversation = self.get_object()
-        user_message = request.data.get('message')
-        page_context = request.data.get('page_context')
+        user_message = request.data.get("message")
+        page_context = request.data.get("page_context")
 
         if not user_message:
             return Response({"error": "Message is required"}, status=400)
 
         # Save user message
-        Message.objects.create(conversation=conversation,
-                               content=user_message, is_user=True)
+        Message.objects.create(
+            conversation=conversation, content=user_message, is_user=True
+        )
 
         # Get ChatGPT response
         chatgpt_response = self.get_chatgpt_response(
-            conversation, user_message, page_context)
+            conversation, user_message, page_context
+        )
 
         # Save ChatGPT response
-        Message.objects.create(conversation=conversation,
-                               content=chatgpt_response, is_user=False)
+        Message.objects.create(
+            conversation=conversation, content=chatgpt_response, is_user=False
+        )
 
         # Generate or update title if it's the first message or empty
         if conversation.messages.count() <= 2 or not conversation.title:
@@ -121,27 +128,35 @@ def continue_conversation(self, request, pk=None):
 
         return Response({"response": chatgpt_response, "title": conversation.title})
 
-    @action(detail=True, methods=['patch'])
+    @action(detail=True, methods=["patch"])
     def update_title(self, request, pk=None):
         conversation = self.get_object()
-        new_title = request.data.get('title')
+        new_title = request.data.get("title")
 
         if not new_title:
-            return Response({"error": "New title is required"}, status=status.HTTP_400_BAD_REQUEST)
+            return Response(
+                {"error": "New title is required"}, status=status.HTTP_400_BAD_REQUEST
+            )
 
         conversation.title = new_title
         conversation.save()
 
-        return Response({"status": "Title updated successfully", "title": conversation.title})
+        return Response(
+            {"status": "Title updated successfully", "title": conversation.title}
+        )
 
     def get_chatgpt_response(self, conversation, user_message, page_context=None):
-        messages = [{
-            "role": "system", 
-            "content": "You are a knowledgeable assistant. Balancer is a powerful tool for selecting bipolar medication for patients. We are open-source and available for free use. Your primary role is to assist licensed clinical professionals with information related to Balancer and bipolar medication selection. If applicable, use the supplied tools to assist the professional."
-        }]
+        messages = [
+            {
+                "role": "system",
+                "content": PromptTemplates.get_conversation_system_prompt(),
+            }
+        ]
 
         if page_context:
-            context_message = f"If applicable, please use the following content to ask questions. If not applicable, please answer to the best of your ability: {page_context}"
+            context_message = PromptTemplates.get_conversation_page_context_prompt(
+                page_context
+            )
             messages.append({"role": "system", "content": context_message})
 
         for msg in conversation.messages.all():
@@ -155,46 +170,50 @@ def get_chatgpt_response(self, conversation, user_message, page_context=None):
                 model="gpt-3.5-turbo",
                 messages=messages,
                 tools=tools,
-                tool_choice="auto"
+                tool_choice="auto",
             )
 
             response_message = response.choices[0].message
-            tool_calls = response_message.get('tool_calls', [])
+            tool_calls = response_message.get("tool_calls", [])
 
             if not tool_calls:
-                return response_message['content']
-
+                return response_message["content"]
 
             # Handle tool calls
             # Add the assistant's message with tool calls to the conversation
-            messages.append({
-                "role": "assistant",
-                "content": response_message.get('content', ''),
-                "tool_calls": tool_calls
-            })
-            
+            messages.append(
+                {
+                    "role": "assistant",
+                    "content": response_message.get("content", ""),
+                    "tool_calls": tool_calls,
+                }
+            )
+
             # Process each tool call
             for tool_call in tool_calls:
-                tool_call_id = tool_call['id']
-                tool_function_name = tool_call['function']['name']
-                tool_arguments = json.loads(tool_call['function'].get('arguments', '{}'))
-                
+                tool_call_id = tool_call["id"]
+                tool_function_name = tool_call["function"]["name"]
+                tool_arguments = json.loads(
+                    tool_call["function"].get("arguments", "{}")
+                )
+
                 # Execute the tool
                 results = execute_tool(tool_function_name, tool_arguments)
-                
+
                 # Add the tool response message
-                messages.append({
-                    "role": "tool",
-                    "content": str(results),  # Convert results to string
-                    "tool_call_id": tool_call_id
-                })
-            
+                messages.append(
+                    {
+                        "role": "tool",
+                        "content": str(results),  # Convert results to string
+                        "tool_call_id": tool_call_id,
+                    }
+                )
+
             # Final API call with tool results
             final_response = openai.ChatCompletion.create(
-                model="gpt-3.5-turbo",
-                    messages=messages
-                )
-            return final_response.choices[0].message['content']
+                model="gpt-3.5-turbo", messages=messages
+            )
+            return final_response.choices[0].message["content"]
         except openai.error.OpenAIError as e:
             logging.error("OpenAI API Error: %s", str(e))
             raise OpenAIAPIException(detail=str(e))
@@ -206,14 +225,17 @@ def generate_title(self, conversation):
         # Get the first two messages
         messages = conversation.messages.all()[:2]
         context = "\n".join([msg.content for msg in messages])
-        prompt = f"Based on the following conversation, generate a short, descriptive title (max 6 words):\n\n{context}"
+        prompt = PromptTemplates.get_title_generation_user_prompt(context)
 
