chore: fix wasm release and benchmark grouping (#30)

Author: raphaelmansuy

.github/workflows/release-wasm.yml

@@ -94,19 +94,9 @@ jobs:
           name: wasm-package
           path: crates/edgeparse-wasm/pkg/*.tgz
 
-      - name: Publish WASM package to npm
-        env:
-          NODE_AUTH_TOKEN: ${{ secrets.NPM_TOKEN }}
+      - name: Skip WASM npm publication
         run: |
-          cd crates/edgeparse-wasm/pkg
-          OUTPUT=$(npm publish --access public 2>&1) && echo "$OUTPUT" || {
-            echo "$OUTPUT"
-            if echo "$OUTPUT" | grep -Eq "cannot publish over the previously published versions|You cannot publish over the previously published version"; then
-              echo "edgeparse-wasm already published at this version — skipping."
-            else
-              exit 1
-            fi
-          }
+          echo "::warning::WASM npm publication is disabled. The package tarball will still be uploaded to the GitHub Release."
 
       - name: Upload npm tarball to GitHub Release
         env:

benchmark/compare_all.py

@@ -7,12 +7,17 @@
 Engine groups:
   Non-OCR (fast, no ML models): edgeparse, opendataloader, pymupdf4llm,
                                  markitdown, liteparse
+  Hybrid (backend-assisted):    edgeparse, opendataloader_hybrid_docling_fast,
+                                 opendataloader_hybrid_hancom
   OCR / ML (model-heavy):       edgeparse, docling, marker, mineru
 
 Usage:
     # Non-OCR comparison (fast, recommended first run):
     uv run python compare_all.py --group non-ocr --install
 
+    # OCR/ML comparison (slow — installs isolated venvs for marker & mineru):
+    uv run python compare_all.py --group hybrid
+
     # OCR/ML comparison (slow — installs isolated venvs for marker & mineru):
     uv run python compare_all.py --group ocr --install
 
@@ -30,6 +35,7 @@
 
 Via Makefile:
     make bench-non-ocr
+    make bench-hybrid
     make bench-ocr
     make bench-ocr OCR_ENGINES=docling
     make bench-compare-all
@@ -53,7 +59,7 @@
 sys.path.insert(0, str(Path(__file__).parent / "src"))
 
 from engine_registry import (
-    ENGINES, ENGINE_META, NON_OCR_ENGINES, OCR_ENGINES,
+    ENGINES, ENGINE_META, NON_OCR_ENGINES, HYBRID_ENGINES, OCR_ENGINES,
     available_engines, display_name,
 )
 from evaluation_schema import missing_evaluation_requirements
@@ -79,6 +85,8 @@
 ALL_ENGINES = [
     # Non-OCR (fast)
     "edgeparse", "opendataloader", "pymupdf4llm", "markitdown", "liteparse",
+    # Hybrid
+    "opendataloader_hybrid_docling_fast", "opendataloader_hybrid_hancom",
     # OCR / ML
     "docling", "marker", "mineru",
 ]
@@ -461,6 +469,7 @@ def _parse_args(argv: Optional[Sequence[str]] = None) -> argparse.Namespace:
         epilog="""
 Examples:
   uv run python compare_all.py --group non-ocr --install
+  uv run python compare_all.py --group hybrid
   uv run python compare_all.py --group ocr --install
   uv run python compare_all.py --engines edgeparse,docling,pymupdf4llm --install
   uv run python compare_all.py --all --no-run
@@ -469,9 +478,9 @@ def _parse_args(argv: Optional[Sequence[str]] = None) -> argparse.Namespace:
     )
     parser.add_argument(
         "--group",
-        choices=["non-ocr", "ocr", "all"],
+        choices=["non-ocr", "hybrid", "ocr", "all"],
         default=None,
-        help="Engine group to benchmark: non-ocr (fast), ocr (ML/model-heavy), all",
+        help="Engine group to benchmark: non-ocr (fast), hybrid (backend-assisted), ocr (ML/model-heavy), all",
     )
     parser.add_argument(
         "--engines",
@@ -542,6 +551,9 @@ def main(argv: Optional[Sequence[str]] = None) -> None:
     elif args.group == "non-ocr":
         engines = list(NON_OCR_ENGINES)
         default_title = "EdgeParse Benchmark — Non-OCR Tools"
+    elif args.group == "hybrid":
+        engines = list(HYBRID_ENGINES)
+        default_title = "EdgeParse Benchmark — Hybrid Tools"
     elif args.group == "ocr":
         engines = list(OCR_ENGINES)
         default_title = "EdgeParse Benchmark — OCR / ML Tools"

benchmark/src/engine_registry.py

@@ -3,6 +3,8 @@
 Engines:
   * ``edgeparse``      — Rust binary built from this repository (always available)
   * ``opendataloader`` — Published Java/Python package (opendataloader-pdf ≥ 2.0)
+  * ``opendataloader_hybrid_docling_fast`` — OpenDataLoader hybrid with Docling Fast backend
+  * ``opendataloader_hybrid_hancom``       — OpenDataLoader hybrid with Hancom backend
   * ``pymupdf4llm``    — PyMuPDF4LLM (pip install pymupdf4llm)
   * ``markitdown``     — Microsoft MarkItDown (pip install markitdown[all])
   * ``liteparse``      — LlamaIndex LiteParse (@llamaindex/liteparse, Node.js CLI)
@@ -14,6 +16,7 @@
 
