MedCard / app.py
LLDDWW's picture
feat: use easyocr and enhance llm prompts
c103d7c
raw
history blame
17.7 kB
import json
import re
from typing import Any, Dict, List, Optional, Sequence
import easyocr
import gradio as gr
import numpy as np
import torch
from PIL import Image, ImageDraw
from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline
# --- OCR pipeline ---------------------------------------------------------
# Use a high-capacity OCR model for better accuracy on prescription labels.
OCR_LANGS = ["ko", "en"]
LLM_MODEL_ID = "Qwen/Qwen2.5-0.5B-Instruct"
def _load_ocr():
use_gpu = torch.cuda.is_available()
return easyocr.Reader(OCR_LANGS, gpu=use_gpu)
ocr_reader = _load_ocr()
def _load_llm():
device_map = "auto" if torch.cuda.is_available() else None
dtype = torch.float16 if torch.cuda.is_available() else torch.float32
model = AutoModelForCausalLM.from_pretrained(LLM_MODEL_ID, device_map=device_map, torch_dtype=dtype)
if device_map is None:
model = model.to(torch.device("cpu"))
tokenizer = AutoTokenizer.from_pretrained(LLM_MODEL_ID)
return model, tokenizer
LLM_MODEL, LLM_TOKENIZER = _load_llm()
# Korean keywords describing time slots on prescription labels.
TIME_KEYWORDS = [
"์•„์นจ",
"์ ์‹ฌ",
"์ €๋…",
"์ทจ์นจ",
"์ž๊ธฐ",
"์‹์ „",
"์‹ํ›„",
"์‹๊ฐ„",
"๊ธฐ์ƒ",
]
# Very small knowledge base for common Korean OTC medications.
MED_KNOWLEDGE: Sequence[Dict[str, Any]] = [
{
"keywords": ["ํƒ€์ด๋ ˆ๋†€", "์•„์„ธํŠธ์•„๋ฏธ๋…ธํŽœ", "acetaminophen"],
"category": "์ง„ํ†ตยทํ•ด์—ด์ œ",
"what_it_does": "๋ชธ์‚ด์ด๋‚˜ ๊ฐ๊ธฐ๋กœ ์—ด์ด ๋‚˜๊ฑฐ๋‚˜ ๋จธ๋ฆฌ๊ฐ€ ์•„ํ”Œ ๋•Œ ํ†ต์ฆ๊ณผ ์—ด์„ ๋‚ฎ์ถฐ ์ค๋‹ˆ๋‹ค.",
"example": "์˜ˆ: ์ˆ˜ํ•™์‹œํ—˜ ์ค€๋น„๋กœ ๊ธด์žฅํ–ˆ๋Š”๋ฐ ๋จธ๋ฆฌ๊ฐ€ ์ง€๋ˆ๊ฑฐ๋ฆด ๋•Œ, ํ•œ ์•Œ ๋ณต์šฉํ•˜๋ฉด ํ†ต์ฆ์ด ์ค„์–ด๋“ญ๋‹ˆ๋‹ค.",
"tip": "์œ„์— ๋ถ€๋‹ด์„ ์ค„์ด๊ธฐ ์œ„ํ•ด ๊ฐ„๋‹จํ•œ ๊ฐ„์‹๊ณผ ํ•จ๊ป˜ ๋ฌผ๊ณผ ๋ณต์šฉํ•˜๊ณ , ํ•˜๋ฃจ ์ด ๋ณต์šฉ ํšŸ์ˆ˜(์ผ๋ฐ˜์ ์œผ๋กœ 4ํšŒ ์ดํ•˜)๋ฅผ ๋„˜๊ธฐ์ง€ ๋งˆ์„ธ์š”.",
},
{
"keywords": ["์ด๋ถ€ํ”„๋กœํŽœ", "๋ถ€๋ฃจํŽœ", "ibuprofen"],
"category": "์ง„ํ†ตยท์†Œ์—ผ์ œ",
