README.md

---
pipeline_tag: other
tags:
  - document-classification
  - medical
---

# ArcaAI Document Classifier

> **Status: in development.** The pipeline, API contracts, label set, and
> endpoints may change without notice. Not yet recommended for production use.

A custom Hugging Face pipeline for classifying medical / health-related
documents into one of six types:

| Label | Meaning |
| --- | --- |
| `bill` | Hospital bill / invoice |
| `discharge_summary` | Discharge summary |
| `echs_card` | ECHS (Ex-Servicemen Contributory Health Scheme) card |
| `lab_report` | Laboratory report |
| `prescription_slip` | Prescription slip |
| `referral_letter_by_polyclinic` | Referral letter from a polyclinic |

Classification is performed by the **ArcaAI hosted inference API** (a
fine-tuned document model served over Triton). This repository publishes a
**weightless** custom pipeline: it loads configuration only and delegates the
actual inference to the API. **No model weights are downloaded.**

## Quickstart

```bash
pip install transformers torch requests

export DOCCLASSIFIER_API_KEY="dc_xxx"   # token from the ArcaAI admin API
```

```python
from transformers import pipeline

classifier = pipeline(
    "document-classification",
    model="akbted/docclassifier",
    trust_remote_code=True,
)

print(classifier("invoice.pdf"))

print(classifier(["invoice.pdf", "lab_report.png", "prescription.pdf"]))
```

`trust_remote_code=True` is required: the repo's `pipeline.py` is executed
locally to make the API call.

## Supported inputs

- PDFs and images (`pdf`, `png`, `jpg`, `jpeg`).
- File paths (`str`), raw bytes, or a list of either.
- Results are **per page**: `[{"file_path", "page", "label", "confidence"}, ...]`.

## Configuration

| Env var | Purpose |
| --- | --- |
| `DOCCLASSIFIER_API_KEY` | API token (required). |
| `DOCCLASSIFIER_API_URL` | Override the endpoint baked into the model config. |

Threshold filtering per call:

```python
classifier("invoice.pdf", confidence_threshold=0.9)
```

## Try it

```python
from transformers import pipeline

classifier = pipeline(
    "document-classification",
    model="akbted/docclassifier",
    trust_remote_code=True,
)

for page in classifier("invoice.pdf"):
    print(f"page={page['page']}  label={page['label']}  confidence={page['confidence']}")
```

## Development status & limitations

- In development; subject to breaking changes.
- Requires network access to the ArcaAI API and a valid API token.
- API requests are rate-limited per token.
- The stand-in architecture ships **no weights** and is a loading vehicle only:
  `AutoModel.from_pretrained("akbted/docclassifier", trust_remote_code=True)`
  returns a weightless shell by design.

## Getting an API token

Contact the ArcaAI team to provision a token for the public API.

## License

See `LICENSE` (to be added).