A fast, helpful, and open-source document parser
fix(wasm): run browser OCR without Tokio
Choudhry Abdullahdf41d4479ee2
diff --git a/crates/liteparse/src/ocr_merge.rs b/crates/liteparse/src/ocr_merge.rsindex ba6adcb..2237c80 100644--- a/crates/liteparse/src/ocr_merge.rs+++ b/crates/liteparse/src/ocr_merge.rs@@ -544,6 +544,31 @@ pub(crate) async fn ocr_and_merge_rendered(num_workers: usize,ocr_failure_fatal: bool,) -> Result<(), LiteParseError> {+ type OcrTaskResult = Result<Vec<OcrResult>, Box<dyn std::error::Error + Send + Sync>>;++ // Browser WASM uses the JavaScript event loop. It has no Tokio runtime or+ // blocking thread pool. Run each JavaScript OCR callback directly so the+ // returned Promise can make progress on the browser event loop.+ #[cfg(target_arch = "wasm32")]+ let task_results: Vec<(usize, usize, f32, OcrTaskResult)> = {+ let _ = num_workers;+ let mut results = Vec::with_capacity(rendered.len());+ for r in rendered {+ let idx = r.idx;+ let page_number = pages[idx].page_number;+ let page_dpi = r.dpi;+ let options = OcrOptions {+ language: ocr_language.to_string(),+ dpi: page_dpi,+ };+ let result = ocr_engine+ .recognize(&r.pixels, r.width, r.height, &options)+ .await;+ results.push((idx, page_number, page_dpi, result));+ }+ results+ };+// Phase 1: spawn one async task per page. A semaphore limits how many run// `recognize` concurrently to `num_workers`.//@@ -557,48 +582,63 @@ pub(crate) async fn ocr_and_merge_rendered(// request never goes out, the permit is never released, and the whole OCR// pass deadlocks. Acquiring the permit asynchronously parks the lightweight// task instead, so only `num_workers` blocking threads are ever consumed.- let num_workers = num_workers.max(1);- let semaphore = Arc::new(tokio::sync::Semaphore::new(num_workers));- let mut handles = Vec::with_capacity(rendered.len());-- let handle = tokio::runtime::Handle::current();-- for r in rendered {- let engine = ocr_engine.clone();- let sem = semaphore.clone();- let language = ocr_language.to_string();- let page_number = pages[r.idx].page_number;- let rt_handle = handle.clone();-- handles.push((- r.idx,- page_number,- r.dpi,- tokio::spawn(async move {- // Park the task (not an OS thread) until a permit is available.- let _permit = sem.acquire_owned().await.expect("semaphore closed");- let options = OcrOptions {- language,- dpi: r.dpi,- };- // Offload the (possibly CPU-blocking, e.g. Tesseract) recognize- // onto a blocking thread. Because the permit is already held,- // at most `num_workers` blocking threads are in use at once,- // leaving the rest of the pool free for the HTTP client's- // internal DNS resolution.- match tokio::task::spawn_blocking(move || {- rt_handle.block_on(engine.recognize(&r.pixels, r.width, r.height, &options))- })- .await- {- Ok(result) => result,- Err(join_err) => {- Err(Box::new(join_err) as Box<dyn std::error::Error + Send + Sync>)+ #[cfg(not(target_arch = "wasm32"))]+ let task_results: Vec<(usize, usize, f32, OcrTaskResult)> = {+ let num_workers = num_workers.max(1);+ let semaphore = Arc::new(tokio::sync::Semaphore::new(num_workers));+ let mut handles = Vec::with_capacity(rendered.len());++ let handle = tokio::runtime::Handle::current();++ for r in rendered {+ let engine = ocr_engine.clone();+ let sem = semaphore.clone();+ let language = ocr_language.to_string();+ let page_number = pages[r.idx].page_number;+ let rt_handle = handle.clone();++ handles.push((+ r.idx,+ page_number,+ r.dpi,+ tokio::spawn(async move {+ // Park the task (not an OS thread) until a permit is available.+ let _permit = sem.acquire_owned().await.expect("semaphore closed");+ let options = OcrOptions {+ language,+ dpi: r.dpi,+ };+ // Offload the (possibly CPU-blocking, e.g. Tesseract) recognize+ // onto a blocking thread. Because the permit is already held,+ // at most `num_workers` blocking threads are in use at once,+ // leaving the rest of the pool free for the HTTP client's+ // internal DNS resolution.