         response = openai.ChatCompletion.create(
             model="gpt-3.5-turbo",
             messages=[
-                {"role": "system", "content": "You are a helpful assistant that generates short, descriptive titles."},
-                {"role": "user", "content": prompt}
-            ]
+                {
+                    "role": "system",
+                    "content": PromptTemplates.get_title_generation_system_prompt(),
+                },
+                {"role": "user", "content": prompt},
+            ],
         )
 
-        return response.choices[0].message['content'].strip()
+        return response.choices[0].message["content"].strip()

server/api/views/embeddings/embeddingsView.py

@@ -6,59 +6,52 @@
 from ...services.embedding_services import get_closest_embeddings
 from ...services.conversions_services import convert_uuids
 from ...services.openai_services import openAIServices
+from ...services.prompt_services import PromptTemplates
 from django.utils.decorators import method_decorator
 from django.views.decorators.csrf import csrf_exempt
 import json
 
 
-@method_decorator(csrf_exempt, name='dispatch')
+@method_decorator(csrf_exempt, name="dispatch")
 class AskEmbeddingsAPIView(APIView):
     permission_classes = [IsAuthenticated]
 
     def post(self, request, *args, **kwargs):
         try:
             user = request.user
-            guid = request.query_params.get('guid')
-            stream = request.query_params.get(
-                'stream', 'false').lower() == 'true'
+            guid = request.query_params.get("guid")
+            stream = request.query_params.get("stream", "false").lower() == "true"
 
-            request_data = request.data.get('message', None)
+            request_data = request.data.get("message", None)
             if not request_data:
-                return Response({"error": "Message data is required."}, status=status.HTTP_400_BAD_REQUEST)
+                return Response(
+                    {"error": "Message data is required."},
+                    status=status.HTTP_400_BAD_REQUEST,
+                )
             message = str(request_data)
 
             embeddings_results = get_closest_embeddings(
-                user=user, message_data=message, guid=guid)
+                user=user, message_data=message, guid=guid
+            )
             embeddings_results = convert_uuids(embeddings_results)
 
             prompt_texts = [
-                f"[Start of INFO {i+1} === GUID: {obj['file_id']}, Page Number: {obj['page_number']}, Chunk Number: {obj['chunk_number']}, Text: {obj['text']} === End of INFO {i+1} ]" for i, obj in enumerate(embeddings_results)]
+                f"[Start of INFO {i + 1} === GUID: {obj['file_id']}, Page Number: {obj['page_number']}, Chunk Number: {obj['chunk_number']}, Text: {obj['text']} === End of INFO {i + 1} ]"
+                for i, obj in enumerate(embeddings_results)
+            ]
 
             listOfEmbeddings = " ".join(prompt_texts)
 
-            prompt_text = (
-                f"""You are an AI assistant tasked with providing detailed, well-structured responses based on the information provided in [PROVIDED-INFO]. Follow these guidelines strictly: 
-                1. Content: Use information contained within [PROVIDED-INFO] to answer the question. 
-                2. Organization: Structure your response with clear sections and paragraphs. 
-                3. Citations: After EACH sentence that uses information from [PROVIDED-INFO], include a citation in this exact format:***[{{file_id}}], Page {{page_number}}, Chunk {{chunk_number}}*** . Only use citations that correspond to the information you're presenting. 
-                4. Clarity: Ensure your answer is well-structured and easy to follow. 
-                5. Direct Response: Answer the user's question directly without unnecessary introductions or filler phrases. 
-                Here's an example of the required response format:
-                ________________________________________ 
-                See's Candy in the context of sales during a specific event. The candy counters rang up 2,690 individual sales on a Friday, and an additional 3,931 transactions on a Saturday ***[16s848as-vcc1-85sd-r196-7f820a4s9de1, Page 5, Chunk 26]***.
-                People like the consumption of fudge and peanut brittle the most ***[130714d7-b9c1-4sdf-b146-fdsf854cad4f, Page 9, Chunk 19]***. 
-                Here is the history of See's Candy: the company was purchased in 1972, and its products have not been materially altered in 101 years ***[895sdsae-b7v5-416f-c84v-7f9784dc01e1, Page 2, Chunk 13]***. 
-                Bipolar disorder treatment often involves mood stabilizers. Lithium is a commonly prescribed mood stabilizer effective in reducing manic episodes ***[b99988ac-e3b0-4d22-b978-215e814807f4, Page 29, Chunk 122]***. For acute hypomania or mild to moderate mania, initial treatment with risperidone or olanzapine monotherapy is suggested ***[b99988ac-e3b0-4d22-b978-215e814807f4, Page 24, Chunk 101]***. 
-                ________________________________________ 
-                Please provide your response to the user's question following these guidelines precisely.
-                [PROVIDED-INFO] = {listOfEmbeddings}"""
-            )
+            prompt_text = PromptTemplates.get_embeddings_query_prompt(listOfEmbeddings)
 
