fix(Haystack): Component Name retrieval

[lambda-science]

Haystack 2.10 Changed internal methods.
Openinference integration of Haystack is not compatible anymore.

File "/code/pythia_api/endpoints/chat.py", line 197, in _run_pipeline
2025-02-13T13:01:28.385766719Z     result = chat_pipeline.run({
2025-02-13T13:01:28.385770416Z              ^^^^^^^^^^^^^^^^^^^
2025-02-13T13:01:28.385774053Z   File "/code/.venv/lib/python3.12/site-packages/openinference/instrumentation/haystack/_wrappers.py", line 189, in __call__
2025-02-13T13:01:28.385777970Z     response = wrapped(*args, **kwargs)
2025-02-13T13:01:28.385782018Z                ^^^^^^^^^^^^^^^^^^^^^^^^
2025-02-13T13:01:28.385785715Z   File "/code/.venv/lib/python3.12/site-packages/haystack/core/pipeline/pipeline.py", line 248, in run
2025-02-13T13:01:28.385789592Z     component_outputs = self._run_component(component, inputs, component_visits, parent_span=span)
2025-02-13T13:01:28.385793259Z                         ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2025-02-13T13:01:28.385797017Z   File "/code/.venv/lib/python3.12/site-packages/openinference/instrumentation/haystack/_wrappers.py", line 73, in __call__
2025-02-13T13:01:28.385808388Z     component_name = pipe_args["name"]
2025-02-13T13:01:28.385812416Z                      ~~~~~~~~~^^^^^^^^
2025-02-13T13:01:28.385816894Z KeyError: 'name'

This MR tries to fix it with a few informations from Haystack maintainers. I'm not sure my MR fix the problems, I don't have so much time to test...

However Haystack Maintener pointed that OpenInference integration uses private methods which are subject to breaking change....
They said:

The more sustainable version would be to implement a proper custom tracing backend as outlined in our docs. The tracing backend is a public API which we try not to break. We really can't guarantee no changes to internal methods though. So moving to the tracing backend would prevent future breakage.

I think it should be re-factored

[lambda-science]

Should update test to use haystack 2.10

[lambda-science]

Well, it looks too broken for me to spend more time on this.
If you want to do a long term integratin that will not be subject to broken changes here is some info:
https://docs.haystack.deepset.ai/docs/tracing#custom-tracing-backend

Have a good day

[jatingarg0908]

Hey also you have to make a deep copy for inputs . and then

• Extract the inputs needed to run for the component
• Update the inputs with the default values for the inputs that are missing
  Then you have to send this processed deep copy of inputs to _get_bounds_arguments.
  This should be the same as Haystack pipeline._run_component method is doing.
  reference : https://github.com/deepset-ai/haystack/blob/0409e5da8f862bfdbb18850b17b4bed48f060cc2/haystack/core/pipeline/pipeline.py#L45

[nate-mar]

@lambda-science we appreciate you taking a pass on this!! I can pick up from here. Thanks again for the effort!

[axiomofjoy]

Hey also you have to make a deep copy for inputs . and then

• Extract the inputs needed to run for the component
• Update the inputs with the default values for the inputs that are missing
  Then you have to send this processed deep copy of inputs to _get_bounds_arguments.
  This should be the same as Haystack pipeline._run_component method is doing.
  reference : https://github.com/deepset-ai/haystack/blob/0409e5da8f862bfdbb18850b17b4bed48f060cc2/haystack/core/pipeline/pipeline.py#L45

Thanks @jatingarg0908. I see what you mean now. The changes from 2.9 make this challenging. Probably going to try to directly instrument the component run method instead if I can.

[axiomofjoy]

This issue should be resolved in openinference-instrumentation-haystack==0.1.18

[axiomofjoy]

Closing this PR out, but thanks so much for your investigations @lambda-science @jatingarg0908 !