Skip to main content
DoctorDocs
Public Accuracy Report · October 2026

OCR Accuracy Benchmarks & Testing Methodology

We believe in complete transparency. We publish our empirical test results across different document categories, failure rates, testing protocols, and known limitations so users and developers know exactly what to expect.

98.4%

Printed Text

High-contrast PDFs & books

88.2%

Neat Handwriting

Student lecture notes & diaries

73.6%

Doctor Slips (Raw)

86.1% with clinical entity linking

97.1%

Indian ID Cards

Aadhaar, PAN & Passport MRZ

Medical OCR Cautionary Notice (YMYL):Doctor prescriptions and clinical notes have high variance in handwriting legibility. While our AI OCR identifies medicine names with up to 86.1% contextual accuracy, misreading a single dosage digit (e.g., 5mg vs 50mg) can cause catastrophic harm. DoctorDocs is strictly an informational and educational aid—never dispense, administer, or modify medication based solely on OCR outputs. Always inspect the original slip and consult a licensed physician or pharmacist.

1. Benchmark Results by Document Category

Evaluated against an empirical test set of 1,950 real-world documents collected between May 2026 and October 2026. Accuracy is measured using industry-standard Character Error Rate (CER) and Word Error Rate (WER) against double-blinded human transcriptions.

Understanding Metric Scope: Character Accuracy vs. Whole-Document AccuracyA 98.4% character accuracy (CER = 1.6%) does not mean 98.4% of entire documents are completely error-free. On a standard 500-word single-page document (~3,000 characters), a 1.6% CER corresponds to approximately 48 misrecognized characters across the page. Word-level and character-level metrics reflect individual token fidelity, which is why human verification remains essential for all sensitive legal contracts and medical prescriptions.
Document CategorySample SizeChar Accuracy (1-CER)Word Accuracy (1-WER)Primary Failure Mode
Clean Printed English
Books, invoices, digital PDFs, legal contracts		500 docs98.4%96.8%Obscure font serifs, tiny punctuation dots
Bilingual Printed Documents (Hindi + English)
Government gazettes, Indian university mark sheets		300 docs94.7%91.2%Devanagari conjunct consonants (halant, matras)
Student Lecture Notes (Neat Handwriting)
Ruled paper, ballpoint/gel pens, clear lighting		250 pages88.2%83.5%Connected cursive loops (e.g. 'u' vs 'v', 'cl' vs 'd')
Indian Medical Prescriptions (OPD Slips)
Cursive shorthand, doctor abbreviations (BD, TDS, SOS)		350 slips73.6% (Raw)
86.1% w/ MedDict		68.2% (Raw)
82.4% w/ MedDict		Extreme abbreviation, overlapping scribbles, brand typos
Faded Paper & Carbon Copies
Thermal billing receipts, faded hospital discharge notes		150 docs64.2% (Raw)
82.5% w/ Enhancer		59.8% (Raw)
78.1% w/ Enhancer		Faded ink contrast drop, background paper grain noise
Indian ID Cards & Passports
Aadhaar QR/Masking, PAN Card, Passport MRZ Lines		400 cards97.1%95.4% (99.8% MRZ)Hologram glare, laminated plastic reflection

2. Engine Architecture Comparison: Local WASM vs Cloud AI

DoctorDocs utilizes a dual-engine architecture to balance absolute privacy against recognition capabilities for challenging handwriting:

Local Engine (Tesseract.js WASM)

  • Privacy: 100% In-Browser. 0 bytes transferred over the internet.
  • Processing Speed: 800ms – 2.5s per standard page.
  • Offline Support: Works fully offline once cached.
  • Best For: Clean printed books, bank statements, high-contrast receipts.
  • Limitations: Struggles with connected cursive handwriting and faded ink.

Cloud AI Engine (Gemini 1.5 Flash)

  • Privacy: Encrypted in transit (TLS 1.3). Zero retention for model training.
  • Processing Speed: 1.5s – 3.2s per page.
  • Offline Support: Requires active internet connection.
  • Best For: Difficult doctor prescriptions, complex cursive notes, lab reports.
  • Limitations: Governed by fair-use API quotas (1 free scan per session; unlimited on Pro).

3. Documented Failure Modes & Edge Cases

1. Severe Motion Blur & Angled Distortion

When images are captured at steep angles (>35 degrees) or with camera shake, character edge detection drops by up to 28%. We strongly recommend using our interactive document scanner camera tool, which detects document borders before snapping.

2. Carbon Paper Ink Bleed-Through

Carbon copies from clinic registers frequently have ink bleed from the reverse side. The local engine may interpret ghost shadows as text. Toggling our built-in Magic Enhancer boosts contrast thresholding and reduces false positive bleed artifacts.

3. Non-Standard Regional Medicine Brands

India has over 100,000 regional branded generic formulations. When a physician writes an uncommon local brand phonetically with cursive abbreviation, OCR cannot guarantee accurate transcription without human verification.

4. Methodology & Reproducibility

Our accuracy evaluation pipeline is open and verifiable:

  1. Ground Truth Formulation: Each test document is independently transcribed by two human operators. Discrepancies are reconciled by a third reviewer.
  2. Normalization: Whitespace, unicode quotes, and case variations are normalized before running Levenshtein edit distance calculations.
  3. CER & WER Metrics:
    CER = (Substitutions + Deletions + Insertions) / Total Reference Characters
    Accuracy = 1 - CER (expressed as a percentage)
  4. Hardware Consistency: Local WASM tests were executed on Chromium V8 (MacBook Air M2, 16GB RAM and Moto G Power Android phone). Cloud AI tests were routed through our stateless edge inference pipeline.

Help Us Improve Our Benchmarks

Found a document that failed recognition or produced an incorrect transcription? Share your feedback to help us refine our preprocessing and model cascades.

Email Benchmark Feedback Test OCR Scanner Live
OCR Accuracy Benchmarks & Testing Methodology — DoctorDocs | DoctorDocs