mirror of
https://github.com/BerriAI/litellm.git
synced 2026-11-16 09:22:22 +00:00
fix(mistral): map tool-call history into Conversations inputs
The Conversations transform built inputs from only role and content, so an assistant message's tool_calls and a tool-role result were silently dropped when a web-search turn landed mid function-calling conversation; the model then lost the prior call context with no error. Each assistant tool_call now becomes a function.call input entry (tool_call_id/name/arguments) and each tool message a function.result entry (tool_call_id/result), matching Mistral's Conversations inputs schema, the way other provider transforms preserve tool history instead of flattening it
This commit is contained in:
parent
e86f74be96
commit
ea340d6d2f
3 changed files with 109 additions and 9 deletions
|
|
@ -21,7 +21,6 @@ from litellm.llms.mistral.chat.transformation import MistralConfig
|
|||
from litellm.llms.mistral.common_utils import OBJ_LIST, STR_OBJ_DICT, WEB_SEARCH_TOOL_TYPES
|
||||
from litellm.types.llms.mistral import (
|
||||
MistralConversationContentChunk,
|
||||
MistralConversationInputMessage,
|
||||
MistralConversationOutput,
|
||||
MistralConversationsResponse,
|
||||
MistralConversationUsage,
|
||||
|
|
@ -66,6 +65,16 @@ def _content_to_str(content: object) -> str:
|
|||
return "" if content is None else str(content)
|
||||
|
||||
|
||||
def _function_call_entry(call: dict[str, object]) -> dict[str, object]:
|
||||
fn = STR_OBJ_DICT.validate_python(call["function"]) if isinstance(call.get("function"), dict) else {}
|
||||
return {
|
||||
"type": "function.call",
|
||||
"tool_call_id": str(call.get("id") or ""),
|
||||
"name": str(fn.get("name") or ""),
|
||||
"arguments": str(fn.get("arguments") or ""),
|
||||
}
|
||||
|
||||
|
||||
def _to_annotation(chunk: MistralConversationContentChunk) -> ChatCompletionAnnotation:
|
||||
url_citation: ChatCompletionAnnotationURLCitation = (
|
||||
{"url": chunk.url or "", "title": chunk.title} if chunk.title else {"url": chunk.url or ""}
|
||||
|
|
@ -131,6 +140,37 @@ class MistralConversationsConfig(MistralConfig):
|
|||
)
|
||||
return {**base, **({"random_seed": random_seed} if random_seed is not None else {})}
|
||||
|
||||
@staticmethod
|
||||
def _input_entries_for_message(message: AllMessageValues) -> tuple[dict[str, object], ...]:
|
||||
"""Map one OpenAI message to its Conversations ``inputs`` entries.
|
||||
|
||||
A ``tool`` message becomes a ``function.result`` entry; an assistant
|
||||
message's ``tool_calls`` each become a ``function.call`` entry (preserving
|
||||
the id/name/arguments binding); text content becomes a plain message
|
||||
entry. Preserving this history keeps a web-search turn mid-conversation
|
||||
from silently dropping prior tool calls the way a role+content flatten would.
