Tesseract Setup

INR recognition requires two Tesseract model files:

eng.traineddata — English script
Devanagari.traineddata — Devanagari script (for Hindi labels common on Indian receipts)

Both come from the Apache 2.0-licensed tessdata_fast project. The package pins their SHA-256 checksums and refuses to run if the files are missing or do not match.

Quick setup (recommended)

Run the setup script from the cloned repository. It downloads, checksums, and installs both models:

./scripts/setup-models.sh

The script places the files under the XDG user data directory:

~/.local/share/payable-receipt-ocr/tessdata/

The package discovers this directory automatically on Linux and macOS.

Custom model directory

If you need to use a different path, supply it explicitly at call time:

result = recognize("receipt.png", tessdata_dir="/opt/tessdata")

Or set the environment variable before the process starts:

export PAYABLE_RECEIPT_OCR_TESSDATA_DIR=/absolute/path/to/tessdata
payable-receipt-ocr receipt.png --currency INR

From the CLI:

payable-receipt-ocr receipt.png --currency INR \
  --tessdata-dir /opt/tessdata

Model discovery order

The package resolves the tessdata directory in this order:

Explicit tessdata_dir argument to recognize() (or --tessdata-dir on the CLI)
PAYABLE_RECEIPT_OCR_TESSDATA_DIR environment variable
~/.local/share/payable-receipt-ocr/tessdata (XDG data home)

Checksum verification

On every recognize() call, both model files are checksummed against the pinned values in runtime-baseline.toml:

Model	SHA-256
`eng.traineddata`	`7d4322bd...170b2`
`Devanagari.traineddata`	`3bbb87c1...c12b7`

A mismatch raises RuntimeBaselineError and no recognition is attempted. This prevents silent accuracy degradation from model version drift.

Runtime environments

Environment	Status
Ubuntu 24.04 / amd64 / tesseract 5.3.4	Conformance target (`baseline_id = linux-noble-amd64-2026-08`)
macOS (Homebrew tesseract)	Supported for development; not the conformance target
Other Linux, ARM, other tesseract versions	Will run in `development` policy; `conformant=false` in result

Latency figures, p95 deadline claims, and accuracy measurements are only valid on the retained conformance runtime. Results on development environments (including macOS) are useful for iteration but do not constitute conformant evaluation.

Troubleshooting

RuntimeBaselineError: model file not found

The model files were not found in any of the three discovery locations. Run ./scripts/setup-models.sh or set PAYABLE_RECEIPT_OCR_TESSDATA_DIR.

RuntimeBaselineError: checksum mismatch

The model files exist but their SHA-256 hashes do not match the pinned values. Re-run ./scripts/setup-models.sh to download the correct versions.

RuntimeBaselineError: Tesseract is not installed or not on PATH

The tesseract executable is not on PATH. Install Tesseract with your package manager and verify with tesseract --version.