coOCR/HTR
Promptotyping Knowledge Vault
Design decisions and project documentation for coOCR/HTR
Design Principles
Critical Expert in the Loop
The AI assists; the human decides. LLMs hallucinate and exhibit sycophancy - expert review with domain knowledge and LLM literacy catches what automation misses.
Hybrid Validation
Deterministic rules catch structural issues (dates, currency, empty cells). LLM Review evaluates semantic plausibility (historical spelling, context). Results categorized as confident / uncertain / problematic.
Workflow-Agnostic
Fits into existing transcription pipelines. Import PAGE-XML, export to multiple formats (TXT, JSON, PAGE-XML, TEI-XML). Use as standalone tool or processing step.
Open Browser Tool
No backend, no installation, no account required. Runs entirely in the browser. Open source (CC BY 4.0), vanilla JavaScript, fully customizable.
Cloud & Local Models
Use cloud LLMs (Gemini, OpenAI, Anthropic) for best quality or local models via Ollama for data privacy. Switch providers anytime - same workflow, your choice of trade-offs.
Development Methodology
Promptotyping
A four-phase iterative context engineering methodology for developing research artifacts with frontier LLMs. Built with agentic coding using Claude Code powered by Claude Opus 4.5. This knowledge vault documents the entire design process. Read more on L.I.S.A. (Gerda Henkel Stiftung)
Documents
Loading document...