Frequently Asked Questions¶

These answers cover the questions that come up most often when teams start using OpenMed in local clinical NLP, PII detection, de-identification, and service deployments.

Hitting a concrete error rather than a conceptual question? See Troubleshooting & Common Errors for a symptom → cause → fix map of the failures users run into most (missing extras, model-download/offline issues, device selection, and REST/MCP setup).

Install and Run¶

Can OpenMed run fully locally or in an air-gapped environment?¶

Yes. OpenMed can run without sending clinical text to an external service. For strict offline use, pre-download the model files, point model_name or model_id at that local directory, and keep the runtime on device="cpu" or another device available inside your environment.

When the identifier is an existing local path, OpenMed asks the underlying loader to use local_files_only=True, so missing tokenizer, config, or weight files fail locally instead of downloading from the model hub. See Loading from a local path.

After warming the configured cache, set OPENMED_OFFLINE=1 or use OpenMedConfig(local_only=True). OpenMed then enables the Hugging Face cache-only flags, passes local_files_only=True to Hub-backed loaders, and blocks outbound sockets during inference and de-identification. See Local-only offline mode and the offline troubleshooting entry. For institutional pip mirrors, HF_ENDPOINT, HTTP proxies, resumable cache warming, and a metered-connection checklist, use the low-bandwidth installation guide.

Which package extras should I install?¶

Use the smallest extra that matches your runtime:

openmed[hf] for the standard Python model runtime.
openmed[hf,service] when you need the REST service.
openmed[mlx] for Python MLX acceleration on Apple Silicon.
openmed[multimodal] for document/image intake and Tesseract OCR; install the system tesseract binary separately (brew install tesseract on macOS or sudo apt-get install tesseract-ocr on Debian/Ubuntu).
openmed[ocr-paddle] for the heavier optional PaddleOCR backend.

Start with the Quick Start, then use Configuration & Validation for cache paths, device selection, profiles, and environment overrides.

Why does a DeBERTa model say that scaled dot-product attention is unsupported?¶

OpenMed 1.7.0 and 1.8.0 could select PyTorch scaled dot-product attention (sdpa) from runtime availability alone, even when the Transformers model architecture did not support it. This affected DeBERTa-v2 token-classification models and could occur on NVIDIA, AMD, or CPU environments.

Upgrade to OpenMed 1.8.1 or later. Automatic attention selection is architecture-safe in those releases. For an environment temporarily pinned to an affected release, select the universally compatible eager implementation before starting Python:

Linux/macOSWindows PowerShellWindows Command Prompt

export OPENMED_TORCH_ATTENTION_BACKEND=eager

$env:OPENMED_TORCH_ATTENTION_BACKEND="eager"

set OPENMED_TORCH_ATTENTION_BACKEND=eager

See PyTorch attention backends for the configuration API and explicit backend options.

Models and Languages¶

Which model should I use?¶

For clinical entity extraction, pick a registry alias that matches the entity family you need, such as disease, drug, anatomy, oncology, gene, or PII. The registry exposes metadata and helper functions for UI dropdowns, text-based suggestions, model sizes, and recommended confidence thresholds. See the Model Registry.

For PII, extract_pii(..., lang="<code>") selects the default model for the requested language when you keep the default model argument. Override model_name only when you need a specific checkpoint, local directory, or privacy-filter family.

Which languages are supported?¶

PII extraction and de-identification support 34 supported PII language codes: am, ar, as, bn, cs, da, de, el, en, es, fr, he, hi, id, it, ja, ko, mr, nl, no, or, pt, ro, ru, sv, sw, ta, te, th, tr, uk, xh, zh, and zu. Russian routing currently uses a documented multilingual default-model placeholder. Bengali, Chinese, and Tamil have dedicated registry entries. Four additional Indic codes (gu, kn, ml, and pa) are opt-in routes through a user-configured OPENMED_INDIC_NER_MODEL; Assamese, Bengali, Hindi, Marathi, Odia, Tamil, and Telugu can use the same adapter. Validator-backed national-ID coverage is broader for specific ID-only locales, including Polish, Latvian, Slovak, Malay, Filipino, and Finnish. The README keeps a short multilingual example set in Multilingual PII.

Clinical NER coverage depends on the selected registry model. Check each model's languages, entity_types, and specialization in the Model Registry before putting it behind an API or batch job.

Privacy and De-identification¶

Is de-identification reversible?¶

It depends on the method:

mask and remove do not preserve the original value in the output.
replace emits locale-aware synthetic surrogates; it is not reversible unless your own workflow stores an external mapping.
hash is one-way, but deterministic for linking repeated values.
shift_dates can be reversed only by someone who knows the shift amount.

Always review outputs before releasing data. PII detection is an assistive control, not a substitute for a privacy review process. See PII Anonymization.

Why are many national IDs grouped under `ID_NUM`?¶

OpenMed normalizes detector-specific labels into 50 canonical PII labels. ID_NUM is the canonical bucket for general identifiers such as medical record numbers, national IDs, CPF/CNPJ, NIR, Steuer-ID, Codice Fiscale, DNI/NIE, BSN, Aadhaar, and NPI. Labels with their own canonical category, such as SSN, ACCOUNT_NUMBER, or CREDIT_CARD, can still stay separate when the detector emits them clearly.

This keeps multilingual model output consistent while policy profiles can still treat identifiers as high-risk direct identifiers. The canonical taxonomy lives in openmed/core/labels.py.

Should I use reversible or irreversible de-identification?¶

Use irreversible output (mask, remove, or one-way hash) when the downstream workflow does not need the original values. Use replace when clinicians, QA reviewers, or demos need realistic-looking synthetic text. Use shift_dates only when preserving relative timelines matters and the offset can be governed like sensitive metadata.

Licensing¶

What license does OpenMed use?¶

The OpenMed package is released under Apache-2.0. See the repository license.

Model checkpoints may carry their own metadata, so verify the specific model card or registry row before redistributing weights or shipping them in a product bundle.

Performance¶

Does OpenMed require a GPU?¶

No. CPU execution is supported and is the default safe baseline for local and CI environments. GPU acceleration can improve latency and throughput for larger workloads:

CUDA devices can be selected with OpenMedConfig(device="cuda").
Apple Silicon systems can use the MLX backend when the relevant extra and model artifacts are available.
Batch processing can improve throughput for repeated extraction or de-identification jobs.

See Configuration & Validation, MLX Backend, Batch Processing, and Performance Profiling.

How do I keep memory usage predictable?¶

Reuse a ModelLoader or batch processor instead of creating a new pipeline for every document. In the REST service, use the model lifecycle endpoints to inspect and unload cached models. See ModelLoader & Pipelines and REST Service.