 Engine groups (for benchmark segmentation):
   NON_OCR_ENGINES — no ML models, no GPU; pure text/geometry extraction
+  HYBRID_ENGINES  — mixed local + backend routing for complex pages
   OCR_ENGINES     — require deep-learning models; GPU optional but recommended
 """
 
@@ -35,6 +38,12 @@
     "liteparse",
 ]
 
+HYBRID_ENGINES: List[str] = [
+    "edgeparse",
+    "opendataloader_hybrid_docling_fast",
+    "opendataloader_hybrid_hancom",
+]
+
 OCR_ENGINES: List[str] = [
     "edgeparse",
     "docling",
@@ -54,15 +63,17 @@
 
 # Engine display metadata: name → (display_name, pip_package, description)
 ENGINE_META: Dict[str, tuple] = {
-    "edgeparse":      ("EdgeParse",      None,                    "Rust PDF engine (this repo)"),
-    "opendataloader": ("OpenDataLoader", "opendataloader-pdf",    "Java/Python PDF engine"),
-    "pymupdf4llm":    ("PyMuPDF4LLM",   "pymupdf4llm",           "PyMuPDF for LLM/RAG"),
-    "markitdown":     ("MarkItDown",     "markitdown[all]",       "Microsoft multi-format converter"),
-    "liteparse":      ("LiteParse",      "@llamaindex/liteparse", "LlamaIndex local PDF parser"),
+    "edgeparse":      ("EdgeParse",                            None,                    "Rust PDF engine (this repo)"),
+    "opendataloader": ("OpenDataLoader",                       "opendataloader-pdf",    "Java/Python PDF engine"),
+    "opendataloader_hybrid_docling_fast": ("OpenDataLoader [hybrid/docling-fast]", None, "OpenDataLoader hybrid with Docling Fast backend"),
+    "opendataloader_hybrid_hancom":       ("OpenDataLoader [hybrid/hancom]",       None, "OpenDataLoader hybrid with Hancom backend"),
+    "pymupdf4llm":    ("PyMuPDF4LLM",                         "pymupdf4llm",           "PyMuPDF for LLM/RAG"),
+    "markitdown":     ("MarkItDown",                           "markitdown[all]",       "Microsoft multi-format converter"),
+    "liteparse":      ("LiteParse",                            "@llamaindex/liteparse", "LlamaIndex local PDF parser"),
     # OCR / ML engines
-    "docling":        ("Docling",        "docling",               "IBM Research document parser [OCR/ML]"),
-    "marker":         ("Marker",         "marker-pdf",            "Marker PDF — Surya OCR [isolated venv]"),
-    "mineru":         ("MinerU",         "mineru[all]",           "OpenDataLab PDF extractor [isolated venv]"),
+    "docling":        ("Docling",                              "docling",               "IBM Research document parser [OCR/ML]"),
+    "marker":         ("Marker",                               "marker-pdf",            "Marker PDF — Surya OCR [isolated venv]"),
+    "mineru":         ("MinerU",                               "mineru[all]",           "OpenDataLab PDF extractor [isolated venv]"),
 }
 
 # ── Auto-register external engines ───────────────────────────────────────────
@@ -77,6 +88,8 @@ def _try_register(name: str, module_name: str, version_label: str = "installed")
         pass
 
 _try_register("opendataloader", "pdf_parser_opendataloader", "published")
+_try_register("opendataloader_hybrid_docling_fast", "pdf_parser_opendataloader_hybrid_docling_fast", "local-hybrid")
+_try_register("opendataloader_hybrid_hancom", "pdf_parser_opendataloader_hybrid_hancom", "local-hybrid")
 _try_register("docling",        "pdf_parser_docling",        "installed")
 _try_register("pymupdf4llm",    "pdf_parser_pymupdf4llm",    "installed")
 _try_register("markitdown",     "pdf_parser_markitdown",     "installed")

benchmark/src/pdf_parser_edgeparse.py

@@ -41,10 +41,14 @@ def to_markdown(document_paths: List[Path], _input_path, output_dir: Path):
         "--quiet",
     ]
 
+    env = dict(**__import__("os").environ)
+    env["EDGEPARSE_RASTER_TABLE_OCR"] = "off"
+
     result = subprocess.run(
         command,
         capture_output=True,
         text=True,
+        env=env,
     )
 
     if result.returncode != 0:

crates/edgeparse-cli/src/main.rs

@@ -84,6 +84,10 @@ struct Cli {
     #[arg(long = "image-dir")]
     image_dir: Option<String>,
 