"what_it_does": "๋ชธ์† ์—ผ์ฆ์„ ๊ฐ€๋ผ์•‰ํžˆ๊ณ  ํ†ต์ฆ์„ ์™„ํ™”ํ•ด์„œ ๊ทผ์œกํ†ต์ด๋‚˜ ์น˜ํ†ต์— ์ž์ฃผ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค.",
"example": "์˜ˆ: ์ฒด์œก ์‹œ๊ฐ„์— ๋ฌด๋ฆŽ์„ ์‚ด์ง ์‚์—ˆ์„ ๋•Œ ๋ถ“๊ธฐ์™€ ์•„ํ””์„ ์ค„์—ฌ ์ค๋‹ˆ๋‹ค.",
"tip": "์‹ํ›„์— ๋ณต์šฉํ•˜๋ฉด ์† ์“ฐ๋ฆผ์„ ์ค„์ผ ์ˆ˜ ์žˆ๊ณ , ๋‹ค๋ฅธ ์†Œ์—ผ์ง„ํ†ต์ œ์™€๋Š” ์‹œ๊ฐ„ ๊ฐ„๊ฒฉ์„ ๋‘์„ธ์š”.",
},
{
"keywords": ["์‹œ์ž˜", "์„ธํ‹ฐ๋ฆฌ์ง„", "cetirizine", "์ง€๋ฅดํ…"],
"category": "์•Œ๋ ˆ๋ฅด๊ธฐ ์™„ํ™”์ œ",
"what_it_does": "์ฝ”๊ฐ€ ๊ฐ„์งˆ๊ฑฐ๋ฆฌ๊ฑฐ๋‚˜ ํ”ผ๋ถ€๊ฐ€ ๊ฐ€๋ ค์šธ ๋•Œ ์•Œ๋ ˆ๋ฅด๊ธฐ ๋ฐ˜์‘์„ ๊ฐ€๋ผ์•‰ํ˜€ ์ค๋‹ˆ๋‹ค.",
"example": "์˜ˆ: ๋ด„์ฒ  ๊ฝƒ๊ฐ€๋ฃจ ๋•Œ๋ฌธ์— ๊ธฐ์นจ๊ณผ ์ฝง๋ฌผ์ด ๋‚˜์˜ฌ ๋•Œ ์ฆ์ƒ์„ ์ค„์—ฌ ์ค๋‹ˆ๋‹ค.",
"tip": "์กธ๋ฆด ์ˆ˜ ์žˆ์œผ๋‹ˆ ์ฒซ ๋ณต์šฉ ํ›„์—๋Š” ์šด์ „์ด๋‚˜ ์ง‘์ค‘์ด ํ•„์š”ํ•œ ํ™œ๋™์€ ํ”ผํ•˜์„ธ์š”.",
},
{
"keywords": ["ํ›ผ์Šคํƒˆ", "pancreatin", "์œ„์žฅ", "์†Œํ™”์ œ"],
"category": "์†Œํ™”์ œ",
"what_it_does": "๊ธฐ๋ฆ„์ง„ ์Œ์‹์„ ๋จน๊ณ  ๋ฐฐ๊ฐ€ ๋”๋ถ€๋ฃฉํ•  ๋•Œ ์†Œํ™”๋ฅผ ๋„์™€ ์†์„ ํŽธํ•˜๊ฒŒ ํ•ด ์ค๋‹ˆ๋‹ค.",
"example": "์˜ˆ: ์น˜ํ‚จ์„ ๋งŽ์ด ๋จน์–ด ์†์ด ๋”๋ถ€๋ฃฉํ•  ๋•Œ ์†์„ ๊ฐ€๋ณ๊ฒŒ ํ•ด ์ค๋‹ˆ๋‹ค.",
"tip": "์‹ํ›„์— ๋ณต์šฉํ•˜๋ฉด ํšจ๊ณผ๊ฐ€ ์ข‹์œผ๋ฉฐ, ๋ณตํ†ต์ด ๊ณ„์†๋˜๋ฉด ๋ณ‘์›์„ ๋ฐฉ๋ฌธํ•˜์„ธ์š”.",
},
{
"keywords": ["๋น„ํƒ€๋ฏผ", "multivitamin", "vitamin"],
"category": "์˜์–‘์ œ",
"what_it_does": "๋ชธ์— ํ•„์š”ํ•œ ๋น„ํƒ€๋ฏผ์„ ์ฑ„์›Œ ํ”ผ๊ณคํ•จ์„ ์ค„์ด๊ณ  ๋ฉด์—ญ๋ ฅ์„ ๋•์Šต๋‹ˆ๋‹ค.",
"example": "์˜ˆ: ์‹œํ—˜ ์ค€๋น„๋กœ ์ž ์„ ์ค„์˜€์„ ๋•Œ ๋ชธ์ด ์ง€์น˜์ง€ ์•Š๋„๋ก ๋„์™€์ค๋‹ˆ๋‹ค.",
"tip": "ํ•˜๋ฃจ ๊ถŒ์žฅ๋Ÿ‰์„ ์ง€์ผœ ๊พธ์ค€ํžˆ ๋ณต์šฉํ•˜๋ฉด ๋” ํšจ๊ณผ์ ์ด๋ฉฐ, ๋ฌผ๊ณผ ํ•จ๊ป˜ ์‚ผํ‚ค์„ธ์š”.",
},
]
def _extract_time_slots(text: str) -> List[str]:
slots = []
for kw in TIME_KEYWORDS:
if kw in text:
slots.append(kw)
# Also capture explicit times like 08:00 ํ˜น์€ 8์‹œ
for match in re.findall(r"(\d{1,2}[:์‹œ]\d{0,2})", text):
norm = match.replace("์‹œ", ":")
if norm.endswith(":"):
norm += "00"
if norm not in slots:
slots.append(norm)
return slots
STOPWORDS = {"์šฉ๋ฒ•", "์šฉ๋Ÿ‰", "๋ณต์šฉ", "๋ฐฉ๋ฒ•", "์•ฝ", "์ •"}