|
||||
"""
|
||||
typed = STR_OBJ_DICT.validate_python(message)
|
||||
role = str(typed.get("role") or "")
|
||||
if role == "tool":
|
||||
return (
|
||||
{
|
||||
"type": "function.result",
|
||||
"tool_call_id": str(typed.get("tool_call_id") or ""),
|
||||
"result": _content_to_str(typed.get("content")),
|
||||
},
|
||||
)
|
||||
raw_calls = typed.get("tool_calls") if role == "assistant" else None
|
||||
calls = OBJ_LIST.validate_python(raw_calls) if isinstance(raw_calls, list) else []
|
||||
content = _content_to_str(typed.get("content"))
|
||||
message_entry: tuple[dict[str, object], ...] = (
|
||||
({"role": role, "content": content},) if content or not calls else ()
|
||||
)
|
||||
call_entries = tuple(
|
||||
_function_call_entry(STR_OBJ_DICT.validate_python(call)) for call in calls if isinstance(call, dict)
|
||||
)
|
||||
return message_entry + call_entries
|
||||
|
||||
def transform_request(
|
||||
self,
|
||||
model: str,
|
||||
|
|
@ -146,12 +186,10 @@ class MistralConversationsConfig(MistralConfig):
|
|||
if message.get("role") == "system" and _content_to_str(message.get("content"))
|
||||
)
|
||||
inputs = [
|
||||
MistralConversationInputMessage(
|
||||
role=str(message.get("role")),
|
||||
content=_content_to_str(message.get("content")),
|
||||
)
|
||||
entry
|
||||
for message in messages
|
||||
if message.get("role") != "system"
|
||||
if str(message.get("role")) != "system"
|
||||
for entry in self._input_entries_for_message(message)
|
||||
]
|
||||
completion_args = self._build_completion_args(params)
|
||||
return {
|
||||
|
|
|
|||
|
|
@ -51,11 +51,41 @@ class TestMistralCompletion(BaseLLMChatTest):
|
|||
|
||||
response = litellm.completion(
|
||||
model=model,
|
||||
messages=[
|
||||
{"role": "user", "content": "What's the weather like in Boston today?"}
|
||||
],
|
||||
messages=[{"role": "user", "content": "What's the weather like in Boston today?"}],
|
||||
web_search_options={},
|
||||
max_tokens=100,
|
||||
)
|
||||
|
||||
assert response is not None
|
||||
|
||||
def test_web_search_with_tool_call_history(self):
|
||||
"""The Conversations API accepts prior function-call history mapped to
|
||||
function.call / function.result input entries alongside a web search
|
||||
request (a 400 here means the mapped wire schema is wrong)."""
|
||||
if not os.getenv("MISTRAL_API_KEY"):
|
||||
pytest.skip("MISTRAL_API_KEY not set")
|
||||
|
||||
response = litellm.completion(
|
||||
model="mistral/mistral-medium-latest",
|
||||
messages=[
|
||||
{"role": "user", "content": "What's the weather in Paris?"},
|
||||
{
|
||||
"role": "assistant",
|
||||
"content": "",
|
||||
"tool_calls": [
|
||||
{
|
||||
"id": "call_1",
|
||||
"type": "function",
|
||||
"function": {"name": "get_weather", "arguments": '{"city": "Paris"}'},
|
||||
}
|
||||
],
|
||||
},
|
||||
{"role": "tool", "tool_call_id": "call_1", "content": "18C and sunny"},
|
||||
{"role": "user", "content": "Now search the web for who won Euro 2024 and cite sources."},
|
||||
],
|
||||
web_search_options={},
|
||||
max_tokens=200,
|
||||
)
|
||||
|
||||
assert response is not None
|
||||
assert response.choices[0].message.content
|
||||
|
|
|
|||
|
|
@ -113,6 +113,38 @@ def test_transform_request_premium_and_function_passthrough():
|
|||
]
|
||||
|
||||
|
||||
def test_transform_request_preserves_tool_call_history():
|
||||
cfg = MistralConversationsConfig()
|
||||
body = cfg.transform_request(
|
||||
model="mistral-medium-latest",
|
||||
messages=[
|
||||
{"role": "user", "content": "weather in Paris?"},
|
||||
{
|
||||
"role": "assistant",
|
||||
"content": "",
|
||||
"tool_calls": [
|
||||
{
|
||||
"id": "call_1",
|
||||
"type": "function",
|
||||
"function": {"name": "get_weather", "arguments": '{"city":"Paris"}'},
|
||||
}
|
||||
],
|
||||
},
|
||||
{"role": "tool", "tool_call_id": "call_1", "content": "sunny"},
|
||||
{"role": "user", "content": "now cite sources"},
|
||||
],
|
||||
optional_params={"web_search_options": {}},
|
||||
litellm_params={},
|
||||
headers={},
|
||||
)
|
||||
assert body["inputs"] == [
|
||||
{"role": "user", "content": "weather in Paris?"},
|
||||
{"type": "function.call", "tool_call_id": "call_1", "name": "get_weather", "arguments": '{"city":"Paris"}'},
|
||||
{"type": "function.result", "tool_call_id": "call_1", "result": "sunny"},
|
||||
{"role": "user", "content": "now cite sources"},
|
||||
]
|
||||
|
||||
|
||||
def test_transform_request_drops_other_builtin_connectors():
|
||||
cfg = MistralConversationsConfig()
|
||||
body = cfg.transform_request(
|
||||
|
|
|
|||
Loading…
Add table
Reference in a new issue