             if stream:
+
                 def stream_generator():
                     try:
                         last_chunk = ""
-                        for chunk in openAIServices.openAI(message, prompt_text, stream=True, raw_stream=False):
+                        for chunk in openAIServices.openAI(
+                            message, prompt_text, stream=True, raw_stream=False
+                        ):
                             # Format as Server-Sent Events for better client handling
                             if chunk and chunk != last_chunk:
                                 last_chunk = chunk
@@ -72,27 +65,29 @@ def stream_generator():
                         yield f"data: {error_data}\n\n"
 
                 response = StreamingHttpResponse(
-                    stream_generator(),
-                    content_type='text/event-stream'
+                    stream_generator(), content_type="text/event-stream"
                 )
                 # Add CORS and caching headers for streaming
-                response['Cache-Control'] = 'no-cache'
-                response['Access-Control-Allow-Origin'] = '*'
+                response["Cache-Control"] = "no-cache"
+                response["Access-Control-Allow-Origin"] = "*"
                 # Disable nginx buffering if behind nginx
-                response['X-Accel-Buffering'] = 'no'
+                response["X-Accel-Buffering"] = "no"
                 return response
             # Non-streaming response
             answer = openAIServices.openAI(
-                userMessage=message,
-                prompt=prompt_text,
-                stream=False
+                userMessage=message, prompt=prompt_text, stream=False
+            )
+            return Response(
+                {
+                    "question": message,
+                    "llm_response": answer,
+                    "embeddings_info": embeddings_results,
+                    "sent_to_llm": prompt_text,
+                },
+                status=status.HTTP_200_OK,
             )
-            return Response({
-                "question": message,
-                "llm_response": answer,
-                "embeddings_info": embeddings_results,
-                "sent_to_llm": prompt_text,
-            }, status=status.HTTP_200_OK)
 
         except Exception as e:
-            return Response({"error": str(e)}, status=status.HTTP_500_INTERNAL_SERVER_ERROR)
+            return Response(
+                {"error": str(e)}, status=status.HTTP_500_INTERNAL_SERVER_ERROR
+            )

server/api/views/text_extraction/views.py

@@ -9,6 +9,7 @@
 from django.views.decorators.csrf import csrf_exempt
 
 from ...services.openai_services import openAIServices
+from ...services.prompt_services import PromptTemplates
 from api.models.model_embeddings import Embeddings
 
 # This is to use openai to extract the rules to save cost
@@ -37,29 +38,7 @@ class RuleExtractionAPIOpenAIView(APIView):
 
     def get(self, request):
         try:
-            user_prompt = """
-            You're analyzing medical text from multiple sources. Each chunk is labeled [chunk-X].
-
-            Act as a seasoned physician or medical professional who treats patients with bipolar disorder.
-
-            Identify rules for medication inclusion or exclusion based on medical history or concerns.
-
-            For each rule you find, return a JSON object using the following format:
-
-            {
-              "rule": "<condition or concern>",
-              "type": "INCLUDE" or "EXCLUDE",
-              "reason": "<short explanation for why this rule applies>",
-              "medications": ["<medication 1>", "<medication 2>", ...],
-              "source": "<chunk-X>"
-            }
-
-            Only include rules that are explicitly stated or strongly implied in the chunk.
-
-            Only use the chunks provided. If no rule is found in a chunk, skip it.
-
-            Return the entire output as a JSON array.
-            """
+            user_prompt = PromptTemplates.get_text_extraction_prompt()
 
             guid = request.query_params.get("guid")
             query = Embeddings.objects.filter(upload_file__guid=guid)