OCR Accuracy and Error Handling

OCR is never 100% accurate. Misread characters (e.g., 'O' vs '0', 'l' vs '1') can break your embeddings and search quality. A production-grade RAG system needs strategies to handle these errors.

Measuring Accuracy (WER and CER)

• Word Error Rate (WER): Measures the ratio of word-level errors.
• Character Error Rate (CER): Measures character-level changes (insertions, deletions, substitutions).

import Levenshtein

def calculate_cer(reference, hypothesis):
    distance = Levenshtein.distance(reference, hypothesis)
    return distance / len(reference)

Common OCR Errors

• Garbage Text: Random characters like ~ or | appearing near lines or edges.
• Merging Words: "TheNew" instead of "The New".
• Broken Formatting: Multi-column text being read line-by-line across columns.

Post-Processing OCR Text

Use NLP models or spell-checkers to clean the output.

from textblob import TextBlob

def spell_check_ocr(text):
    b = TextBlob(text)
    return str(b.correct())

LLM-Based Correction

Since we are building a RAG system, we can use an LLM (like Claude) to "fix" the OCR noise:

prompt = f"Fix the common OCR errors in the following text while preserving the original meaning: {noisy_text}"

Building a Resilient Pipeline

1. Thresholding: Reject chunks with an average_ocr_confidence below a certain level (e.g., 0.70).
2. Double-Indexing: Index the "raw" OCR text and the "cleaned" version.
3. Fuzzy Search: Use vector embeddings that are robust to small spelling variations.

Error Handling Scenarios

Scenario	Strategy
Hand-written notes	Use a high-quality Vision model (e.g. Claude)
Low-resolution scan	Upscale and use OCR_DPI optimization
Mixed languages	Use OCR engines with multi-language support (Tesseract --lang)

Exercises

1. Intentionally smudge a piece of paper, scan it, and run OCR.
2. Identify 3 consistent errors it makes.
3. Write a small script to regex-replace or "clean" those specific errors.