Skip to main content

19. Pack: doc.ocr

Status: Accepted Date: 2026-04-07 Domain: api-design

Context​

OCR pipelines require image preprocessing, Tesseract invocation, and language-pack management — boilerplate that should not be in agent code (PRD §6.6).

Decision​

Ship doc.ocr as a built-in pack.

Input: { file_url: string, language?: string (ISO 639-2, default "eng"), preprocess?: boolean } Output: { text: string, confidence: number, page_count: integer } Errors: not_found, timeout, internal_error

The handler downloads the file, optionally runs preprocessing (deskew, denoise, contrast), invokes Tesseract with the requested language pack, and returns extracted text plus an aggregate confidence score. PDFs are split per-page automatically.

Consequences​

Positive: OCR becomes a single typed call; language packs are bundled in the session image. Negative: image bloat from extra languages; quality varies by source document.

§6.6 Capability Packs