Datasets:

bep40
/

grob-products-updated

import requests, re, json, os
H = "https://huggingface.co"
from pymupdf import Document as MuDoc

def norm(s): return re.sub(r'[^0-9]','',str(s))

# Simulate frontend search + verify each returned Imudex product's image maps to the
# correct column-aware region on its page (the exact fix logic).
pdf=requests.get(H+"/datasets/bep40/grob-products-updated/resolve/main/pdfs/imudex_2026.pdf", timeout=600).content
open("/tmp/imudex.pdf","wb").write(pdf)
doc=MuDoc("/tmp/imudex.pdf")

# page images per pdf index
def page_imgs(pgn):
    pg=doc[pgn]; out=[]
    for ip in pg.get_image_info():
        bb=ip["bbox"]; w=ip.get("width") or 0; h=ip.get("height") or 0
        if w>=160 and h>=110:
            out.append({"x0":bb[0],"y0":bb[1],"x1":bb[2],"y1":bb[3],"w":w,"h":h})
    return out
PAGE_IMGS={pgn:page_imgs(pgn) for pgn in range(len(doc))}

def side(x): return 'L' if x<567 else 'R'
def expected_img(sku_page_no, code_x, code_y):
    # find the image whose column matches code side, nearest vertically
    pgn=sku_page_no-1
    imgs=PAGE_IMGS[pgn]; s=side(code_x)
    cand=[im for im in imgs if side((im["x0"]+im["x1"])/2)==s]
    if not cand: return None
    bd=1e15; best=None
    for im in cand:
        icy=(im["y0"]+im["y1"])/2
        dd=abs(icy-code_y)
        if dd<bd: bd=dd; best=im
    return best

# re-dump page 81 code rows to validate mapping expectation for key SKUs
code_pat=re.compile(r'^[\d\s.]{6,}$')
d=doc[80].get_text("dict")
rows={}
for b in d["blocks"]:
    if b["type"]!=0: continue
    for l in b["lines"]:
        txt="".join(s["text"] for s in l["spans"]).strip()
        if txt and code_pat.match(txt.strip()):
            x0,y0,x1,y1=l["bbox"]
            rows[norm(txt)]=(round(x0,1),round(y0,1))
for sk,row in rows.items():
    im=expected_img(81, row[0], row[1])
    print(f"page81 code {sk} -> img bbox {[round(im['x0']),round(im['y0']),round(im['x1']),round(im['y1'])] if im else None} label={ 'L' if row[0]<567 else 'R'}")

	import requests, re, json, os
	H = "https://huggingface.co"
	from pymupdf import Document as MuDoc

	def norm(s): return re.sub(r'[^0-9]','',str(s))

	# Simulate frontend search + verify each returned Imudex product's image maps to the
	# correct column-aware region on its page (the exact fix logic).
	pdf=requests.get(H+"/datasets/bep40/grob-products-updated/resolve/main/pdfs/imudex_2026.pdf", timeout=600).content
	open("/tmp/imudex.pdf","wb").write(pdf)
	doc=MuDoc("/tmp/imudex.pdf")

	# page images per pdf index
	def page_imgs(pgn):
	pg=doc[pgn]; out=[]
	for ip in pg.get_image_info():
	bb=ip["bbox"]; w=ip.get("width") or 0; h=ip.get("height") or 0
	if w>=160 and h>=110:
	out.append({"x0":bb[0],"y0":bb[1],"x1":bb[2],"y1":bb[3],"w":w,"h":h})
	return out
	PAGE_IMGS={pgn:page_imgs(pgn) for pgn in range(len(doc))}

	def side(x): return 'L' if x<567 else 'R'
	def expected_img(sku_page_no, code_x, code_y):
	# find the image whose column matches code side, nearest vertically
	pgn=sku_page_no-1
	imgs=PAGE_IMGS[pgn]; s=side(code_x)
	cand=[im for im in imgs if side((im["x0"]+im["x1"])/2)==s]
	if not cand: return None
	bd=1e15; best=None
	for im in cand:
	icy=(im["y0"]+im["y1"])/2
	dd=abs(icy-code_y)
	if dd<bd: bd=dd; best=im
	return best

	# re-dump page 81 code rows to validate mapping expectation for key SKUs
	code_pat=re.compile(r'^[\d\s.]{6,}$')
	d=doc[80].get_text("dict")
	rows={}
	for b in d["blocks"]:
	if b["type"]!=0: continue
	for l in b["lines"]:
	txt="".join(s["text"] for s in l["spans"]).strip()
	if txt and code_pat.match(txt.strip()):
	x0,y0,x1,y1=l["bbox"]
	rows[norm(txt)]=(round(x0,1),round(y0,1))
	for sk,row in rows.items():
	im=expected_img(81, row[0], row[1])
	print(f"page81 code {sk} -> img bbox {[round(im['x0']),round(im['y0']),round(im['x1']),round(im['y1'])] if im else None} label={ 'L' if row[0]<567 else 'R'}")