Super Extract

Structured information extraction with character-level source grounding. Extract entities from PDFs, DOCX, and text with exact position mapping and interactive visualization.

Inspired by google/langextract.

What It Does

Super Extract is an AI coding agent skill that:

1. Reads documents (PDF, DOCX, plain text) via companion pdf/docx skills
2. Extracts structured entities using LLM few-shot prompting
3. Aligns extractions to exact character positions in the source text using token-based alignment (difflib)
4. Visualizes results as highlighted interactive HTML

Skill Structure

skill/
  SKILL.md        — Agent instructions (extraction pipeline, visualization template)
  bin/
    align.py      — Tokenizer + WordAligner (ported from google/langextract)

align.py

The alignment engine is a standalone Python script (~380 lines, zero dependencies beyond stdlib) that:

• Tokenizes text via RegexTokenizer (spaced languages) or UnicodeTokenizer (CJK, Thai, etc.)
• Aligns extraction strings to source text using WordAligner (difflib SequenceMatcher)
• Supports exact and fuzzy matching with configurable threshold
• Outputs character intervals (start_pos, end_pos) for each extraction

Usage

# Pipe JSON
echo '{"source":"Patient takes Aspirin 100mg daily.","extractions":[{"extraction_class":"medication","extraction_text":"Aspirin"}]}' \
  | python3 skill/bin/align.py --stdin

# From files
python3 skill/bin/align.py --source doc.txt --extractions entities.json

# CJK text
python3 skill/bin/align.py --stdin --unicode < input.json

Output

{
  "source": "Patient takes Aspirin 100mg daily.",
  "extractions": [
    {
      "extraction_class": "medication",
      "extraction_text": "Aspirin",
      "char_interval": {"start_pos": 14, "end_pos": 21},
      "alignment_status": "match_exact",
      "token_interval": {"start_index": 2, "end_index": 3}
    }
  ]
}

How It Differs from google/langextract

google/langextract is a full Python library that bundles its own LLM inference, provider system, and batch processing. Super Extract strips that down to just the alignment engine — because the AI agent is the LLM. The agent handles prompting, chunking, and visualization natively.

Component	google/langextract	super-extract
LLM inference	Built-in (Gemini, OpenAI, Ollama)	Agent is the LLM
Prompt template	Q&A format in code	SKILL.md instructions
Tokenizer	RegexTokenizer + UnicodeTokenizer	Same (ported)
Alignment	WordAligner (difflib)	Same (ported)
Visualization	Jupyter HTML widget	Agent-generated HTML
Batch processing	ThreadPoolExecutor	Agent parallelism

License

Apache 2.0 — alignment code ported from google/langextract (Apache 2.0).