fix(groq): record only the parameters the caller passed

[feiiiiii5]

Description

The groq instrumentor records parameters that the caller never passed. With groq==1.0.0, this call:

client.chat.completions.create(
    messages=[{"role": "user", "content": "hi"}],
    model="llama-3.1-8b-instant",
)

records 33 entries in llm.invocation_parameters (1 expected) and 36 in input.value (2 expected):

{"model": "llama-3.1-8b-instant",
 "citation_options": "<groq.Omit object at 0x102ee8390>",
 "frequency_penalty": "<groq.Omit object at 0x102ee8390>",
 "max_tokens": "<groq.Omit object at 0x102ee8390>",
 "response_format": "<groq.Omit object at 0x102ee8390>",
 "seed": "<groq.Omit object at 0x102ee8390>",
 "stop": "<groq.Omit object at 0x102ee8390>",
 "temperature": "<groq.Omit object at 0x102ee8390>",
 "timeout": "NOT_GIVEN",
 "...": "24 more"}

So the trace claims the caller set temperature, max_tokens, top_p, seed, stop, tools, response_format, stream, … when they passed nothing, and 33 of those values are Python object addresses, so the same call records different text in every process.

Why it happens. _parse_args() binds the call, applies signature defaults, then relies on one identity test to drop the SDK's "not given" markers:

bound_signature.apply_defaults()
bound_arguments = bound_signature.arguments  # Defaults empty to NOT_GIVEN
...
if value is not None and value is not NOT_GIVEN:

That was true when the code was written and still holds on the version CI pins (test-requirements.txt → groq==0.9.0, where the optional defaults are the exported NOT_GIVEN singleton). The current SDK uses two other shapes, and neither passes the test:

parameter	default type	is groq.NOT_GIVEN	isinstance of
temperature, tools, stream, … (33)	Omit	False	groq.Omit
timeout	NotGiven	False (not the exported singleton)	groq.NotGiven

pyproject.toml declares "groq >= 0.9.0" with no upper bound, so the 1.x line is supported and ci-groq-latest installs it. This is the case python/AGENTS.md names directly: "Request parameters: filter out the SDK's 'not given' sentinels (e.g. Omit / NotGiven in Stainless-generated SDKs) so unset defaults don't pollute input.value and llm.invocation_parameters."

The change. Bind without applying defaults, so an unsupplied parameter never enters the recorded request — the approach cohere/_wrappers.py already documents for the same reason ("Map the call's arguments to parameter names, keeping only values the caller set"). The existing None / NOT_GIVEN guards stay, so a caller who passes a sentinel explicitly is still filtered. No sentinel class is imported, so nothing here depends on which marker a given groq release uses — which also matters because groq < 1.0 does not export Omit publicly at all.

Only the values of two existing attributes change, and only for groq>=1.0; no attribute key is added, renamed or removed, and no span name or hierarchy is touched. That is why there is no trace screenshot below: the before/after recorded values above are the whole observable difference.

Verification

Python 3.11, one venv with the three workspace packages installed editable (openinference-semantic-conventions, openinference-instrumentation, openinference-instrumentation-groq) plus opentelemetry-sdk, wrapt, responses, httpx<0.28. Commands run from python/instrumentation/openinference-instrumentation-groq.

$ python -m pytest tests/ -q

installed groq	origin/main (b737360)	this branch (25bb7ea)
1.0.0 (current; ci-groq-latest)	5 failed, 16 passed	21 passed
0.9.0 (CI pin; ci-groq)	21 passed	21 passed

The 5 added tests (tests/test_invocation_parameters.py) fail on main under groq>=1.0 for the reason described above, and the existing 16 keep passing before and after. Stated plainly: under the pinned groq==0.9.0 the new tests pass on main too, because that version's defaults really are the NOT_GIVEN singleton — they bite in the groq-latest matrix environment, which python-CI.yaml does run for this package.

What they cover: omitted parameters absent from llm.invocation_parameters; absent from input.value; no <... object at 0x...> in either recorded value; temperature=0.5, max_tokens=10, stop="END" still recorded (the fix does not over-filter); and the same result on the async path.

$ ruff==0.9.2  ruff format --diff src tests   → 12 files already formatted
$ ruff==0.9.2  ruff check --no-fix src tests  → All checks passed!
$ mypy==1.11.2 mypy .                         → Success: no issues found in 12 source files

Not verified here: no live groq call (no API key), no UI rendering, and no whole-repo tox run-parallel — the diff is confined to this package, so other packages' suites were not run.

Resolves #3781

Checklist:

• Follows OpenInference configuration to hide sensitive info — the diff only removes recorded data; it adds no field and no new value.
• Spans properly inherit from context attributes — untouched by this diff; the context-attribute and using_attributes tests in tests/test_instrumentor.py pass on this branch.
• Properly respects suppress tracing context — untouched; the suppress path in _start_as_current_span is not part of this diff and its tests pass.

Prepared with an AI coding assistant; the reproduction, tests and every command above were executed and inspected locally before submitting.