Safe to merge; all findings are P2 style/quality suggestions with no blocking correctness issues.

The archetype resolution logic is sound, tests cover remote/local/GPU paths and hardware-based dispatch, and the public API is cleanly exported. The only findings are unreachable dead code from an incomplete cleanup and a potential silent zip truncation that depends on model behavior.

nemo_retriever/src/nemo_retriever/audio/asr_actor.py — unreachable _transcribe_local path and silent zip truncation in _call_local_batch.

1. nemo_retriever/src/nemo_retriever/audio/asr_actor.py, line 372-376 (link)

   _transcribe_local is unreachable dead code

   _transcribe_one is only ever called from _call_remote_batch (line 177), which itself is only called when self._client is not None (line 166). Therefore the else branch here — and the entire _transcribe_local method — can never be reached in normal execution. This appears to be an incomplete refactoring: local-model transcription is now handled exclusively in _call_local_batch, but the per-row fallback in _transcribe_one was not removed.

   Consider either removing _transcribe_local and the else branch, or restructuring _call_remote_batch to reuse _call_local_batch for rows where the remote call fails.

   Prompt To Fix With AI

   This is a comment left during a code review.
   Path: nemo_retriever/src/nemo_retriever/audio/asr_actor.py
   Line: 372-376
   
   Comment:
   **`_transcribe_local` is unreachable dead code**
   
   `_transcribe_one` is only ever called from `_call_remote_batch` (line 177), which itself is only called when `self._client is not None` (line 166). Therefore the `else` branch here — and the entire `_transcribe_local` method — can never be reached in normal execution. This appears to be an incomplete refactoring: local-model transcription is now handled exclusively in `_call_local_batch`, but the per-row fallback in `_transcribe_one` was not removed.
   
   Consider either removing `_transcribe_local` and the `else` branch, or restructuring `_call_remote_batch` to reuse `_call_local_batch` for rows where the remote call fails.
   
   How can I resolve this? If you propose a fix, please make it concise.

2. nemo_retriever/src/nemo_retriever/audio/asr_actor.py, line 238-247 (link)

   Silent row truncation if model returns fewer transcripts than paths

   zip(rows_list, transcripts) silently drops tail rows when transcripts is shorter than rows_list. This can happen if self._model.transcribe skips or omits results for empty-string paths ("") that are appended to paths_for_model when path resolution fails (e.g. both path and bytes are missing). Adding a length check with a warning would surface this early.

   Prompt To Fix With AI

   This is a comment left during a code review.
   Path: nemo_retriever/src/nemo_retriever/audio/asr_actor.py
   Line: 238-247
   
   Comment:
   **Silent row truncation if model returns fewer transcripts than paths**
   
   `zip(rows_list, transcripts)` silently drops tail rows when `transcripts` is shorter than `rows_list`. This can happen if `self._model.transcribe` skips or omits results for empty-string paths (`""`) that are appended to `paths_for_model` when path resolution fails (e.g. both `path` and `bytes` are missing). Adding a length check with a warning would surface this early.
   
   How can I resolve this? If you propose a fix, please make it concise.

3. nemo_retriever/src/nemo_retriever/audio/asr_actor.py, line 420-433 (link)

   apply_asr_to_df kwargs are accepted but ignored

   The **kwargs: Any parameter is declared but never forwarded to actor() or ASRParams. Either pass it through or remove the parameter to avoid silently swallowing caller-supplied options.

   Prompt To Fix With AI

   This is a comment left during a code review.
   Path: nemo_retriever/src/nemo_retriever/audio/asr_actor.py
   Line: 420-433
   
   Comment:
   **`apply_asr_to_df` kwargs are accepted but ignored**
   
   The `**kwargs: Any` parameter is declared but never forwarded to `actor()` or `ASRParams`. Either pass it through or remove the parameter to avoid silently swallowing caller-supplied options.
   
   How can I resolve this? If you propose a fix, please make it concise.

Safe to merge after addressing the local-batch error-handling gap; the archetype resolution logic and GPU/CPU split are otherwise sound.

One P1 issue: _call_local_batch has no per-row error isolation (unlike the remote path) and passes empty-string sentinel paths to the model, which can crash an entire batch; the zip truncation can also silently drop rows. The rest of the design is correct and well-tested.

nemo_retriever/src/nemo_retriever/audio/asr_actor.py — specifically _call_local_batch (lines 188–253) and _transcribe_local/else-branch in _transcribe_one (lines 270–376)

1. nemo_retriever/src/nemo_retriever/audio/asr_actor.py, line 238-246 (link)

   Missing per-row error isolation in local batch path

   The remote batch path (_call_remote_batch) wraps each row in a try/except so one bad file only skips that row. The local batch path has no equivalent guard: if self._model.transcribe(paths_for_model) raises on any input (including the "" strings appended for rows where both path and raw were unavailable — line 235), the entire batch crashes. Additionally, when the model returns fewer transcripts than input paths (e.g., it silently skips empty-path entries), the bare zip on line 246 silently truncates output rows, causing undetected data loss.

   Consider filtering empty paths before calling the model, and using itertools.zip_longest (with a fallback of "") to ensure every input row produces an output row.

   Prompt To Fix With AI

   This is a comment left during a code review.
   Path: nemo_retriever/src/nemo_retriever/audio/asr_actor.py
   Line: 238-246
   
   Comment:
   **Missing per-row error isolation in local batch path**
   
   The remote batch path (`_call_remote_batch`) wraps each row in a `try/except` so one bad file only skips that row. The local batch path has no equivalent guard: if `self._model.transcribe(paths_for_model)` raises on *any* input (including the `""` strings appended for rows where both `path` and `raw` were unavailable — line 235), the entire batch crashes. Additionally, when the model returns fewer transcripts than input paths (e.g., it silently skips empty-path entries), the bare `zip` on line 246 silently truncates output rows, causing undetected data loss.
   
   Consider filtering empty paths before calling the model, and using `itertools.zip_longest` (with a fallback of `""`) to ensure every input row produces an output row.
   
   How can I resolve this? If you propose a fix, please make it concise.

2. nemo_retriever/src/nemo_retriever/audio/asr_actor.py, line 270-287 (link)

   _transcribe_local is unreachable dead code

   _transcribe_local is only invoked from the else branch of _transcribe_one (lines 372–376), but _transcribe_one is exclusively called from _call_remote_batch, which is itself guarded by if self._client is not None in process. This means the else branch in _transcribe_one — and therefore _transcribe_local entirely — can never execute. The local single-row transcription logic is also re-implemented (with differences) inside _call_local_batch, adding to the confusion.

   Prompt To Fix With AI

   This is a comment left during a code review.
   Path: nemo_retriever/src/nemo_retriever/audio/asr_actor.py
   Line: 270-287
   
   Comment:
   **`_transcribe_local` is unreachable dead code**
   
   `_transcribe_local` is only invoked from the `else` branch of `_transcribe_one` (lines 372–376), but `_transcribe_one` is exclusively called from `_call_remote_batch`, which is itself guarded by `if self._client is not None` in `process`. This means the `else` branch in `_transcribe_one` — and therefore `_transcribe_local` entirely — can never execute. The local single-row transcription logic is also re-implemented (with differences) inside `_call_local_batch`, adding to the confusion.
   
   How can I resolve this? If you propose a fix, please make it concise.