[BUG MODEL]: magistral-medium-latest returns erratic ThinkChunks and TextChunks #252
Description
Model
magistral-medium-latest
Expected Behavior
I expect a single ThinkChunk and a single TextChunk in the content as indicated on the documentation. However, what I think I'm seeing instead is that the TextChunk is itself split up into three parts: a TextChunk, an empty ThinkChunk, and another TextChunk, resulting in four chunks in total.
(The tool_calls part seems fine.)
Additional Context
This occurs in the context of SigmundAI when using magistral-medium-latest.
Suggested Solutions
My hunch is that the model response is parsed incorrectly server-side because of the workspace tags that are included in the response.
Request Payload
[{'content': 'You are Sigmund, a brilliant AI assistant. You always put code '
'between triple backticks (```), and never use triple backticks '
'for anything else. You sometimes use emojis.\n'
'\n'
'When providing examples or updated text or code to the user '
'always do this through the workspace. You set workspace content '
'by including `<workspace language="language">text or '
'code</workspace>` in your reply or by using tools.\n'
'\n'
'Here is an example:\n'
'\n'
'```\n'
'Sure, I can write a hello world function for you! I added the '
'code to the workspace.\n'
'\n'
'<workspace language="python">\n'
'def hello_world():\n'
" print('Hello world!')\n"
'</workspace>\n'
'```\n'
'\n'
'Available languages: css, html, javascript, opensesame, python, '
'r, markdown\n'
'\n'
'Important: always use the workspace as shown above, and do *not* '
'include long examples of text or code in your reply.\n',
'role': 'system'},
{'content': 'Can you show me how to calculate the square root of 7 in Python?',
'role': 'user'}]Output
ChatCompletionResponse(
id="06ccb1bf42a44619b13b83678b95ca22",
object="chat.completion",
model="magistral-medium-latest",
usage=UsageInfo(
prompt_tokens=339,
completion_tokens=245,
total_tokens=584,
prompt_audio_seconds=Unset(),
),
created=1753539878,
choices=[
ChatCompletionChoice(
index=0,
message=AssistantMessage(
content=[
ThinkChunk(
thinking=[
TextChunk(
text="Alright, I need to show how to calculate the square root of 7 in Python. I remember that Python has a module called `math` that contains a function for calculating square roots. The function is called `sqrt`.\n\nFirst, I need to import the `math` module. Then, I can use the `sqrt` function to calculate the square root of 7.\n\nHere's how I can do it:\n\n```python\nimport math\nresult = math.sqrt(7)\nprint(result)\n```\n\nBut before I show this to the user, I should test it to make sure it works correctly. I can use the `execute_code` tool to run this code and see the output.",
type="text",
)
],
closed=None,
type="thinking",
),
TextChunk(
text='Sure! I can show you how to calculate the square root of 7 in Python. Here\'s the code:\n\n```python\nimport math\nresult = math.sqrt(7)\nprint(result)\n```\n\nLet me execute this code to show you the result.\n\n<workspace language="python">\nimport math\nresult = math.sqrt(7)\nprint(result)\n',
type="text",
),
ThinkChunk(thinking=[], closed=None, type="thinking"),
TextChunk(text="workspace>\n\n", type="text"),
],
tool_calls=[
ToolCall(
function=FunctionCall(
name="execute_code",
arguments='{"language": "python", "code": "import math\\nresult = math.sqrt(7)\\nprint(result)"}',
),
id="y7FYBCEod",
type=None,
index=0,
metadata=Unset(),
)
],
prefix=False,
role="assistant",
),
finish_reason="tool_calls",
)
],
)
```