def _extract_medications(text: str) -> List[Dict[str, Optional[str]]]:
meds: List[Dict[str, Optional[str]]] = []
pattern = re.compile(
r"([๊ฐ€-ํžฃA-Za-z]{2,})[\sยท]*(\d+[\./]?\d*\s*(?:mg|mL|ML|ml|์ •|์บก์А))?"
)
seen: set[str] = set()
for match in pattern.finditer(text):
name = match.group(1)
if name in STOPWORDS or len(name) <= 1:
continue
if any(sw in name for sw in STOPWORDS):
continue
name_norm = name.strip()
if name_norm in seen:
continue
seen.add(name_norm)
dose = match.group(2).strip() if match.group(2) else None
meds.append({"name": name_norm, "dose": dose})
return meds
def parse_fields(raw: str) -> Dict[str, Any]:
"""Extract drug name and dosage information from OCR text."""
collapsed = raw.replace("\n", " ")
collapsed = re.sub(r"\s+", " ", collapsed)
medications = _extract_medications(collapsed)
first = medications[0] if medications else {"name": None, "dose": None}
drug_name = first.get("name")
dose_per_intake = first.get("dose")
times_per_day: Optional[int] = None
times_match = re.search(r"(?:1์ผ|ํ•˜๋ฃจ)\s*(\d+)\s*ํšŒ", collapsed)
if times_match:
times_per_day = int(times_match.group(1))
time_slots = _extract_time_slots(collapsed)
return {
"drug_name": drug_name,
"dose_per_intake": dose_per_intake,
"times_per_day": times_per_day,
"time_slots": time_slots or None,
"medications": medications,
}
def ocr_and_parse(image: Image.Image) -> Dict[str, Any]:
np_img = np.array(image.convert("RGB"))
results = ocr_reader.readtext(np_img, detail=1, paragraph=False)
segments: List[Dict[str, Any]] = []
lines: List[str] = []
for bbox, text, confidence in results:
cleaned = text.strip()
if not cleaned:
continue
lines.append(cleaned)
segments.append({
"text": cleaned,
"confidence": float(confidence),
"bbox": bbox,
})
raw_text = "\n".join(lines)
fields = parse_fields(raw_text)
warnings: List[str] = []
if not fields["drug_name"]:
warnings.append("์•ฝ ์ด๋ฆ„ ์ธ์‹์ด ๋ถˆํ™•์‹คํ•ฉ๋‹ˆ๋‹ค.")
if not fields["times_per_day"]:
warnings.append("1์ผ ํšŸ์ˆ˜๋ฅผ ์ฐพ์ง€ ๋ชปํ–ˆ์Šต๋‹ˆ๋‹ค (์˜ˆ: 1์ผ 3ํšŒ).")
