Skip to content

VISION.md

Latest commit

 

History

History
135 lines (99 loc) · 5.99 KB

VISION.md

File metadata and controls

135 lines (99 loc) · 5.99 KB
Raw
type knowledge
created 2026-02-03
tags
coocr-htr
vision
goals
status active

Project Vision

Mission Statement

coOCR/HTR is a browser-based tool that helps domain experts verify, validate, and correct OCR/HTR results.

Core Problem

Standard OCR/HTR pipelines often produce erroneous results on historical documents:

  • Unusual script forms (Kurrent, Fraktur, historical handwriting)
  • Complex layouts (tables, marginalia, strikethroughs)
  • Domain-specific vocabulary (technical terms, historical concepts)

These errors require human expertise to correct - but existing tools are often:

  • Complex and difficult to use
  • Not optimized for the correction workflow
  • Without AI support for difficult passages

Solution

coOCR/HTR positions itself as an Editor-in-the-Loop tool:

┌─────────────────────────────────────────────────────────────────┐
│                        WORKFLOW                                  │
├─────────────────────────────────────────────────────────────────┤
│                                                                  │
│   [Image/PAGE-XML]  ──►  [coOCR/HTR]  ──►  [Correct OCR/HTR]    │
│                              │                                   │
│                              ▼                                   │
│                     ┌─────────────────┐                         │
│                     │ Expert          │                         │
│                     │ - verifies      │                         │
│                     │ - validates     │                         │
│                     │ - corrects      │                         │
│                     └─────────────────┘                         │
│                                                                  │
└─────────────────────────────────────────────────────────────────┘

Two Input Modes

Mode Input Use Case
Generate OCR Upload image Document has no transcription yet
Correct OCR Upload PAGE-XML Transcription exists (e.g., from Transkribus)

AI Support

  • LLM Transcription: For difficult documents where standard OCR fails
  • Hybrid Validation: Deterministic rules + LLM Review for quality assessment
  • Visual Interface: Synchronized view of image, text, and validation

Target Audience

User Need
Digital Humanists OCR correction for edition projects
Archivists Fast transcription of holdings
Historians Source access with AI support
Citizen Scientists Accessible transcription work

Success Criteria

The product is complete when:

  1. Self-explanatory: Someone unfamiliar with the tool can use it without instructions
  2. Complete Workflow:
    • Upload own documents (image OR PAGE-XML)
    • Generate OCR or edit existing transcription
    • Validate and correct
    • Export in usable format (PAGE-XML, TXT, JSON)
  3. Workflow Integration: Output can be used in other processes
  4. Quality Assurance: "Good, correct OCR/HTR comes out the other side"

Adoption & Open Development

coOCR/HTR is positioned as an open-developed, community-driven tool. The development model follows an Open Development approach: the codebase is public, contributions are welcome, and institutional forks are explicitly encouraged.

Institutional Adoption

Institution Context Status
ZBZ (Zentralbibliothek Zürich) Jeanne Hersch Edition, 289 documents Fork planned (GitLab Uni Zürich, Podman)
ÖAW (Klugseder) Medieval music manuscripts Fork exists (reference implementation)

The ZBZ plans to fork co-ocr-htr for their OCR/TEI pipeline (see zbz-ocr-tei). Their deployment uses Podman (daemonless Docker alternative) on institutional infrastructure with Azure-based LLM access. This validates the tool's architecture: browser-based, no backend dependencies, configurable API endpoints.

Community of Experts

The Open Development vision positions domain experts (archivists, philologists, historians) as peer reviewers of LLM output — not just users, but co-developers of quality standards. Each institutional fork can contribute validation rules, prompts, and domain knowledge back to the main project.

Non-Goals

  • Not a replacement for specialized HTR models (Transkribus, eScriptorium)
  • No training tool for custom models

Current Status

Phase Status Description
Phase 1: Core Application [x] LLM integration, Viewer, Editor, Validation
Phase 2: Multi-Page & Docs [x] Page navigation, Help/About Pages
Phase 3: Batch Processing [x] Transcribe/validate all pages automatically
Phase 4: Polish & Release [x] Tests, PAGE-XML Export, UI refinements

Live Demo: dhcraft.org/co-ocr-htr

Design Principles

Principle Meaning
Expert-in-the-Loop Machine assists, human decides
Categorical Confidence confident/uncertain/problematic instead of 0-100%
Constructive UI Helps with work, doesn't get in the way
Workflow Tool Input in, correct output out
Zero Dependencies Runs in browser, no installation

References:

Updated: 2026-02-14