-
[OPIK-7263] [SDK] feat: add Mistral AI Python integration (track_mistral) (#7385)
frostbyte_neo 发布于
2026-07-08 08:22:21 +00:00 - [OPIK-7263] [SDK] feat: add Mistral AI Python integration (track_mistral)
Add a native Opik integration for the Mistral AI Python SDK (mistralai),
exposed as track_mistral(client). Wraps chat.complete / complete_async and
chat.stream / stream_async, aggregating streamed chunks into a single span.
Structured-output calls (chat.parse and variants) delegate to these
primitives, so they are tracked as a single span named chat_completion_parse
via the response_format kwarg — no double-logging.Captures input messages, output choices, token usage, model, and provider
("mistral"), enabling automatic cost tracking. Adds MISTRALAI to the
LLMProvider enum. Includes an offline test suite (fake_backend), a CI
workflow wired into the library-integration runner, and a documentation page
that leads with the native integration and demotes LiteLLM to an alternative.Implements OPIK-7263: Mistral AI Python SDK integration.
Co-Authored-By: Claude Opus 4.8 (1M context) noreply@anthropic.com
- fix(mistral): name spans by wrapped method, not response_format
Naming the span chat_completion_parse whenever response_format was present
misclassified direct complete(response_format=...) / stream(response_format=...)
calls as parse. Name spans after the wrapped primitive instead
(complete -> chat_completion_create, stream -> chat_completion_stream); parse
delegates to those and shares the name. Also rename the provider-enum test to
the required test_ convention and fix an "opik" ->
"Opik" doc typo. Addresses PR #7385 review feedback.Co-Authored-By: Claude Opus 4.8 (1M context) noreply@anthropic.com
- fix(mistral): pin mistralai to v1 in tests and docs
CI installed mistralai unpinned and resolved to 2.x, which is a namespace
package wherefrom mistralai import Mistralno longer works (Mistral moved
under mistralai.client), breaking test collection. The integration targets the
mainstream v1 API, so pin the test requirement to mistralai>=1.0.0,<2.0.0 and
update the docs install command to "mistralai<2".Co-Authored-By: Claude Opus 4.8 (1M context) noreply@anthropic.com
- refactor(mistral): use only public mistralai API + guard min version
- Stop importing internal modules (mistralai.utils.eventstreaming,
mistralai.models.chatcompletionresponse) that move between releases. Use the
public top-level ChatCompletionResponse, and detect streams by their iterator
protocol (anext/next) operating on type(stream) rather than the
private EventStream classes. - Add a runtime version guard in track_mistral: raise a clear error when the
installed mistralai is older than 1.3.0 (the release that introduced the
EventStream streaming API), read via importlib.metadata. Pin the test
requirement and docs install to mistralai>=1.3.0,<2.
Co-Authored-By: Claude Opus 4.8 (1M context) noreply@anthropic.com
- feat(llm_usage): add dedicated Mistral usage format
Add MistralUsage to the llm_usage namespace (parses UsageInfo:
prompt/completion/total tokens + prompt_tokens_details.cached_tokens) with an
OpikUsage.from_mistral_dict builder registered under LLMProvider.MISTRALAI.
The Mistral decorator now parses usage with LLMProvider.MISTRALAI instead of
piggybacking on the OpenAI completions converter, so it's not coupled to
OpenAI's format. Adds unit tests for the parser and factory path.Co-Authored-By: Claude Opus 4.8 (1M context) noreply@anthropic.com
- docs(skills): add dedicated-usage-format guidance to opik-integrations
Capture the lesson from the Mistral usage work: add a first-class
llm_usage format per provider instead of reusing another provider's parser.Co-Authored-By: Claude Opus 4.8 (1M context) noreply@anthropic.com
- fix(mistral): address review nits (dead log arg, import order)
- Drop the unused str(exception) positional passed to LOGGER.debug with no
format placeholder in the stream patchers; exc_info=True already captures the
full exception + traceback. - Reorder mistral_usage imports to stdlib-first (typing, then pydantic, then
local), matching the anthropic/bedrock usage modules.
Co-Authored-By: Claude Opus 4.8 (1M context) noreply@anthropic.com
- fix(mistral): accumulate streamed tool calls by index
The stream aggregator overwrote message.tool_calls on every chunk that carried
them, which would drop data if a tool call were split across chunks. Accumulate
by tool-call index instead, concatenating streamed argument fragments and
filling in id/name/type as they arrive. Mistral currently emits each tool call
complete in a single chunk, so this is defensive, but it removes the
overwrite-and-keep-last behavior. Adds a streaming tool-call test asserting the
aggregated span output keeps the complete tool call (name + valid-JSON args).Co-Authored-By: Claude Opus 4.8 (1M context) noreply@anthropic.com
Co-authored-by: Claude Opus 4.8 (1M context) noreply@anthropic.com
下载附件