return {
"raw_text": raw_text,
"fields": fields,
"warnings": warnings,
"segments": segments,
}
def render_card(fields: Dict[str, Any]) -> Image.Image:
width, height = 720, 400
img = Image.new("RGB", (width, height), "white")
draw = ImageDraw.Draw(img)
header_text = "์˜ค๋Š˜ ๋ณต์šฉ ์ผ์ •"
draw.rectangle((0, 0, width, 60), fill=(230, 240, 255))
draw.text((24, 18), header_text, fill=(0, 0, 0))
y = 90
def add_line(label: str, value: Optional[str]):
nonlocal y
draw.text((24, y), label, fill=(60, 60, 60))
display = value if value else "-"
draw.text((180, y), f": {display}", fill=(0, 0, 0))
y += 34
add_line("์•ฝ ์ด๋ฆ„", fields.get("drug_name"))
add_line("1ํšŒ ์šฉ๋Ÿ‰", fields.get("dose_per_intake"))
add_line("1์ผ ํšŸ์ˆ˜", str(fields.get("times_per_day") or ""))
slots = fields.get("time_slots") or []
add_line("์‹œ๊ฐ„๋Œ€", ", ".join(slots) if slots else None)
footer = "โ€ป ์˜๋ฃŒ์ง„ ์ฒ˜๋ฐฉ์ด ์šฐ์„ ์ด๋ฉฐ, ๋ณธ ์•ฑ์€ ์ฐธ๊ณ ์šฉ์ž…๋‹ˆ๋‹ค."
draw.text((24, height - 60), footer, fill=(120, 120, 120))
return img
def to_csv_row(output: Dict[str, Any]) -> str:
fields = output["fields"]
row = [
fields.get("drug_name") or "",
fields.get("dose_per_intake") or "",
str(fields.get("times_per_day") or ""),
";".join(fields.get("time_slots") or []),
]
return ",".join(row)
def _match_knowledge(name: str) -> Optional[Dict[str, Any]]:
lowered = name.lower()
for info in MED_KNOWLEDGE:
for kw in info["keywords"]:
if kw.lower() in lowered or lowered in kw.lower():
return info
return None
def build_kb_explanations(output: Dict[str, Any]) -> str:
meds = output["fields"].get("medications") or []
if not meds:
return (
"### ์•ฝ ์„ค๋ช…\n"
"- ์•ฝ ์ด๋ฆ„์„ ์ •ํ™•ํžˆ ์ธ์‹ํ•˜์ง€ ๋ชปํ–ˆ์–ด์š”. ์‚ฌ์ง„์„ ๋‹ค์‹œ ์ฐ๊ฑฐ๋‚˜ ์•ฝ์‚ฌ์—๊ฒŒ ์ง์ ‘ ํ™•์ธํ•ด ์ฃผ์„ธ์š”.\n"
"\n> โš ๏ธ ์˜๋ฃŒ์ง„ ์ฒ˜๋ฐฉ๊ณผ ๋ณต์•ฝ ์ง€์‹œ๊ฐ€ ๊ฐ€์žฅ ์šฐ์„ ์ž…๋‹ˆ๋‹ค."
)
lines = ["### ์‰ฝ๊ฒŒ ์•Œ์•„๋ณด๋Š” ์•ฝ ์„ค๋ช…"]
for med in meds:
name = med.get("name") or "์ด๋ฆ„ ๋ฏธํ™•์ธ"
info = _match_knowledge(name) if name else None
dose = med.get("dose")
if info:
lines.append(
f"- **{name}** ({info['category']})"
)
if dose:
lines.append(f" - ์•ฝ ๋ด‰ํˆฌ์— ์ ํžŒ ์šฉ๋Ÿ‰: `{dose}`")
lines.append(f" - ํ•˜๋Š” ์ผ: {info['what_it_does']}")
lines.append(f" - ์ค‘ํ•™์ƒ ์˜ˆ์‹œ: {info['example']}")
lines.append(f" - ๋ณต์šฉ ํŒ: {info['tip']}")
else:
lines.append(f"- **{name}**")
if dose:
lines.append(f" - ์•ฝ ๋ด‰ํˆฌ ์šฉ๋Ÿ‰: `{dose}`")
lines.append(
" - ์•„์ง ๋ฐ์ดํ„ฐ๊ฐ€ ์—†์–ด์š”. ์•ฝ ์ด๋ฆ„์„ ๋‹ค์‹œ ํ™•์ธํ•˜๊ฑฐ๋‚˜ ์•ฝ์‚ฌ์—๊ฒŒ ๋ฌผ์–ด๋ณด์„ธ์š”."
