Improve OpenAI error handling (#164)

Show an error to the client when the openai content filter activates

app/llm/external/openai_chat.py

@@ -1,10 +1,16 @@
 import logging
 import time
-import traceback
 from datetime import datetime
 from typing import Literal, Any
 
-from openai import OpenAI
+from openai import (
+    OpenAI,
+    APIError,
+    APITimeoutError,
+    RateLimitError,
+    InternalServerError,
+    ContentFilterFinishReasonError,
+)
 from openai.lib.azure import AzureOpenAI
 from openai.types.chat import ChatCompletionMessage, ChatCompletionMessageParam
 from openai.types.shared_params import ResponseFormatJSONObject
@@ -82,35 +88,49 @@ def chat(
     ) -> PyrisMessage:
         print("Sending messages to OpenAI", messages)
         # noinspection PyTypeChecker
-        retries = 10
+        retries = 5
         backoff_factor = 2
         initial_delay = 1
+        # Maximum wait time: 1 + 2 + 4 + 8 + 16 = 31 seconds
+
+        messages = convert_to_open_ai_messages(messages)
 
         for attempt in range(retries):
             try:
                 if arguments.response_format == "JSON":
                     response = self._client.chat.completions.create(
                         model=self.model,
-                        messages=convert_to_open_ai_messages(messages),
+                        messages=messages,
                         temperature=arguments.temperature,
                         max_tokens=arguments.max_tokens,
                         response_format=ResponseFormatJSONObject(type="json_object"),
                     )
                 else:
                     response = self._client.chat.completions.create(
                         model=self.model,
-                        messages=convert_to_open_ai_messages(messages),
+                        messages=messages,
                         temperature=arguments.temperature,
                         max_tokens=arguments.max_tokens,
                     )
-                return convert_to_iris_message(response.choices[0].message)
-            except Exception as e:
+                choice = response.choices[0]
+                if choice.finish_reason == "content_filter":
+                    # I figured that an openai error would be automatically raised if the content filter activated,
+                    # but it seems that that is not the case.
+                    # We don't want to retry because the same message will likely be rejected again.
+                    # Raise an exception to trigger the global error handler and report a fatal error to the client.
+                    raise ContentFilterFinishReasonError()
+                return convert_to_iris_message(choice.message)
+            except (
+                APIError,
+                APITimeoutError,
+                RateLimitError,
+                InternalServerError,
+            ):
                 wait_time = initial_delay * (backoff_factor**attempt)
-                logging.warning(f"Exception on attempt {attempt + 1}: {e}")
-                traceback.print_exc()
+                logging.exception(f"OpenAI error on attempt {attempt + 1}:")
                 logging.info(f"Retrying in {wait_time} seconds...")
                 time.sleep(wait_time)
-        logging.error("Failed to interpret image after several attempts.")
+        raise Exception(f"Failed to get response from OpenAI after {retries} retries")
 
 
 class DirectOpenAIChatModel(OpenAIChatModel):

app/llm/external/openai_embeddings.py

@@ -1,6 +1,12 @@
 import logging
 from typing import Literal, Any
-from openai import OpenAI
+from openai import (
+    OpenAI,
+    APIError,
+    APITimeoutError,
+    RateLimitError,
+    InternalServerError,
+)
 from openai.lib.azure import AzureOpenAI
 
 from ...llm.external.model import EmbeddingModel
@@ -13,9 +19,10 @@ class OpenAIEmbeddingModel(EmbeddingModel):
     _client: OpenAI
 
     def embed(self, text: str) -> list[float]:
-        retries = 10
+        retries = 5
         backoff_factor = 2
         initial_delay = 1
+        # Maximum wait time: 1 + 2 + 4 + 8 + 16 = 31 seconds
 
         for attempt in range(retries):
             try:
@@ -25,15 +32,17 @@ def embed(self, text: str) -> list[float]:
                     encoding_format="float",
                 )
                 return response.data[0].embedding
-            except Exception as e:
+            except (
+                APIError,
+                APITimeoutError,
+                RateLimitError,
+                InternalServerError,
+            ):
                 wait_time = initial_delay * (backoff_factor**attempt)
-                logging.warning(f"Rate limit exceeded on attempt {attempt + 1}: {e}")
+                logging.exception(f"OpenAI error on attempt {attempt + 1}")
                 logging.info(f"Retrying in {wait_time} seconds...")
                 time.sleep(wait_time)
-        logging.error(
-            "Failed to get embedding after several attempts due to rate limit."
-        )
-        return []
+        raise Exception(f"Failed to get embedding from OpenAI after {retries} retries.")
 
 
 class DirectOpenAIEmbeddingModel(OpenAIEmbeddingModel):