Use chat model for query rewriting (#1659)

* Use chat model for query rewriting

* Snapshots

Author: pamelafox

app/backend/app.py

@@ -520,6 +520,8 @@ async def setup_clients():
             auth_helper=auth_helper,
             vision_endpoint=AZURE_VISION_ENDPOINT,
             vision_token_provider=token_provider,
+            chatgpt_model=OPENAI_CHATGPT_MODEL,
+            chatgpt_deployment=AZURE_OPENAI_CHATGPT_DEPLOYMENT,
             gpt4v_deployment=AZURE_OPENAI_GPT4V_DEPLOYMENT,
             gpt4v_model=AZURE_OPENAI_GPT4V_MODEL,
             embedding_model=OPENAI_EMB_MODEL,

app/backend/approaches/chatreadretrievereadvision.py

@@ -32,6 +32,8 @@ def __init__(
         blob_container_client: ContainerClient,
         openai_client: AsyncOpenAI,
         auth_helper: AuthenticationHelper,
+        chatgpt_model: str,
+        chatgpt_deployment: Optional[str],  # Not needed for non-Azure OpenAI
         gpt4v_deployment: Optional[str],  # Not needed for non-Azure OpenAI
         gpt4v_model: str,
         embedding_deployment: Optional[str],  # Not needed for non-Azure OpenAI or for retrieval_mode="text"
@@ -48,6 +50,8 @@ def __init__(
         self.blob_container_client = blob_container_client
         self.openai_client = openai_client
         self.auth_helper = auth_helper
+        self.chatgpt_model = chatgpt_model
+        self.chatgpt_deployment = chatgpt_deployment
         self.gpt4v_deployment = gpt4v_deployment
         self.gpt4v_model = gpt4v_model
         self.embedding_deployment = embedding_deployment
@@ -107,8 +111,10 @@ async def run_until_final_call(
         user_query_request = "Generate search query for: " + original_user_query
 
         query_response_token_limit = 100
+        query_model = self.chatgpt_model
+        query_deployment = self.chatgpt_deployment
         query_messages = build_messages(
-            model=self.gpt4v_model,
+            model=query_model,
             system_prompt=self.query_prompt_template,
             few_shots=self.query_prompt_few_shots,
             past_messages=past_messages,
@@ -117,7 +123,7 @@ async def run_until_final_call(
         )
 
         chat_completion: ChatCompletion = await self.openai_client.chat.completions.create(
-            model=self.gpt4v_deployment if self.gpt4v_deployment else self.gpt4v_model,
+            model=query_deployment if query_deployment else query_model,
             messages=query_messages,
             temperature=0.0,  # Minimize creativity for search query generation
             max_tokens=query_response_token_limit,
@@ -197,9 +203,9 @@ async def run_until_final_call(
                     "Prompt to generate search query",
                     [str(message) for message in query_messages],
                     (
-                        {"model": self.gpt4v_model, "deployment": self.gpt4v_deployment}
-                        if self.gpt4v_deployment
-                        else {"model": self.gpt4v_model}
+                        {"model": query_model, "deployment": query_deployment}
+                        if query_deployment
+                        else {"model": query_model}
                     ),
                 ),
                 ThoughtStep(

tests/snapshots/test_app/test_chat_vision/client1/result.json

@@ -24,7 +24,8 @@
                             "{'role': 'user', 'content': 'Generate search query for: Are interest rates high?'}"
                         ],
                         "props": {
-                            "model": "gpt-4"
+                            "deployment": "test-chatgpt",
+                            "model": "gpt-35-turbo"
                         },
                         "title": "Prompt to generate search query"
                     },

tests/snapshots/test_app/test_chat_vision_vectors/client1/result.json

@@ -24,7 +24,8 @@
                             "{'role': 'user', 'content': 'Generate search query for: Are interest rates high?'}"
                         ],
                         "props": {
-                            "model": "gpt-4"
+                            "deployment": "test-chatgpt",
+                            "model": "gpt-35-turbo"
                         },
                         "title": "Prompt to generate search query"
                     },

tests/test_chatvisionapproach.py

@@ -52,6 +52,8 @@ def chat_approach(openai_client, mock_confidential_client_success):
         blob_container_client=None,
         vision_endpoint="endpoint",
         vision_token_provider=lambda: "token",
+        chatgpt_model="gpt-35-turbo",
+        chatgpt_deployment="chat",
         gpt4v_deployment="gpt-4v",
         gpt4v_model="gpt-4v",
         embedding_deployment="embeddings",