hermes-config/skills/.archive/ocr-data-extraction/SKILL.md

name, description, version, author, license, metadata

name	description	version	author	license	metadata
ocr-data-extraction	Extract structured data (names, phones, VINs, dates) from OCR'd text using rule-based regex parsing — no AI/API needed.	1.0.0	Hermes Agent	MIT	hermes tags related_skills OCR parsing regex data-extraction screenshots browser Tesseract ocr-and-documents

OCR → Structured Data Extraction (Rule-Based)

Extract structured fields from OCR'd text using pure regex — no API keys, no network calls, no AI. Runs entirely on CPU, either in-browser (Tesseract.js CDN) or server-side (Tesseract CLI).

When to use this approach vs AI/API:

• Data follows predictable patterns (phones, VINs, dates, names)
• Latency matters (instant vs 1-3s API call)
• Cost matters (free vs per-call billing)
• Privacy matters (data stays local)
• User explicitly prefers CPU-based — honor this signal immediately

Quick Start (Browser)

<script src="https://cdn.jsdelivr.net/npm/tesseract.js@5/dist/tesseract.min.js"></script>

Then call Tesseract.recognize(file, 'eng', { logger }) to get text, then run the parser.

Parser Architecture

The parser uses layered extraction in this order:

1. Block Splitting

Split OCR text on blank lines first. If that fails, detect table-like structures (consistent word counts across lines).

2. Structured Field Extraction (per block)

Extract and REMOVE these from the text first (order matters — early extraction simplifies later parsing):

Phone:  /\(?\d{3}\)?[\s.\-]*\d{3}[\s.\-]*\d{4}/
VIN:    /\b[A-HJ-NPR-Z0-9OIQ]{17}\b/i   ← lenient: accepts O→0, I→1, Q→0
Date:   YYYY-MM-DD, MM/DD/YYYY, "Jan 15, 2024"
Time:   9:00 AM, 1:30PM, 14:00, 8am
Duration: /\b(\d+)\s*(?:min|minutes?|hrs?|hours?)\b/i

OCR VIN tolerance: Tesseract commonly confuses 0→O, 1→I. Accept O/I/Q in the regex, then normalize:

function fixVin(v) { return v.toUpperCase().replace(/O/g,'0').replace(/I/g,'1').replace(/Q/g,'0'); }

3. Remaining Text → Name / Vehicle / Service

After removing structured fields, split remaining text on multi-spaces (preserves column structure). Then:

• Name: first part, or consecutive capitalized words without digits
• Vehicle: part containing year pattern (19xx/20xx) or known makes (ford|toyota|honda|bmw|...)
• Service: everything else

4. Advisor + RO Code Stripping

If the source has AdvisorName [RO_CODE] prefixes, strip them:

var m = part.match(/^([A-Z][a-z]+)\s+(\[[^\]]+\])\s+(.*)/);
if (m) { roCode = m[2]; return m[3]; }  // return service, save RO as note

5. Multi-Appointment Boundary Detection

When multiple appointments appear without blank-line separation, detect boundaries by:

• Find the next phone/VIN in the remaining text
• Walk back to the nearest multi-space gap or capitalized name before it
• Split there — text before goes to current appointment, text after recurses
• Regex for gap walk-back: /\s{2,}(?=\S+(?:\s+\S+){0,1}\s*$)/
• Fallback (no gap): detect last 1-2 capitalized words before phone: /([A-Z][A-Za-z]+(?:\s+\S+){0,1})\s*$/

6. Header/Noise Stripping

• Separator lines: /^[\-–=_*#]{3,}$/ → remove
• Short all-caps headers: < 25 chars, all uppercase letters/spaces → remove
• Table header row: contains name|phone|date|time|service|vehicle|vin → skip first line
• Known header words: schedule|appointments|roster|calendar|upcoming → filter from parts

Pitfalls

• Don't collapse whitespace before splitting: Split on \s{2,} FIRST, then clean each part. If you collapse to single spaces first, multi-space splitting silently breaks.
• Trailing space before boundary breaks regex: Phone/VIN markers are often preceded by a space in the OCR. Trim or use \s*$ in lookahead regexes.
• "O" in VIN ≠ letter O: Always use lenient VIN regex and normalize. OCR will turn zeros into O's.
• Advisor names vs customer names: In shop management systems, "Word [CODE]" is advisor+RO, not customer. Strip it.
• Block recursion can re-include old text: When recursing, pass only the text AFTER the boundary split point, not the full remaining b variable.
• Table detection false positives: Only use line-by-line table parsing when lines have consistent word counts (≤3 word difference from average). Exclude header line before checking consistency.

Reference File

See references/parser-patterns.md for the full extraction regex catalog and test cases.