)
lines.append("\n> โš ๏ธ ์‹ค์ œ ๋ณต์•ฝ์€ ์˜์‚ฌยท์•ฝ์‚ฌ์˜ ์ง€์‹œ์— ๋ฐ˜๋“œ์‹œ ๋”ฐ๋ฅด์„ธ์š”.")
return "\n".join(lines)
def generate_llm_explanations(output: Dict[str, Any]) -> str:
meds = output["fields"].get("medications") or []
if not meds:
return (
"์•ฝ ์ด๋ฆ„์„ ์ œ๋Œ€๋กœ ์ธ์‹ํ•˜์ง€ ๋ชปํ–ˆ์–ด์š”. ์‚ฌ์ง„์„ ๋‹ค์‹œ ์ฐ๊ฑฐ๋‚˜ ์•ฝ์‚ฌ์—๊ฒŒ ์ง์ ‘ ํ™•์ธํ•ด ์ฃผ์„ธ์š”."
)
med_lines = []
for idx, med in enumerate(meds, 1):
name = med.get("name") or "์ด๋ฆ„ ๋ฏธํ™•์ธ"
dose = med.get("dose") or "์šฉ๋Ÿ‰ ์ •๋ณด ์—†์Œ"
med_lines.append(f"{idx}. {name} โ€” {dose}")
context = "\n".join(med_lines)
raw_text = output.get("raw_text", "")
system_prompt = (
"๋‹น์‹ ์€ ์•ฝ์‚ฌ ์„ ์ƒ๋‹˜์ž…๋‹ˆ๋‹ค. ์–ด๋ ค์šด ์˜ํ•™ ์šฉ์–ด๋ฅผ ์“ฐ์ง€ ๋ง๊ณ , ์ค‘ํ•™์ƒ๋„ ์ดํ•ดํ•  ์ˆ˜ ์žˆ๋Š” ๋งํˆฌ๋กœ ์นœ์ ˆํ•˜๊ฒŒ ์„ค๋ช…ํ•˜์„ธ์š”."
)
user_prompt = (
"๋‹ค์Œ์€ ์•ฝ๋ด‰ํˆฌ์—์„œ OCR๋กœ ์ถ”์ถœํ•œ ์ „์ฒด ํ…์ŠคํŠธ์ž…๋‹ˆ๋‹ค. ์•ฝ ์ด๋ฆ„๊ณผ ๋ณต์šฉ ์ง€์‹œ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ๊ฐ ์•ฝ์˜ ์ •๋ณด๋ฅผ ์•„์ฃผ ์‰ฝ๊ฒŒ ์ •๋ฆฌํ•ด ์ฃผ์„ธ์š”.\n"
"์š”๊ตฌ ์‚ฌํ•ญ:\n"
"1. ๊ฐ ์•ฝ๋งˆ๋‹ค ์•„๋ž˜ ํ•ญ๋ชฉ์„ bullet ํ˜•์‹์œผ๋กœ ์ž‘์„ฑํ•ฉ๋‹ˆ๋‹ค.\n"
" - ์•ฝ ์ด๋ฆ„: (๊ฐ€๋Šฅํ•˜๋ฉด ํ•œ๊ธ€/์˜๋ฌธ ๋ณ‘๊ธฐ)\n"
" - ์–ด๋–ค ์•ฝ์ธ์ง€ ํ•œ ์ค„ ์„ค๋ช…\n"
" - ๋ณต์šฉ ์˜ˆ์‹œ: ์–ธ์ œ, ์–ด๋–ค ์ƒํ™ฉ์—์„œ ๋ณต์šฉํ•˜๋ฉด ์ข‹์€์ง€ ์˜ˆ์‹œ\n"
" - ๋ณต์šฉ ๋ฐฉ๋ฒ• ์˜ˆ์‹œ: 1ํšŒ ์šฉ๋Ÿ‰/ํ•˜๋ฃจ ํšŸ์ˆ˜๊ฐ€ ์žˆ๋‹ค๋ฉด ์–ธ๊ธ‰\n"
" - ๋ถ€์ž‘์šฉ ๋˜๋Š” ์ฃผ์˜์‚ฌํ•ญ: ํ”ํ•œ ๋ถ€์ž‘์šฉ, ํ”ผํ•ด์•ผ ํ•  ํ–‰๋™\n"
"2. ์–ด๋ ค์šด ์˜ํ•™ ์šฉ์–ด๋Š” ํ”ผํ•˜๊ณ , ์ค‘ํ•™์ƒ๋„ ์ดํ•ดํ•  ์ˆ˜ ์žˆ๋Š” ๋งํˆฌ๋กœ ์ž‘์„ฑํ•ฉ๋‹ˆ๋‹ค.\n"
