MUDIDI¶

About MUDIDI¶

MUDIDI's documentation is published at naarm-nlp.github.io/mudidi from the repository's main branch.

Pipeline¶

MUDIDI digitizes scanned multilingual dictionaries with a two-stage language-model pipeline:

1. Stage 1 transcribes each page faithfully.
2. Stage 2 discovers dictionary structure and emits SIL Toolbox MDF records.

Installation and quickstart¶

Install MUDIDI with uv, configure the API key for your model provider, and then follow the quickstart to run the complete pipeline. Use the YAML configuration guide when you need to override the defaults.

Production inference¶

Use Production Inference to digitize your own dictionary from a directory of page images or a source PDF. The production guide also covers stage selection, agentic retries, and the generated output layout.

Benchmarking and evaluation¶

Use Benchmarking & Evaluation for reproducible dataset experiments, multi-configuration sweeps, and Stage 1 or Stage 2 evaluation. Start with the extraction workflows, then review the evaluation guide. Specialized OCR and VLM backends are documented separately under Advanced VLM backends.

Configuration¶

MUDIDI uses strict, versioned YAML configurations for advanced workflows. See the configuration guide for practical usage and the generated YAML configuration reference for every field and subfield.

Reference¶

Consult the CLI reference for the complete command tree and the Python API reference for maintainer-facing interfaces. Contributors can continue with the architecture overview and CLI migration guide.