# text.extract

Plain-text extraction from PDF and DOCX documents. Pure-Rust,
in-WASM, no permissions required. The canonical first step in
any flow that needs raw text — anonymise, summarise, translate,
LLM-classify, audit-redact.

## Capability

- `text.extract@0.1.0`

## Inputs

| Name       | Type  | Description                                            |
| ---------- | ----- | ------------------------------------------------------ |
| `document` | bytes | The document to extract. Accepted MIME types below.    |

Accepted MIME types (declared in `module.yaml#accepts_mime`,
read by Studio as the file-picker filter):

- `application/pdf`
- `application/x-pdf`
- `application/vnd.openxmlformats-officedocument.wordprocessingml.document`

Other MIME types are rejected with a `invalid_input` error
before any parsing happens.

## Output

| Name        | Type | Description                                       |
| ----------- | ---- | ------------------------------------------------- |
| `extracted` | json | `ExtractedDocument` schema below.                 |

```json
{
  "engine": "pdf-extract",
  "engine_version": "0.7",
  "pages": [
    { "number": 1, "text": "…" }
  ]
}
```

Per-page `confidence` and `bbox` fields are reserved in the
schema for a future service-mode variant (pdfium / mupdf via
host services) that surfaces layout information. v0.1.0 does
not populate them; the schema accepts them as optional.

## Permissions

None. Pure-Rust pipeline (`pdf-extract`, `zip`, `quick-xml`).
No filesystem, no network. The empty `permissions: []` in
`module.yaml` makes that explicit; the hub enforces it.

## Limits in v0.1.0

- **Layout-naive PDF.** Multi-column or table-heavy documents
  may produce reordered text. Single-column body text is solid.
- **No OCR.** Scanned (image-only) PDFs return empty text per
  page. Pair with a future `text.ocr@^0` step for those.
- **DOCX-only Office.** `.doc` (legacy binary), `.rtf`, `.odt`
  are not yet supported.

Bumping to `1.0.0` requires either the service-mode variant
(pdfium-quality output + layout) or a confirmed-good extraction
score on a representative corpus.

## Example flow

```yaml
name: extract-anonymise
inputs:
  document: bytes
steps:
  - id: extract
    use: text.extract@^0
    with:
      document: $inputs.document
  - id: redact
    use: text.anonymize@^0
    with:
      text: $extract.extracted.pages[*].text
outputs:
  redacted: $redact.anonymized
```

## Build

```bash
cargo build --release --target wasm32-wasip2
# Output: target/wasm32-wasip2/release/text_extract.wasm
```