+ match tokio::task::spawn_blocking(move || {+ rt_handle.block_on(engine.recognize(&r.pixels, r.width, r.height, &options))+ })+ .await+ {+ Ok(result) => result,+ Err(join_err) => {+ Err(Box::new(join_err) as Box<dyn std::error::Error + Send + Sync>)+ }}+ }),+ ));+ }++ let mut results = Vec::with_capacity(handles.len());+ for (idx, page_number, page_dpi, handle) in handles {+ let result = match handle.await {+ Ok(result) => result,+ Err(join_err) => {+ Err(Box::new(join_err) as Box<dyn std::error::Error + Send + Sync>)}- }),- ));- }+ };+ results.push((idx, page_number, page_dpi, result));+ }+ results+ };// Phase 3: collect results and merge into pages.@@ -614,32 +654,20 @@ pub(crate) async fn ocr_and_merge_rendered(// enrichment but already has all its text. We must only fail loud when OCR// failure destroyed a sparse page's likely primary text source — otherwise// a broken OCR setup would abort perfectly good native-text documents.- let total_tasks = handles.len();+ let total_tasks = task_results.len();let mut failed_tasks = 0usize;let mut failed_sparse_text_page = false;let mut first_error: Option<String> = None;- for (idx, page_number, page_dpi, handle) in handles {- let ocr_results: Vec<OcrResult> = match handle.await {- Ok(Ok(results)) => results,- Ok(Err(e)) => {- failed_tasks += 1;- failed_sparse_text_page |= page_has_sparse_native_text(&pages[idx]);- // Only log the first failure to avoid flooding stderr with an- // identical message for every page.- if first_error.is_none() {- let msg = e.to_string();- eprintln!("[ocr] failed for page {}: {}", page_number, msg);- first_error = Some(msg);- }- continue;- }+ for (idx, page_number, page_dpi, result) in task_results {+ let ocr_results: Vec<OcrResult> = match result {+ Ok(results) => results,Err(e) => {failed_tasks += 1;failed_sparse_text_page |= page_has_sparse_native_text(&pages[idx]);if first_error.is_none() {let msg = e.to_string();- eprintln!("[ocr] task panicked for page {}: {}", page_number, msg);+ eprintln!("[ocr] failed for page {}: {}", page_number, msg);first_error = Some(msg);}continue;diff --git a/scripts/browser-compat/wasm-test.html b/scripts/browser-compat/wasm-test.htmlindex 88445f5..e16137d 100644--- a/scripts/browser-compat/wasm-test.html+++ b/scripts/browser-compat/wasm-test.html@@ -3,6 +3,7 @@<head><meta charset="UTF-8" /><title>LiteParse WASM Browser Test</title>+ <link rel="icon" href="data:," /></head><body><div id="status">loading</div>@@ -14,6 +15,32 @@const resultEl = document.getElementById("result");const errorEl = document.getElementById("error");+ function buildBlankPdf() {+ const objects = [+ "<< /Type /Catalog /Pages 2 0 R >>",+ "<< /Type /Pages /Kids [3 0 R] /Count 1 >>",+ "<< /Type /Page /Parent 2 0 R /MediaBox [0 0 72 72] /Resources << >> /Contents 4 0 R >>",+ "<< /Length 0 >>\nstream\n\nendstream",+ ];++ let pdf = "%PDF-1.4\n%1234\n";+ const offsets = [0];+ for (let index = 0; index < objects.length; index += 1) {+ offsets.push(pdf.length);+ pdf += `${index + 1} 0 obj\n${objects[index]}\nendobj\n`;+ }++ const xrefOffset = pdf.length;+ pdf += `xref\n0 ${objects.length + 1}\n`;+ pdf += "0000000000 65535 f \n";+ for (const offset of offsets.slice(1)) {+ pdf += `${String(offset).padStart(10, "0")} 00000 n \n`;+ }+ pdf += `trailer\n<< /Size ${objects.length + 1} /Root 1 0 R >>\n`;+ pdf += `startxref\n${xrefOffset}\n%%EOF\n`;+ return new TextEncoder().encode(pdf);+ }+try {status.textContent = "loading wasm";const { default: init, LiteParse } = await import("/packages/wasm/pkg/liteparse_wasm.js");@@ -37,11 +64,56 @@if (pageCount === 0) throw new Error("No pages parsed");if (!hasText) throw new Error(`Text too short: ${textLength} chars`);+ status.textContent = "testing OCR callback";+ let ocrCallbackCalls = 0;+ let ocrByteLength = 0;+ const ocrParser = new LiteParse({+ ocrEnabled: true,+ ocrLanguage: "eng",+ outputFormat: "text",+ dpi: 72,+ quiet: true,+ ocrEngine: {+ async recognize(imageData, width, height, language) {+ ocrCallbackCalls += 1;+ if (!(imageData instanceof Uint8Array)) {+ throw new Error("OCR input is not a Uint8Array");+ }+ if (imageData.length === 0) {+ throw new Error("OCR input is empty");+ }+ if (width <= 0 || height <= 0) {+ throw new Error(`Invalid OCR dimensions: ${width}x${height}`);+ }+ if (language !