"3. ์•ฝ ์ด๋ฆ„์„ ํ™•์‹คํžˆ ๋ชจ๋ฅด๋ฉด โ€˜์ด๋ฆ„ ๋ฏธํ™•์ธโ€™์ด๋ผ๊ณ  ์“ฐ๊ณ , ์•ฝ์‚ฌ์—๊ฒŒ ํ™•์ธํ•˜๋ผ๊ณ  ์•ˆ๋‚ดํ•ฉ๋‹ˆ๋‹ค.\n"
"4. ๋งˆ์ง€๋ง‰ ๋ฌธ๋‹จ์— ๋ฐ˜๋“œ์‹œ โ€˜์‹ค์ œ ๋ณต์•ฝ์€ ์˜์‚ฌยท์•ฝ์‚ฌ์˜ ์ง€์‹œ๋ฅผ ๋”ฐ๋ฅด์„ธ์š”โ€™ ๋ฌธ์žฅ์„ ํฌํ•จํ•˜์„ธ์š”.\n"
f"\n์•ฝ ๋ชฉ๋ก(์ถ”์ถœ ์š”์•ฝ):\n{context}\n\nOCR ์›๋ฌธ ์ „์ฒด:\n{raw_text}\n"
)
messages = [
{"role": "system", "content": system_prompt},
{"role": "user", "content": user_prompt},
]
input_ids = LLM_TOKENIZER.apply_chat_template(
messages,
add_generation_prompt=True,
return_tensors="pt",
)
input_ids = input_ids.to(LLM_MODEL.device)
with torch.no_grad():
output_ids = LLM_MODEL.generate(
input_ids,
max_new_tokens=480,
temperature=0.7,
top_p=0.9,
do_sample=True,
eos_token_id=LLM_TOKENIZER.eos_token_id,
)
generated_ids = output_ids[0][input_ids.shape[1]:]
text = LLM_TOKENIZER.decode(generated_ids, skip_special_tokens=True).strip()
return text
def build_explanations(output: Dict[str, Any]) -> str:
try:
llm_text = generate_llm_explanations(output)
if llm_text:
return llm_text
except Exception as err: # pragma: no cover - safe fallback
print(f"[WARN] LLM generation failed: {err}", flush=True)
return build_kb_explanations(output)
def format_warnings(warnings: List[str]) -> str:
if not warnings:
return "โœ… ์ธ์‹๋œ ์ •๋ณด๊ฐ€ ์ถฉ๋ถ„ํ•ด์š”. ๋ณต์•ฝ ์‹œ๊ฐ„๋งŒ ์ž˜ ์ง€์ผœ ์ฃผ์„ธ์š”."
lines = ["### ํ™•์ธํ•ด ์ฃผ์„ธ์š”"]
for warn in warnings:
lines.append(f"- {warn}")
lines.append("\n> ์˜๋ฃŒ์ง„์˜ ์ง€์‹œ๊ฐ€ ๊ฐ€์žฅ ์ •ํ™•ํ•ฉ๋‹ˆ๋‹ค.")
