Skip to content

Instantly share code, notes, and snippets.

@snowfluke

snowfluke/index.ts

Last active March 8, 2026 03:16
Show Gist options

  • Select an option

    Learn more about clone URLs
  • Save snowfluke/6faa92771ceaf41945c1821829b0a301 to your computer and use it in GitHub Desktop.

Select an option

Learn more about clone URLs
Save snowfluke/6faa92771ceaf41945c1821829b0a301 to your computer and use it in GitHub Desktop.
Download ZIP
Convert flatten/scanned pdf into searchable pdf
Raw
index.ts
// curl -fsSL https://bun.sh/install | bash
// bun add ppu-pdf ppu-paddle-ocr onnxruntime-node
// Run it: bun run index.ts
import { PaddleOcrService } from "ppu-paddle-ocr";
import { PdfReader } from "ppu-pdf";
export const MODEL_BASE_URL =
"https://media.githubusercontent.com/media/PT-Perkasa-Pilar-Utama/ppu-paddle-ocr-models/main";
export const DICT_BASE_URL =
"https://raw.githubusercontent.com/PT-Perkasa-Pilar-Utama/ppu-paddle-ocr-models/main";
const pdfReader = new PdfReader({ verbose: false });
// Tweak the model variant and dictionary to balance the accuracy and performance.
// Note that the dictionary should match the recognition model, otherwise the OCR results will be inaccurate.
const ocr = new PaddleOcrService({
model: {
detection: `${MODEL_BASE_URL}/detection/PP-OCRv5_mobile_det_infer.onnx`,
recognition: `${MODEL_BASE_URL}/recognition/PP-OCRv5_mobile_rec_infer.onnx`,
charactersDictionary: `${DICT_BASE_URL}/recognition/ppocrv5_dict.txt`,
},
});
await ocr.initialize();
// Download OCR model and warm up cache
console.log("Warming up OCR model...");
{
const testBuffer = await Bun.file("./assets/opposite-expectation-scan.pdf").arrayBuffer();
const testDoc = pdfReader.open(testBuffer);
const testCanvas = await pdfReader.renderAll(testDoc);
await pdfReader.getTextsScanned(ocr, testCanvas);
pdfReader.destroy(testDoc);
}
console.log("Warmup complete.\n");
console.time("Normal inference")
{
// 1. Reading the file from disk
const fileScan = Bun.file("./assets/test_japanese.pdf");
const bufferScan = await fileScan.arrayBuffer();
// 2. Open and Render
const pdfScan = pdfReader.open(bufferScan);
const canvasMap = await pdfReader.renderAll(pdfScan);
pdfReader.destroy(pdfScan);
// 3. Extract OCR Texts
const texts = await pdfReader.getTextsScanned(ocr, canvasMap);
// 4. Rebuild Searchable PDF
const pdfForRebuild = pdfReader.open(bufferScan);
const rebuiltPdfBuffer = await pdfReader.rebuild(pdfForRebuild, texts);
pdfReader.destroy(pdfForRebuild);
// 5. Save onto disk
await Bun.write("./test_japanese_searchable.pdf", rebuiltPdfBuffer);
}
console.timeEnd("Normal inference")
// import { bench, group, run } from "mitata";
// console.log("\nStarting benchmarking")
// group("ppu-pdf e2e processing", () => {
// bench("Extract Texts and Rebuild PDF", async () => {
// const fileScan = Bun.file("./assets/test_japanese.pdf");
// const bufferScan = await fileScan.arrayBuffer();
// const pdfScan = pdfReader.open(bufferScan);
// const canvasMap = await pdfReader.renderAll(pdfScan);
// pdfReader.destroy(pdfScan);
// const texts = await pdfReader.getTextsScanned(ocr, canvasMap);
// const pdfForRebuild = pdfReader.open(bufferScan);
// const rebuiltPdfBuffer = await pdfReader.rebuild(pdfForRebuild, texts);
// pdfReader.destroy(pdfForRebuild);
// await Bun.write("./test_japanese_searchable.pdf", rebuiltPdfBuffer);
// });
// });
// await run({
// colors: true,
// });
await ocr.destroy();
// BENCHMARK RESULT
// benchmark avg (min … max) p75 / p99 (min … top 1%)
// -------------------------------------------- -------------------------------
// • ppu-pdf e2e processing
// -------------------------------------------- -------------------------------
// japan_PP-OCRv3_mobile_rec_infer.onnx + japan_dict.txt
// Extract Texts and Rebuild PDF 798.30 ms/iter 799.05 ms █ █
// (783.87 ms … 850.33 ms) 817.52 ms █ █
// (224.00 kb … 18.47 mb) 9.74 mb █▁█▁▁▁█▁██▁▁▁▁▁█▁▁▁▁█
// PP-OCRv5_mobile_rec_infer.onnx + ppocrv5_dict.txt
// Extract Texts and Rebuild PDF 802.18 ms/iter 803.59 ms █ █ █
// (792.74 ms … 825.94 ms) 817.62 ms █ █▅ █▅ ▅ ▅ ▅
// ( 16.00 kb … 15.58 mb) 7.87 mb █▁██▁██▁▁█▁▁▁█▁▁▁▁▁▁█
// PP-OCRv5_server_rec_infer.onnx + ppocrv5_dict.txt
// Extract Texts and Rebuild PDF 802.84 ms/iter 804.37 ms █
// (797.71 ms … 819.77 ms) 808.87 ms ▅█▅▅▅▅▅ ▅ ▅ ▅
// (384.00 kb … 33.72 mb) 11.90 mb ███████▁▁▁▁▁█▁▁▁█▁▁▁█
@snowfluke
Copy link
Author

