Extract-HTML Skill

A robust skill for converting HTML documents into strictly valid JSON based on a user-provided JSON Schema.

Capabilities

1. Schema Compliance: Guarantees output conforms to the provided JSON Schema (using Schematron-3B + validation loop).
2. Deterministic Tables: Extracts HTML tables using pandas.read_html and injects them as context, preventing hallucination of data.
3. Media Text Extraction: Identifies images, filters by pixel size, and optionally uses a Vision API (OpenAI-compatible) to extract text/OCR.
4. Self-Correction: Validates model output and retries with error feedback if schema validation fails.

Usage

Basic Conversion (Local Only)

./run.sh convert \
  --html input.html \
  --schema target.schema.json \
  --out result.json

Advanced (With Vision & Remote Fetch)

./run.sh convert \
  --html input.html \
  --schema target.schema.json \
  --out result.json \
  --fetch-remote-media \
  --vision-api-base "https://glhf.chat/api/openai/v1" \
  --vision-api-key "sk-..." \
  --vision-model "gpt-4o-mini"

Options

• --max-attempts <int>: Number of self-correction retries.
• --extract-tables / --no-extract-tables: Toggle deterministic table extraction.
• --extract-media-text: Enable image processing.
• --min-image-px, --max-image-px: Filter images by size.
• --include-sections: detailed H1-H6 hierarchy in context.

Dependencies

• Ollama running schematron-3b (or compatible model).
• Python 3.11+
• See pyproject.toml for python deps.