pengjunjie/ragflow
1
0
Fork 1
mirror of https://github.com/infiniflow/ragflow.git synced 2025-12-08 20:42:30 +08:00
Code Issues Packages Projects Releases Wiki Activity

Feat: init dataflow. (#9791)

Browse Source 
### What problem does this PR solve?

#9790

Close #9782

### Type of change

- [x] New Feature (non-breaking change which adds functionality)
This commit is contained in:
Kevin Hu
2025-08-28 18:40:32 +08:00
committed by GitHub
parent a246949b77
commit c27172b3bc
19 changed files with 1020 additions and 166 deletions
Download Patch File Download Diff File Expand all files Collapse all files

106
deepdoc/parser/pdf_parser.py
View File

@ -93,6 +93,7 @@ class RAGFlowPdfParser:
model_dir, "updown_concat_xgb.model"))
self.page_from = 0
self.column_num = 1
def __char_width(self, c):
return (c["x1"] - c["x0"]) // max(len(c["text"]), 1)
@ -427,10 +428,18 @@ class RAGFlowPdfParser:
i += 1
self.boxes = bxs
def _naive_vertical_merge(self):
def _naive_vertical_merge(self, zoomin=3):
bxs = Recognizer.sort_Y_firstly(
self.boxes, np.median(
self.mean_height) / 3)
column_width = np.median([b["x1"] - b["x0"] for b in self.boxes])
self.column_num = int(self.page_images[0].size[0] / zoomin / column_width)
if column_width < self.page_images[0].size[0] / zoomin / self.column_num:
logging.info("Multi-column................... {} {}".format(column_width,
self.page_images[0].size[0] / zoomin / self.column_num))
self.boxes = self.sort_X_by_page(self.boxes, column_width / self.column_num)
i = 0
while i + 1 < len(bxs):
b = bxs[i]
@ -1139,20 +1148,94 @@ class RAGFlowPdfParser:
need_image, zoomin, return_html, False)
return self.__filterout_scraps(deepcopy(self.boxes), zoomin), tbls
def parse_into_bboxes(self, fnm, callback=None, zoomin=3):
start = timer()
self.__images__(fnm, zoomin)
if callback:
callback(0.40, "OCR finished ({:.2f}s)".format(timer() - start))
start = timer()
self._layouts_rec(zoomin)
if callback:
callback(0.63, "Layout analysis ({:.2f}s)".format(timer() - start))
start = timer()
self._table_transformer_job(zoomin)
if callback:
callback(0.83, "Table analysis ({:.2f}s)".format(timer() - start))
start = timer()
self._text_merge()
self._concat_downward()
self._naive_vertical_merge(zoomin)
if callback:
callback(0.92, "Text merged ({:.2f}s)".format(timer() - start))
start = timer()
tbls, figs = self._extract_table_figure(True, zoomin, True, True, True)
def insert_table_figures(tbls_or_figs, layout_type):
def min_rectangle_distance(rect1, rect2):
import math
pn1, left1, right1, top1, bottom1 = rect1
pn2, left2, right2, top2, bottom2 = rect2
if (right1 >= left2 and right2 >= left1 and
bottom1 >= top2 and bottom2 >= top1):
return 0 + (pn1-pn2)*10000
if right1 < left2:
dx = left2 - right1
elif right2 < left1:
dx = left1 - right2
else:
dx = 0
if bottom1 < top2:
dy = top2 - bottom1
elif bottom2 < top1:
dy = top1 - bottom2
else:
dy = 0
return math.sqrt(dx*dx + dy*dy) + (pn1-pn2)*10000
for (img, txt), poss in tbls_or_figs:
bboxes = [(i, (b["page_number"], b["x0"], b["x1"], b["top"], b["bottom"])) for i, b in enumerate(self.boxes)]
dists = [(min_rectangle_distance((pn, left, right, top, bott), rect),i) for i, rect in bboxes for pn, left, right, top, bott in poss]
min_i = np.argmin(dists, axis=0)[0]
min_i, rect = bboxes[dists[min_i][-1]]
if isinstance(txt, list):
txt = "\n".join(txt)
self.boxes.insert(min_i, {
"page_number": rect[0], "x0": rect[1], "x1": rect[2], "top": rect[3], "bottom": rect[4], "layout_type": layout_type, "text": txt, "image": img
})
for b in self.boxes:
b["position_tag"] = self._line_tag(b, zoomin)
b["image"] = self.crop(b["position_tag"], zoomin)
insert_table_figures(tbls, "table")
insert_table_figures(figs, "figure")
if callback:
callback(1, "Structured ({:.2f}s)".format(timer() - start))
return deepcopy(self.boxes)
@staticmethod
def remove_tag(txt):
return re.sub(r"@@[\t0-9.-]+?##", "", txt)
def crop(self, text, ZM=3, need_position=False):
imgs = []
@staticmethod
def extract_positions(txt):
poss = []
for tag in re.findall(r"@@[0-9-]+\t[0-9.\t]+##", text):
for tag in re.findall(r"@@[0-9-]+\t[0-9.\t]+##", txt):
pn, left, right, top, bottom = tag.strip(
"#").strip("@").split("\t")
left, right, top, bottom = float(left), float(
right), float(top), float(bottom)
poss.append(([int(p) - 1 for p in pn.split("-")],
left, right, top, bottom))
return poss
def crop(self, text, ZM=3, need_position=False):
imgs = []
poss = self.extract_positions(text)
if not poss:
if need_position:
return None, None
@ -1296,8 +1379,8 @@ class VisionParser(RAGFlowPdfParser):
def __call__(self, filename, from_page=0, to_page=100000, **kwargs):
callback = kwargs.get("callback", lambda prog, msg: None)
self.__images__(fnm=filename, zoomin=3, page_from=from_page, page_to=to_page, **kwargs)
zoomin = kwargs.get("zoomin", 3)
self.__images__(fnm=filename, zoomin=zoomin, page_from=from_page, page_to=to_page, callback=callback)
total_pdf_pages = self.total_page
@ -1311,16 +1394,19 @@ class VisionParser(RAGFlowPdfParser):
if pdf_page_num < start_page or pdf_page_num >= end_page:
continue
docs = picture_vision_llm_chunk(
text = picture_vision_llm_chunk(
binary=img_binary,
vision_model=self.vision_model,
prompt=vision_llm_describe_prompt(page=pdf_page_num+1),
callback=callback,
)
if kwargs.get("callback"):
kwargs["callback"](idx*1./len(self.page_images), f"Processed: {idx+1}/{len(self.page_images)}")
if docs:
all_docs.append(docs)
return [(doc, "") for doc in all_docs], []
if text:
width, height = self.page_images[idx].size
all_docs.append((text, f"{pdf_page_num+1} 0 {width/zoomin} 0 {height/zoomin}"))
return all_docs, []
if __name__ == "__main__":