+    /// Raster table OCR recovery (on, off)
+    #[arg(long = "raster-table-ocr", default_value = "on")]
+    raster_table_ocr: String,
+
     /// Pages to extract (e.g., "1,3,5-7")
     #[arg(long = "pages")]
     pages: Option<String>,
@@ -206,6 +210,11 @@ fn build_config(cli: &Cli) -> edgeparse_core::api::config::ProcessingConfig {
     use edgeparse_core::api::config::*;
     use edgeparse_core::api::filter::FilterConfig;
 
+    let raster_table_ocr = std::env::var("EDGEPARSE_RASTER_TABLE_OCR")
+        .ok()
+        .map(|value| !matches!(value.as_str(), "off" | "false" | "0"))
+        .unwrap_or_else(|| !matches!(cli.raster_table_ocr.as_str(), "off" | "false" | "0"));
+
     let formats = if let Some(ref fmt) = cli.format {
         fmt.split(',')
             .filter_map(|s| match s.trim() {
@@ -258,6 +267,7 @@ fn build_config(cli: &Cli) -> edgeparse_core::api::config::ProcessingConfig {
             _ => ImageFormat::Png,
         },
         image_dir: cli.image_dir.clone(),
+        raster_table_ocr,
         pages: cli.pages.clone(),
         include_header_footer: cli.include_header_footer,
         hybrid: match cli.hybrid.as_str() {

crates/edgeparse-core/src/api/config.rs

@@ -114,6 +114,8 @@ pub struct ProcessingConfig {
     pub image_format: ImageFormat,
     /// Directory for extracted images
     pub image_dir: Option<String>,
+    /// Enable raster table OCR recovery on image-based tables
+    pub raster_table_ocr: bool,
     /// Pages to extract (e.g., "1,3,5-7")
     pub pages: Option<String>,
     /// Include headers/footers in output
@@ -150,6 +152,7 @@ impl Default for ProcessingConfig {
             image_output: ImageOutput::External,
             image_format: ImageFormat::Png,
             image_dir: None,
+            raster_table_ocr: true,
             pages: None,
             include_header_footer: false,
             hybrid: HybridBackend::Off,
@@ -175,6 +178,7 @@ mod tests {
         assert_eq!(config.table_method, TableMethod::Default);
         assert_eq!(config.image_output, ImageOutput::External);
         assert_eq!(config.image_format, ImageFormat::Png);
+        assert!(config.raster_table_ocr);
         assert_eq!(config.hybrid, HybridBackend::Off);
         assert_eq!(config.hybrid_timeout, 30000);
     }

crates/edgeparse-core/src/lib.rs

@@ -58,17 +58,19 @@ pub fn convert(
     for (&page_num, &page_id) in &pages_map {
         let page_chunks = extract_page_chunks(&raw_doc.document, page_num, page_id)?;
         let mut recovered_tables = Vec::new();
-        if let Some(page_info) = page_info_list
-            .iter()
-            .find(|info| info.page_number == page_num)
-        {
-            recovered_tables = recover_raster_table_borders(
-                input_path,
-                &page_info.crop_box,
-                page_num,
-                &page_chunks.text_chunks,
-                &page_chunks.image_chunks,
-            );
+        if config.raster_table_ocr {
+            if let Some(page_info) = page_info_list
+                .iter()
+                .find(|info| info.page_number == page_num)
+            {
+                recovered_tables = recover_raster_table_borders(
+                    input_path,
+                    &page_info.crop_box,
+                    page_num,
+                    &page_chunks.text_chunks,
+                    &page_chunks.image_chunks,
+                );
+            }
         }
         let mut elements: Vec<ContentElement> = page_chunks
             .text_chunks
@@ -124,14 +126,16 @@ pub fn convert(
     doc.creation_date = raw_doc.metadata.creation_date;
     doc.modification_date = raw_doc.metadata.modification_date;
 
-    for (page_idx, page) in pipeline_state.pages.iter_mut().enumerate() {
-        if let Some(page_info) = page_info_list.get(page_idx) {
-            recover_page_raster_table_cell_text(
-                input_path,
-                &page_info.crop_box,
-                page_info.page_number,
-                page,
-            );
+    if config.raster_table_ocr {
+        for (page_idx, page) in pipeline_state.pages.iter_mut().enumerate() {
+            if let Some(page_info) = page_info_list.get(page_idx) {
+                recover_page_raster_table_cell_text(
+                    input_path,
+                    &page_info.crop_box,
+                    page_info.page_number,
+                    page,
+                );
+            }
         }
     }
 

docs/07-cicd-publishing.md

@@ -235,7 +235,7 @@ fast on mismatches.
 
 - Builds the browser-targeted WASM package with `wasm-pack`
 - Syncs the npm package version from the tag
-- Publishes `edgeparse-wasm`
+- npm publication is currently disabled
 - Uploads the generated npm tarball to the GitHub Release
 
 ### `release-cli.yml`
@@ -278,7 +278,7 @@ and Homebrew. Docker publishing remains CI-driven through `release-docker.yml`.
 
 Crates.io versions are immutable. Bump the version and retag.
 
-### npm publish fails on platform packages or the WASM package
+### npm publish fails on platform packages
 
 Use a Classic Automation token for `NPM_TOKEN`. Granular tokens often miss one
 or more package names and produce `E403 Forbidden`.