snowfluke commented Mar 7, 2026

OCR Benchmark Report

Results

Model Avg p75 p99 Avg Mem Max Mem
japan_PP-OCRv3_mobile 798.30 ms 799.05 ms 817.52 ms 9.74 mb 18.47 mb
PP-OCRv5_mobile 802.18 ms 803.59 ms 817.62 ms 7.87 mb 15.58 mb
PP-OCRv5_server 802.84 ms 804.37 ms 808.87 ms 11.90 mb 33.72 mb

Analysis

Latency

All three models complete in ~800ms for the full document with no meaningful difference. The gap between fastest and slowest is only 4.54ms — within measurement noise.

Memory

PP-OCRv5_mobile is the most efficient at 7.87mb average. PP-OCRv5_server consumes the most at 11.90mb average with a spike to 33.72mb, reflecting larger model weights loaded at runtime. japan_PP-OCRv3_mobile shows the widest swing (224kb → 18.47mb), suggesting inconsistent allocation, possibly from lazy model initialization on the first iteration.

Latency Consistency

PP-OCRv5_server has the tightest spread (797–819ms, p99 808ms), making it the most predictable under load. PP-OCRv5_mobile is close behind. japan_PP-OCRv3_mobile has the widest range (783–850ms), with occasional spikes visible in the histogram.

Recommendation

Use PP-OCRv5_mobile for production.

  • Lowest memory footprint (7.87mb avg, 15.58mb max) — important for WebAssembly where heap is constrained
  • Speed is on par with the other models (~800ms)
  • Better latency consistency than PP-OCRv3_mobile
  • No overhead cost of server model weights unlike PP-OCRv5_server

PP-OCRv5_server is worth considering only if accuracy on complex Japanese layouts proves insufficient with the mobile model — the memory tradeoff is significant in a browser context.

@snowfluke
Copy link
Author

snowfluke commented Mar 7, 2026

Hypotetichal 180-page in Bun.js run time

Model Est. Duration Est. Avg Mem Est. Max Mem
japan_PP-OCRv3_mobile ~14.4 s ~174 mb ~332 mb
PP-OCRv5_mobile ~14.4 s ~142 mb ~280 mb
PP-OCRv5_server ~14.5 s ~214 mb ~607 mb

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment