fix: add frequency penalty for gpt-4o-mini (#3166)

letta/llm_api/openai_client.py

@@ -216,6 +216,10 @@ class OpenAIClient(LLMClientBase):
             # NOTE: the reasoners that don't support temperature require 1.0, not None
             temperature=llm_config.temperature if supports_temperature_param(model) else 1.0,
         )
+
+        if llm_config.frequency_penalty is not None:
+            data.frequency_penalty = llm_config.frequency_penalty
+
         if tools and supports_parallel_tool_calling(model):
             data.parallel_tool_calls = False
 

letta/schemas/llm_config.py

@@ -77,6 +77,10 @@ class LLMConfig(BaseModel):
     max_reasoning_tokens: int = Field(
         0, description="Configurable thinking budget for extended thinking, only used if enable_reasoner is True. Minimum value is 1024."
     )
+    frequency_penalty: Optional[float] = Field(
+        None,  # Can also deafult to 0.0?
+        description="Positive values penalize new tokens based on their existing frequency in the text so far, decreasing the model's likelihood to repeat the same line verbatim. From OpenAI: Number between -2.0 and 2.0.",
+    )
 
     # FIXME hack to silence pydantic protected namespace warning
     model_config = ConfigDict(protected_namespaces=())

letta/schemas/providers.py

@@ -324,18 +324,25 @@ class OpenAIProvider(Provider):
             else:
                 handle = self.get_handle(model_name)
 
-            configs.append(
-                LLMConfig(
-                    model=model_name,
-                    model_endpoint_type="openai",
-                    model_endpoint=self.base_url,
-                    context_window=context_window_size,
-                    handle=handle,
-                    provider_name=self.name,
-                    provider_category=self.provider_category,
-                )
+            llm_config = LLMConfig(
+                model=model_name,
+                model_endpoint_type="openai",
+                model_endpoint=self.base_url,
+                context_window=context_window_size,
+                handle=handle,
+                provider_name=self.name,
+                provider_category=self.provider_category,
             )
 
+            # gpt-4o-mini has started to regress with pretty bad emoji spam loops
+            # this is to counteract that
+            if "gpt-4o-mini" in model_name:
+                llm_config.frequency_penalty = 1.0
+            if "gpt-4.1-mini" in model_name:
+                llm_config.frequency_penalty = 1.0
+
+            configs.append(llm_config)
+
         # for OpenAI, sort in reverse order
         if self.base_url == "https://api.openai.com/v1":
             # alphnumeric sort