feat(python): openai instrumentator #35
Conversation
RogerHYang
changed the title
python/instrumentation/openinference-instrumentation-openai/README.rst
Outdated
Show resolved
Hide resolved
| """ | ||
| Phoenix collector should be running in the background. | ||
| """ |
There was a problem hiding this comment.
Non-blocking. - I think what might make sense long term is to have small example apps that have a README and project toml. That way we can stack multiple instrumentations together.
There was a problem hiding this comment.
Yea, that's a good idea!
|
|
||
| resource = Resource(attributes={}) | ||
| tracer_provider = trace_sdk.TracerProvider(resource=resource) | ||
| span_exporter = OTLPSpanExporter(endpoint="http://127.0.0.1:6006/v1/traces") |
There was a problem hiding this comment.
thought: Let's brainstorm how to maybe just run some of these examples via docker-compose. That way it will be easier to showcase phoenix.
python/instrumentation/openinference-instrumentation-openai/integration_tests/completions.py
Outdated
Show resolved
Hide resolved
| cast_to: type, | ||
| request_options: Mapping[str, Any], | ||
| ) -> Iterator[_WithSpan]: | ||
| span_kind = _EMBEDDING_SPAN_KIND if cast_to is CreateEmbeddingResponse else _LLM_SPAN_KIND |
There was a problem hiding this comment.
thought: I get that in this context everything is probably an LLM or Embedding but this read a bit presumptive. Maybe this could be in a function _get_span_kind_from_type or something like that that makes it a bit more clear? Falling through to _LLM_SPAN_KIND feels a tad dangerous - I'm guessing assistant API doesn't hit this but might be better to fallback to unknown?
There was a problem hiding this comment.
yea, agree. i'll update. On the other hand maybe the span name should come from a higher level caller: by the time the Request function is called, we don't really know who called it and for what purpose. An alternative is to also monkey patch those higher level callers, but instead of tracing them per se, just have them put the real span_name into the context and pass it down to request (and then use the fallback in request if span_name doesn't exist.
...on/openinference-instrumentation-openai/src/openinference/instrumentation/openai/_request.py
Outdated
Show resolved
Hide resolved
...tion/openinference-instrumentation-openai/src/openinference/instrumentation/openai/_utils.py
Show resolved
Hide resolved
...tion/openinference-instrumentation-openai/src/openinference/instrumentation/openai/_utils.py
Outdated
Show resolved
Hide resolved
resolves #1961