henry3260 opened a new pull request, #74751:
URL: https://github.com/apache/airflow/pull/74751
## Why
`LLMBatchOperator.request_params` carries free-form provider body parameters
(`temperature`, `top_k`, `seed`, `thinking`, ...) merged into every request.
It is the batch
counterpart of `LLMOperator.agent_params` / `AgentOperator.agent_params`,
both of which are in
`template_fields`, so a Dag author who writes
`request_params={"temperature": "{{ params.temp }}"}`
reasonably expects it to be rendered.
It was not in `template_fields`. The Jinja expression stayed a literal
string and was sent to the
provider API as the parameter value, which fails at the provider with
nothing pointing back at the
operator parameter as the cause.
`max_tokens` is deliberately left alone. Every scalar numeric parameter in
this provider
(`max_tokens`, `poll_interval`, `timeout`,
`LLMFileAnalysisOperator.max_files`) is untemplated;
only free-form dicts (`agent_params`, `embedding_kwargs`) and dicts with a
known numeric schema
(`usage_limits`, which also coerces each rendered value) are templated.
Templating a bare `int`
would hand the provider API the string Jinja produced.
## What
-
`providers/common/ai/src/airflow/providers/common/ai/operators/llm_batch.py` —
add
`request_params` to `template_fields`, and document the templating in its
`:param:` entry.
No coercion is applied, matching `agent_params`: the keys are provider
body parameters whose
types the operator does not know.
- `providers/common/ai/docs/operators/llm_batch.rst` — document that
`request_params` is
templated, that Jinja hands back a string (so a numeric value written as a
template needs
`render_template_as_native_obj=True` on the Dag), and that the rendered
values feed the
re-attach fingerprint — the same stability rule that already applies to
`result_path`.
- `providers/common/ai/tests/unit/common/ai/operators/test_llm_batch.py` —
add
`TestTemplateFields`: one test pinning the `template_fields` set, one
rendering
`{"temperature": "{{ params.temp }}"}` through `render_template_fields`.
Both fail without the
operator change.
Verified locally: the 50 tests in `test_llm_batch.py` pass; `prek run
--from-ref upstream/main`
is clean on both the `pre-commit` and `manual` stages. The
Docker-image-based checks `breeze
verify` lists (provider test suites, `build-docs`, `mypy-providers`) were
not run locally — the
Breeze image was not built on this machine — so CI covers those.
---
##### Was generative AI tooling used to co-author this PR?
- [X] Yes — Claude Code (Opus 5)
Generated-by: Claude Code (Opus 5) following [the
guidelines](https://github.com/apache/airflow/blob/main/contributing-docs/05_pull_requests.rst#gen-ai-assisted-contributions)
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
