openai.py doesn't handle cases where LLM does not return an expected tool call gracefully

[steveternium]

I'm noticing that when I have dynamic routes set up, there are cases where the routing selects a function call, but openai code throws an error due to (I think) the LLM not actually returning a tool call in the completion. Possibly this is because I am using the default LLM (I can't seem to get rid of this warning: WARNING semantic_router.utils.logger No LLM provided for dynamic route, will use OpenAI LLM default. Ensure API key is set in OPENAI_API_KEY environment variable.)

Error is thrown here:
if function_schemas:
tool_calls = completion.choices[0].message.tool_calls
if tool_calls is None:
raise ValueError("Invalid output, expected a tool call.")
if len(tool_calls) < 1:
raise ValueError(
"Invalid output, expected at least one tool to be specified."
)

function_schemas = {'name': 'metacog_doc_loader', 'description': 'Loads a metacognition supplemental document related to the topic as context for the language model.\n\n:param topic: The topic which determines which file to load, and must be\n either "student-study", "time-management", or "none". Do NOT use any\n other topic.\n:type topic: str\n:return: The complete text of the loaded document or none if the topic was not known.', 'parameters': {'type': 'object', 'properties': {'topic': {'type': 'string', 'description': 'The topic which determines which file to load, and must be\n either "student-study", "time-management", or "none". Do NOT use any\n other topic.'}, 'config': {'type': 'object', 'description': 'No description available.'}}, 'required': ['topic']}}

But the chat completion is:
ChatCompletion(id='chatcmpl-AnV8oWIVf9mo2G5fUb5mc6Txisbvn', choices=[Choice(finish_reason='length', index=0, logprobs=None, message=ChatCompletionMessage(content="Setting realistic goals involves a few key steps:\n\n1. Self-Assessment: Understand your strengths, weaknesses, interests, and values. Reflect on past experiences to identify what has worked well for you and what hasn't.\n\n2. Research and Information Gathering: Learn about the requirements and challenges associated with your goal. This might involve talking to others who have achieved similar goals or reading relevant materials.\n\n3. SMART Criteria: Ensure your goals are Specific, Measurable, Achievable, Relevant, and Time-bound. This framework helps in setting clear and realistic objectives.\n\n4. Break Down Goals: Divide larger goals into smaller, manageable tasks. This makes it easier to track progress and adjust as needed.\n\n5. Seek Feedback: Discuss your goals with mentors, peers, or professionals who can provide insights and advice.\n\n6. Flexibility and Adaptation: Be prepared to adjust your goals based on new information or changes in circumstances.\n\n7. **Time Management", refusal=None, role='assistant', audio=None, function_call=None, tool_calls=None))], created=1736361142, model='gpt-4o-2024-08-06', object='chat.completion', service_tier=None, system_fingerprint='fp_d28bcae782', usage=CompletionUsage(completion_tokens=200, prompt_tokens=199, total_tokens=399, completion_tokens_details=CompletionTokensDetails(accepted_prediction_tokens=0, audio_tokens=0, reasoning_tokens=0, rejected_prediction_tokens=0), prompt_tokens_details=PromptTokensDetails(audio_tokens=0, cached_tokens=0)))

This is still very new to me, so it's entirely possible this is due to human error, but within a chat session that uses langgraph and semantic router, I'm see this a lot on short user messages that are vague, short, and probably should be out of scope of the main routes.