feat: add sycamore

Dockerfile

@@ -31,7 +31,7 @@ ENV HOME=/home/user \
     PYTHONUNBUFFERED=1 \
     GRADIO_SERVER_NAME=0.0.0.0
 
-RUN pip3 install --no-cache-dir --upgrade -r /code/requirements.txt
+RUN pip3 install --use-deprecated=legacy-resolver --no-cache-dir --upgrade -r /code/requirements.txt
 
 # Set the working directory to the user's home directory
 WORKDIR $HOME/app

app.py

@@ -22,6 +22,7 @@ from backends import (  # convert_zerox,
     convert_img2table,
     convert_marker,
     convert_mineru,
+    convert_sycamore,
     convert_unstructured,
 )
 from backends.settings import ENABLE_DEBUG_MODE
@@ -65,6 +66,8 @@ def convert_document(path, method, start_page=0, enabled=True):
         text, debug_image_paths = convert_mineru(path, file_name)
     elif method == "Gemini (API)":
         text, debug_image_paths = convert_gemini(path, file_name)
+    elif method == "Sycamore":
+        text, debug_image_paths = convert_sycamore(path, file_name)
     # elif method == "Zerox":
     #     text, debug_image_paths = convert_zerox(path, file_name)
     elif method == "Img2Table":
@@ -155,6 +158,7 @@ SUPPORTED_METHODS = [
     "Gemini (API)",
     "Img2Table",
     "GMFT",
+    "Sycamore",
     # "Zerox"
 ]
 

backends/__init__.py

@@ -4,6 +4,7 @@ from .gmft import convert_gmft
 from .img2table import convert_img2table
 from .marker import convert_marker
 from .mineru import convert_mineru
+from .syca import convert_sycamore
 from .unstructured import convert_unstructured
 
 # from .zerox import convert_zerox
@@ -17,4 +18,5 @@ __all__ = [
     # "convert_zerox",
     "convert_img2table",
     "convert_gmft",
+    "convert_sycamore",
 ]

backends/syca.py

@@ -0,0 +1,51 @@
+import logging
+from pathlib import Path
+
+import sycamore
+from sycamore import ExecMode
+from sycamore.data import Document
+from sycamore.data.document import DocumentPropertyTypes
+from sycamore.functions.document import DrawBoxes, split_and_convert_to_image
+from sycamore.transforms.partition import ArynPartitioner
+from sycamore.utils.markdown import elements_to_markdown
+
+from .settings import ENABLE_DEBUG_MODE
+
+logging.getLogger().setLevel(logging.INFO)
+SYCAMORE_DEBUG_PATH = Path("/tmp/sycamore")
+SYCAMORE_DEBUG_PATH.mkdir(exist_ok=True)
+
+
+paritioner = ArynPartitioner(
+    use_partitioning_service=False,
+    extract_table_structure=True,
+    use_ocr=True,
+    extract_images=True,
+)
+context = sycamore.init(
+    exec_mode=ExecMode.LOCAL,
+)
+
+
+def image_page_filename_fn(doc: Document) -> str:
+    page_num = doc.properties[DocumentPropertyTypes.PAGE_NUMBER]
+    return f"page_{page_num}.png"
+
+
+def convert_sycamore(path: str, file_name: str):
+    docset = context.read.binary(paths=path, binary_format="pdf").partition(
+        partitioner=paritioner,
+    )
+    debug_path = SYCAMORE_DEBUG_PATH / file_name
+    debug_path.mkdir(exist_ok=True)
+    image_paths = []
+
+    doc = docset.take_all()[0]
+    md = elements_to_markdown(doc.elements)
+
+    if ENABLE_DEBUG_MODE:
+        docset.flat_map(split_and_convert_to_image).map_batch(
+            DrawBoxes, f_constructor_kwargs={"draw_table_cells": True}
+        ).write.files(str(debug_path), filename_fn=image_page_filename_fn)
+        image_paths = [str(path) for path in debug_path.glob("*.png")]
+    return md, image_paths

requirements.txt

@@ -19,3 +19,4 @@ openai
 opencv-contrib-python
 gmft
 img2table
+sycamore-ai[local-inference]