Try the model step by step
- Open the OCR demo. A fictional GTC receipt is already visible; it contains no client records.
- Choose Load model and wait for Ready. This explicitly starts the model and runtime downloads.
- Choose Read text, then compare the recognized text with the original image.
- Try your own permitted English PNG, JPG or WebP image within the stated limits. Copy or download the result after reviewing it.
- Use Reset input to clear the image and result. Use sample receipt restores the fictional example.
Two small models perform different jobs
The selected PaddlePaddle models are PP-OCRv6_tiny_det_onnx and PP-OCRv6_tiny_rec_onnx. The detector locates text regions in the image. The recognizer reads characters from those regions, using a supplied character dictionary. The resulting lines become the recognized text shown beside your image.
The Lab uses a credited adaptation of runonweb’s OCR code with the paddleocr library. ONNX Runtime Web executes the models through WebAssembly in a dedicated worker. These are existing open models; Garrich Torres Consultants built the interface, input checks and operating controls around them. The current tool offers English document processing, with English and Arabic interfaces.
Read the output against the source
Imagine an administrator receives a photographed supplier receipt. OCR can provide text to copy into a review worksheet. The reviewer should compare names, reference numbers, dates, decimal separators and currency marks with the image. A sharp photograph and a simple layout make this comparison easier; faint characters, shadows and crowded columns can create uncertainty.
The displayed confidence score is a model signal. It does not establish that every character or amount is correct. Keep the image available while checking the text, and resolve unclear values from an authorized source. This public tool reads text; it does not identify invoice fields, check totals or authorize a payment.
Use an image that fits the current tool
The demo accepts one PNG, JPG or WebP image of up to 5 MiB and 12 million pixels. It checks the file header before decoding and checks the decoded dimensions again. PDF files and batch folders are outside this input path. For a longer document, prepare an appropriate image separately and review whether this tool fits the task.
An Arabic interface changes the labels and navigation. It does not establish Arabic OCR quality for this build. Use English text for the offered demonstration. A production project involving other languages, handwriting or difficult scans needs its own representative examples and evaluation.
Understand the download and data boundary
The first OCR load uses approximately 6.3 MB of model files and 14.3 MB of runtime files, plus app files. The visible sample receipt is ordinary page content; displaying it does not load an AI model. After you choose Load model, GTC serves the required files and the browser performs recognition on your device.
Your selected image and recognized text stay in this tab’s memory. They are not uploaded, logged or used for model training. Normal hosting request metadata is still generated when the app and model files are requested. There is no cloud inference fallback if local processing fails. Browser and hardware conditions affect whether loading and recognition can finish.
Keep control of device resources and saved files
The Lab permits one model at a time. Supported browsers coordinate this across Lab tabs; the page explains when coordination is limited to the current tab. Models unload after five minutes without activity or one minute in the background. Loading stops after five minutes, and processing stops after two minutes. Stop and Unload clear the current result while keeping the input for retry.
Your browser may cache model files for later use, and may evict them. Clear saved models removes the Lab’s saved model caches after active models are stopped; ordinary HTTP cache remains browser-managed. Cached models do not promise fully offline access. Reset or closing the tool clears its current input and result. Copy and Download create outputs you then manage yourself.
Connect a reviewed result to a useful workflow
An administration team might evaluate OCR for preparing receipt text; a finance team for reviewed capture; a law firm for turning an approved scanned page into searchable working text. These are possible integration scopes, not connected services in this demo. Each needs a clear destination, permitted inputs and a responsible reviewer.
Start with representative, authorized examples and record the corrections people make. Agree which errors prevent acceptance, how exceptions are handled and who may approve a downstream write. Garrich Torres Consultants can scope that path into your business systems. The next step is one measurable document-intake task, with review preserved from the image through to the approved record.