feat: Add ability to configure OpenAI base URL in ChatGPTAgentConfig

tests/streaming/agent/test_base_agent.py

@@ -177,3 +177,16 @@ async def test_action_response_agent_input(mocker: MockerFixture):
     # TODO: assert that the canned response is optionally sent if the action is not quiet
     # and that it goes through the normal flow when the action is not quiet
     pass
+
+@pytest.fixture
+def agent_config():
+    return ChatGPTAgentConfig(
+        openai_api_key="test_key",
+        model_name="llama3-8b-8192",
+        base_url="https://api.groq.com/openai/v1",
+        prompt_preamble="Test prompt",
+    )
+
+def test_chat_gpt_agent_base_url(agent_config):
+    agent = ChatGPTAgent(agent_config)
+    assert agent.openai_client.base_url == "https://api.groq.com/openai/v1"

vocode/streaming/agent/chat_gpt_agent.py

@@ -39,9 +39,11 @@ def instantiate_openai_client(agent_config: ChatGPTAgentConfig, model_fallback:
     else:
         if agent_config.openai_api_key is not None:
             logger.info("Using OpenAI API key override")
+        if agent_config.base_url is not None:
+            logger.info(f"Using OpenAI base URL override: {agent_config.base_url}")
         return AsyncOpenAI(
             api_key=agent_config.openai_api_key or os.environ["OPENAI_API_KEY"],
-            base_url="https://api.openai.com/v1",
+            base_url=agent_config.base_url or "https://api.openai.com/v1",
             max_retries=0 if model_fallback else OPENAI_DEFAULT_MAX_RETRIES,
         )
 

vocode/streaming/agent/token_utils.py

@@ -90,7 +90,7 @@ def get_tokenizer_info(model: str) -> Optional[TokenizerInfo]:
     try:
         encoding = tiktoken.encoding_for_model(model)
     except KeyError:
-        logger.warning("Warning: model not found. Using cl100k_base encoding.")
+        logger.warning(f"Warning: model not found. Using cl100k_base encoding for {model}.")
         encoding = tiktoken.get_encoding("cl100k_base")
     if model in {
         "gpt-3.5-turbo-0613",
@@ -117,6 +117,14 @@ def get_tokenizer_info(model: str) -> Optional[TokenizerInfo]:
         )
         tokens_per_message = 3
         tokens_per_name = 1
+    elif "llama" in model.lower():
+        logger.warning(
+            f"Warning: you are using a llama model with an OpenAI compatible endpoint. \
+            Llama models are not supported natively support for token counting in tiktoken. \
+            Using cl100k_base encoding for {model} as an APPROXIMATION of token usage."
+        )
+        tokens_per_message = 3
+        tokens_per_name = 1
     else:
         return None
 
@@ -250,3 +258,4 @@ def format_default(schema):
         result += "_: " + formatted
     result += ") => any;\n\n"
     return result
+

vocode/streaming/models/agent.py

@@ -115,6 +115,7 @@ class ChatGPTAgentConfig(AgentConfig, type=AgentType.CHAT_GPT.value):  # type: i
     openai_api_key: Optional[str] = None
     prompt_preamble: str
     model_name: str = CHAT_GPT_AGENT_DEFAULT_MODEL_NAME
+    base_url: Optional[str] = None
     temperature: float = LLM_AGENT_DEFAULT_TEMPERATURE
     max_tokens: int = LLM_AGENT_DEFAULT_MAX_TOKENS
     azure_params: Optional[AzureOpenAIConfig] = None