This is a comment left during a code review.
Path: nemo_retriever/src/nemo_retriever/ocr/shared.py
Line: 802

Comment:
**Breaking rename of `ocr_v1` → `ocr` column**

Renaming the output column from `ocr_v1` to `ocr` is a breaking change for any caller that reads this column from the result DataFrame. Both `OCRActor` (v1) and the new `OCRV2Actor` share this function, so v1 consumers lose their existing column name without a deprecation cycle. Per the `api-backward-compatibility` rule, fields must not be removed or renamed without a deprecation cycle and documentation notice. Consider keeping `ocr_v1` for the v1 path (passing a parameter, as suggested in the previous review) and using `ocr` only for the new v2 path — or at minimum documenting this as a breaking change in the PR description and changelog.

**Rule Used:** Changes to public API surfaces (FastAPI endpoints,... ([source](https://app.greptile.com/review/custom-context?memory=api-backward-compatibility))

How can I resolve this? If you propose a fix, please make it concise.

---

This is a comment left during a code review.
Path: nemo_retriever/src/nemo_retriever/ocr/gpu_ocrv2.py
Line: 1-20

Comment:
**No unit tests for new OCRV2 actors**

`gpu_ocrv2.py`, `cpu_ocrv2.py`, and `nemotron_ocr_v2.py` are new source files with zero dedicated tests. The test file changes only update column-name assertions in pre-existing v1 tests. Per the `test-mirrors-source-structure` and `test-coverage-new-code` rules, new business logic must include unit tests covering at least the happy path and the error path. At a minimum, `OCRV2Actor.__call__` error recovery, `OCRV2CPUActor` initialization defaults, and `NemotronOCRV2.invoke` for the tensor/base64 paths should be covered.

**Rule Used:** New functionality must include corresponding unit ... ([source](https://app.greptile.com/review/custom-context?memory=test-coverage-new-code))

How can I resolve this? If you propose a fix, please make it concise.

---

This is a comment left during a code review.
Path: nemo_retriever/src/nemo_retriever/model/local/nemotron_ocr_v2.py
Line: 178-179

Comment:
**Docstring claims file-path detection that the code does not perform**

The `invoke` docstring states `str` is treated as base64 "unless it is an existing file path", but lines 178-179 unconditionally encode any string to bytes and forward it to the model — no `os.path.exists()` check is performed. Either update the docstring to reflect the actual behavior (strings are always treated as base64), or add the file-path check as described.

```suggestion
        if isinstance(input_data, str):
            if os.path.isfile(input_data):
                with open(input_data, "rb") as f:
                    return self._model(f.read(), merge_level=merge_level)
            return self._model(input_data.encode("utf-8"), merge_level=merge_level)
```

How can I resolve this? If you propose a fix, please make it concise.