Skip to content

Self-improving AI Employees for everyone

fix(llm): read endpoint error messages without a doubled article

trohitg41b85f9c9266
diff --git a/server/services/llm/protocol.py b/server/services/llm/protocol.py
index 692dcc5c..ccb1bc8c 100644
--- a/server/services/llm/protocol.py
+++ b/server/services/llm/protocol.py
@@ -372,16 +372,22 @@ class LLMError(Exception):
"lmstudio": "LM Studio",
}
# Common nouns take an article, so they read "The <noun>" as a
- # subject and "the <noun>" as an object.
+ # subject and "the <noun>" as an object. A name also works as an
+ # adjective ("the configured OpenAI model"); a noun with its article
+ # does not, so the two model-scoped messages are phrased around it.
generic_nouns = {"openai_compatible": "OpenAI-compatible endpoint"}
# A named endpoint's reference is "openai_compatible:<slug>".
provider_key = str(self.provider or "").strip().lower().partition(":")[0]
if provider_key in provider_names:
provider = provider_object = provider_names[provider_key]
+ not_found = f"The configured {provider} model or endpoint was not found."
+ context_length = f"The request exceeds the {provider} model context window."
else:
noun = generic_nouns.get(provider_key, "language model provider")
provider = f"The {noun}"
provider_object = f"the {noun}"
+ not_found = f"{provider} did not find the configured model."
+ context_length = f"The request exceeds the context window of the model at {provider_object}."
category = (
self.category.value
if isinstance(self.category, LLMErrorCategory)
@@ -403,14 +409,8 @@ class LLMError(Exception):
LLMErrorCategory.INVALID_REQUEST.value: (
f"{provider} rejected the model request configuration."
),
- LLMErrorCategory.NOT_FOUND.value: (
- f"The configured {provider_object} model or endpoint "
- "was not found."
- ),
- LLMErrorCategory.CONTEXT_LENGTH.value: (
- f"The request exceeds the {provider_object} model "
- "context window."
- ),
+ LLMErrorCategory.NOT_FOUND.value: not_found,
+ LLMErrorCategory.CONTEXT_LENGTH.value: context_length,
LLMErrorCategory.TIMEOUT.value: (
f"The request to {provider_object} timed out."
),
diff --git a/server/tests/llm/test_protocol_codec.py b/server/tests/llm/test_protocol_codec.py
index 2c0ff698..55bd902d 100644
--- a/server/tests/llm/test_protocol_codec.py
+++ b/server/tests/llm/test_protocol_codec.py
@@ -179,6 +179,43 @@ def test_structured_error_exposes_only_category_based_user_message():
assert "Bearer secret" not in error.user_message
+ ("provider", "not_found", "context_length"),
+ [
+ (
+ "openai",
+ "The configured OpenAI model or endpoint was not found.",
+ "The request exceeds the OpenAI model context window.",
+ ),
+ (
+ "openai_compatible:home",
+ "The OpenAI-compatible endpoint did not find the configured model.",
+ "The request exceeds the context window of the model at the OpenAI-compatible endpoint.",
+ ),
+ (
+ "some-unregistered-provider",
+ "The language model provider did not find the configured model.",
+ "The request exceeds the context window of the model at the language model provider.",
+ ),
+ ],
+)
+def test_model_scoped_messages_read_for_names_and_common_nouns(provider, not_found, context_length):
+ def message(category: LLMErrorCategory) -> str:
+ return LLMError(message="raw", provider=provider, category=category).user_message
+
+ assert message(LLMErrorCategory.NOT_FOUND) == not_found
+ assert message(LLMErrorCategory.CONTEXT_LENGTH) == context_length
+
+
[email protected]("provider", ["openai_compatible:home", "some-unregistered-provider"])
[email protected]("category", list(LLMErrorCategory))
+def test_no_message_doubles_the_article_of_a_common_noun(provider, category):
+ text = LLMError(message="raw", provider=provider, category=category).user_message.lower()
+
+ assert "the the " not in text
+ assert "configured the " not in text
+
+
def test_structured_error_preserves_http_date_retry_after():
class Response:
status_code = 429