== "eng") {+ throw new Error(`Unexpected OCR language: ${language}`);+ }++ ocrByteLength = imageData.length;+ return [{+ text: "WASM_OCR_CALLBACK_OK",+ bbox: [4, 4, Math.max(8, width - 4), Math.min(24, height - 4)],+ confidence: 0.99,+ }];+ },+ },+ });+ const ocrResult = await ocrParser.parse(buildBlankPdf());++ if (ocrCallbackCalls !== 1) {+ throw new Error(`Expected one OCR callback, received ${ocrCallbackCalls}`);+ }+ if (!ocrResult.text.includes("WASM_OCR_CALLBACK_OK")) {+ throw new Error("OCR callback text was not merged into the parse result");+ }+status.textContent = "done";resultEl.style.display = "block";resultEl.setAttribute("data-pages", pageCount);resultEl.setAttribute("data-text-length", textLength);- resultEl.textContent = `OK: ${pageCount} pages, ${textLength} chars`;+ resultEl.setAttribute("data-ocr-callbacks", ocrCallbackCalls);+ resultEl.setAttribute("data-ocr-bytes", ocrByteLength);+ resultEl.textContent = `OK: ${pageCount} pages, ${textLength} chars, OCR callback passed`;} catch (err) {status.textContent = "error";errorEl.style.display = "block";diff --git a/scripts/browser-compat/wasm-test.mjs b/scripts/browser-compat/wasm-test.mjsindex 01128df..e66c821 100644--- a/scripts/browser-compat/wasm-test.mjs+++ b/scripts/browser-compat/wasm-test.mjs@@ -7,6 +7,7 @@* Requires: playwright (npx playwright install chromium)* Expects: packages/wasm/pkg/ to contain the built WASM files* demo/docs/apple-10k-2024.pdf to exist+ * Set PLAYWRIGHT_CHROMIUM_EXECUTABLE to use an existing browser executable.*/import { createServer } from "node:http";@@ -67,35 +68,54 @@ async function main() {let browser;try {- browser = await chromium.launch();+ browser = await chromium.launch({+ executablePath: process.env.PLAYWRIGHT_CHROMIUM_EXECUTABLE,+ });const page = await browser.newPage();- // Collect console errorsconst errors = [];- page.on("pageerror", (err) => errors.push(err.message));+ let reportBrowserFailure;+ const browserFailure = new Promise((resolvePromise) => {+ reportBrowserFailure = resolvePromise;+ });+ page.on("pageerror", (error) => {+ errors.push(error.message);+ reportBrowserFailure(error);+ });+ page.on("console", (message) => {+ if (message.type() === "error") {+ const error = new Error(message.text());+ errors.push(error.message);+ reportBrowserFailure(error);+ }+ });console.log("Navigating to test page...");await page.goto(`${baseUrl}/scripts/browser-compat/wasm-test.html`);- // Wait for either #result or #error to appear (up to 120s for large PDFs)- await page.waitForSelector("#result[style*='block'], #error[style*='block']", {- timeout: 120_000,- });+ const completion = page+ .waitForSelector("#result[style*='block'], #error[style*='block']", { timeout: 120_000 })+ .then(() => null)+ .catch((error) => error);+ const failure = await Promise.race([completion, browserFailure]);+ if (failure) throw failure;const errorText = await page.locator("#error").textContent();if (errorText) {- console.error(`FAIL: ${errorText}`);- if (errors.length) console.error("Console errors:", errors);- process.exit(1);+ throw new Error(`${errorText}${errors.length ? `\nPage errors: ${errors.join("\n")}` : ""}`);}const resultText = await page.locator("#result").textContent();const pages = await page.locator("#result").getAttribute("data-pages");const textLength = await page.locator("#result").getAttribute("data-text-length");+ const ocrCallbacks = await page.locator("#result").getAttribute("data-ocr-callbacks");+ const ocrBytes = await page.locator("#result").getAttribute("data-ocr-bytes");console.log(`PASS: ${resultText}`);console.log(` Pages: ${pages}`);console.log(` Text length: ${textLength}`);+ console.log(` OCR callback calls: ${ocrCallbacks}`);+ console.log(` OCR input bytes: ${ocrBytes}`);if (errors.length) {console.warn("Browser console errors (non-fatal):", errors);@@ -106,7 +126,7 @@ async function main() {}}-main().catch((err) => {- console.error("Test runner failed:", err);+main().catch((error) => {+ console.error("Test runner failed:", error);process.exit(1);});