Merge ecdb70e133 into 252c71c0b2

litellm/utils.py

@@ -7260,6 +7260,16 @@ def shorten_message_to_fit_limit(message, tokens_needed, model: str | None, rais
     return message
 
 
+def _extract_system_message_content(content: Union[str, list, Any]) -> str:
+    if isinstance(content, str):
+        return content
+    if isinstance(content, list):
+        return "".join(
+            part.get("text", "") for part in content if isinstance(part, dict) and part.get("type") == "text"
+        )
+    return ""
+
+
 # LiteLLM token trimmer
 # this code is borrowed from https://github.com/KillianLucas/tokentrim/blob/main/tokentrim/tokentrim.py
 # Credits for this code go to Killian Lucas
@@ -7305,7 +7315,7 @@ def trim_messages(
         for message in messages:
             if message["role"] == "system":
                 system_message += "\n" if system_message else ""
-                system_message += message["content"]
+                system_message += _extract_system_message_content(message.get("content", ""))
 
         ## Handle Tool Call ## - check if last message is a tool response, return as is - https://github.com/BerriAI/litellm/issues/4931
         tool_messages: Final = []

tests/test_litellm/test_utils.py

@@ -58,6 +58,7 @@ from litellm.utils import (
     get_prompt_cache_min_tokens,
     is_cached_message,
     is_prompt_caching_valid_prompt,
+    trim_messages,
 )
 
 # Adds the parent directory to the system path
@@ -122,6 +123,26 @@ def test_usage_openai_cache_write_tokens_populates_both_names():
     assert usage.prompt_tokens_details.cache_creation_tokens == 800
 
 
+def test_trim_messages_system_message_list_content():
+    messages = [
+        {
+            "role": "system",
+            "content": [
+                {"type": "text", "text": "You are a helpful assistant."},
+                {"type": "text", "text": "Always answer in French."},
+            ],
+        },
+        {"role": "user", "content": "Hello there, how are you?"},
+    ]
+
+    trimmed_messages = trim_messages(messages, model="gpt-4", max_tokens=5)
+
+    assert (
+        trimmed_messages[0]["content"]
+        == "You are a helpful assistant.Always answer in French."
+    )
+
+
 def test_usage_anthropic_cache_creation_maps_to_cache_write_tokens():
     """Anthropic/Bedrock report the top-level cache_creation_input_tokens field.
     It must be normalized onto the OpenAI cache_write_tokens name as well as the