fix prompt too long

Author: jrobertboos

src/utils/agents/streaming.py

@@ -221,12 +221,29 @@ async def generate_agent_response(
         context.query_request.conversation_id is None
         and bool(context.query_request.generate_topic_summary)
     )
-    topic_summary = await maybe_get_topic_summary(
-        generate_topic_summary=should_generate_topic_summary,
-        input_text=context.query_request.query,
-        client=context.client,
-        model_id=responses_params.model,
-    )
+    try:
+        topic_summary = await maybe_get_topic_summary(
+            generate_topic_summary=should_generate_topic_summary,
+            input_text=context.query_request.query,
+            client=context.client,
+            model_id=responses_params.model,
+        )
+    except HTTPException as exc:
+        logger.warning(
+            "Topic summary failed for request %s: %s",
+            context.request_id,
+            exc.detail,
+        )
+        detail: dict[str, str] = exc.detail if isinstance(exc.detail, dict) else {}
+        yield serialize_event(
+            ErrorStreamPayload.create(
+                status_code=exc.status_code,
+                response=detail.get("response", "Internal server error"),
+                cause=detail.get("cause", str(exc.detail)),
+            ),
+            media_type,
+        )
+        return
     logger.info("Consuming tokens")
     consume_query_tokens(
         user_id=context.user_id,

src/utils/query.py

@@ -574,10 +574,9 @@ def handle_known_apistatus_errors(
     Returns:
         AbstractErrorResponse: The error response model.
     """
-    if error.status_code == 400:
-        error_message = getattr(error, "message", str(error))
-        if is_context_length_error(error_message):
-            return PromptTooLongResponse(model=model_id)
-    elif error.status_code == 429:
+    error_message = getattr(error, "message", str(error))
+    if is_context_length_error(error_message):
+        return PromptTooLongResponse(model=model_id)
+    if error.status_code == 429:
         return QuotaExceededResponse.model(model_id)
     return InternalServerErrorResponse.generic()