OCR
Extract text from images using TrOCR, GLM-OCR, and LightOnOCR-2 models.
Extract text from images using OCR models — from lightweight line-level TrOCR to document-level vision-language models like GLM-OCR and LightOnOCR-2. Works with handwritten text, printed documents, tables, formulas, and screenshots.
For full API reference (extractText(), extractTextMany(), options, result types, and custom providers), see the Core OCR guide.
See it in action
Try the Knowledge Base block — its ingest pipeline runs OCR on scanned documents before embedding and search.
Recommended Models
| Model | Size | Type | Strength | Use Case |
|---|---|---|---|---|
Xenova/trocr-small-printed | ~120MB | Line-level | Printed text | Quick extraction, screenshots (recommended for getting started) |
Xenova/trocr-small-handwritten | ~120MB | Line-level | Handwritten text | Notes, forms, handwriting |
onnx-community/GLM-OCR-ONNX | ~652MB | Document-level | Tables, formulas, structured data | Complex documents, invoices, receipts |
onnx-community/LightOnOCR-2-1B-ONNX | ~700MB | Document-level | Speed, 11 languages | High-throughput document scanning |
Line-level vs Document-level
TrOCR models process one line of text at a time (~120MB, fast). GLM-OCR and LightOnOCR-2 are vision-language models that process full document images (~650-700MB, WebGPU recommended) with support for tables, formulas, and structured extraction via prompt-based modes.
Document Scanner Example
Based on the Knowledge Base block:
import { transformers } from '@localmode/transformers';
import { extractText } from '@localmode/core';
// Line-level OCR (TrOCR)
const model = transformers.ocr('Xenova/trocr-small-printed');
const { text } = await extractText({
model,
image: dataUrl,
});Generative OCR (GLM-OCR / LightOnOCR-2)
// Document-level OCR with prompt-based modes
const model = transformers.ocr('onnx-community/GLM-OCR-ONNX');
// Text extraction (default)
const { text } = await extractText({ model, image: documentImage });
// Table recognition
const { text: table } = await extractText({
model,
image: tableImage,
prompt: 'Table Recognition:',
});
// Formula recognition
const { text: formula } = await extractText({
model,
image: mathImage,
prompt: 'Formula Recognition:',
});Image Input Formats
The image parameter accepts:
string— Data URL (data:image/jpeg;base64,...)Blob— Image blob from file input
Best Practices
OCR Tips
- Choose the right model — TrOCR for quick single-line extraction, GLM-OCR/LightOnOCR-2 for full documents
- Image quality matters — Clear, well-lit images give much better results
- Use prompt modes — Generative models support
'Text Recognition:','Table Recognition:', and'Formula Recognition:'prompts - WebGPU for generative models — GLM-OCR and LightOnOCR-2 run significantly faster with WebGPU; WASM works but is slower
For full PDF text extraction, use @localmode/pdfjs instead — it extracts text directly from PDF
structure without OCR. Use OCR only for scanned documents or images of text.
Composed Block
| Block | Description | Links |
|---|---|---|
| Semantic Search | Text/PDF/OCR ingest with chunking, then reranked vector search with category facets and calibrated thresholds | Live · Install: npx shadcn add @localmode/ui/blocks/knowledge/semantic-search |