return "\n".join(lines)
def run_pipeline(image: Optional[Image.Image]):
if image is None:
return (
"์ด๋ฏธ์ง€๋ฅผ ์—…๋กœ๋“œํ•˜์„ธ์š”.",
None,
None,
"์ด๋ฏธ์ง€๋ฅผ ๋จผ์ € ์—…๋กœ๋“œํ•ด ์ฃผ์„ธ์š”.",
"๐Ÿ“ท ์•ฝ ๋ด‰ํˆฌ ์‚ฌ์ง„์„ ์˜ฌ๋ฆฌ๋ฉด ์ธ์‹์ด ์‹œ์ž‘๋ผ์š”.",
"",
)
output = ocr_and_parse(image)
card = render_card(output["fields"])
csv_row = to_csv_row(output)
json_text = json.dumps(output, ensure_ascii=False, indent=2)
explanations = build_explanations(output)
warnings_md = format_warnings(output.get("warnings", []))
return json_text, card, csv_row, explanations, warnings_md, output.get("raw_text", "")
CUSTOM_CSS = """
body {background: radial-gradient(circle at top left, #f5f0ff 0%, #fff7ec 60%, #ffffff 100%);}
.gradio-container {max-width: 1180px !important; margin: auto; font-family: 'Noto Sans KR', sans-serif;}
.hero {
background: linear-gradient(120deg, rgba(123, 97, 255, 0.12), rgba(255, 207, 117, 0.18));
border-radius: 28px;
padding: 36px 44px;
box-shadow: 0 20px 40px rgba(66, 46, 138, 0.08);
margin-bottom: 32px;
}
.hero h1 {font-size: 2.4rem; font-weight: 700; color: #1f1c3b; margin-bottom: 12px;}
.hero p {color: #514c7b; font-size: 1.05rem; line-height: 1.6; max-width: 640px;}
.glass-panel {background: rgba(255, 255, 255, 0.72); backdrop-filter: blur(18px); border-radius: 26px; padding: 28px; box-shadow: 0 12px 32px rgba(80, 60, 160, 0.12);}
.panel-title {font-weight: 700; font-size: 1.2rem; margin-bottom: 18px; color: #2f2355;}
.primary-btn button {background: linear-gradient(120deg, #7c62ff, #ffa74d); border: none; color: white; font-weight: 600; border-radius: 999px; padding: 12px 22px; box-shadow: 0 12px 24px rgba(124, 98, 255, 0.25);}
.primary-btn button:hover {opacity: 0.95; transform: translateY(-1px);}
.output-card {background: rgba(255, 255, 255, 0.88); border-radius: 22px; padding: 24px; box-shadow: inset 0 0 0 1px rgba(124, 98, 255, 0.08), 0 14px 30px rgba(49, 32, 114, 0.12);}
.notice {background: rgba(255, 247, 226, 0.9); border-radius: 18px; padding: 18px; color: #7a4b00; box-shadow: inset 0 0 0 1px rgba(255, 193, 96, 0.3);}
.csv-box textarea {font-family: 'JetBrains Mono', monospace;}
.gr-image {border-radius: 20px !important; box-shadow: 0 10px 20px rgba(60, 40, 120, 0.15);}
.accordion {border-radius: 20px !important;}
"""
HERO_HTML = """
<div class="hero">
<h1>MedCard-KR ยท ์•ฝ๋ด‰ํˆฌ ํ•œ ์ปท์œผ๋กœ ์ดํ•ดํ•˜๋Š” ๋ณต์šฉ ์•ˆ๋‚ด</h1>
<p>์‚ฌ์ง„ ์† ์•ฝ ์ด๋ฆ„์„ OCR๋กœ ์ฝ์–ด ๋“ค์ด๊ณ , Qwen LLM์ด ์ค‘ํ•™์ƒ๋„ ์ดํ•ดํ•  ์ˆ˜ ์žˆ๋Š” ๋งํˆฌ๋กœ ์•ฝ์„ ์„ค๋ช…ํ•ด ๋“œ๋ฆฝ๋‹ˆ๋‹ค.
๋ณต์šฉ ์ผ์ • ์นด๋“œ์™€ CSV๊นŒ์ง€ ํ•œ ๋ฒˆ์— ๋ฐ›์•„ ๋ณด์„ธ์š”.</p>
</div>
"""
with gr.Blocks(theme=gr.themes.Soft(), css=CUSTOM_CSS) as demo:
gr.HTML(HERO_HTML)
with gr.Row():
with gr.Column(scale=4, elem_classes=["glass-panel"]):
gr.Markdown("### 1. ์•ฝ ๋ด‰ํˆฌ ์‚ฌ์ง„์„ ์—…๋กœ๋“œํ•˜์„ธ์š”")
img_in = gr.Image(type="pil", label="์•ฝ ๋ด‰ํˆฌ/๋ผ๋ฒจ ์‚ฌ์ง„", height=360)
warn_md = gr.Markdown("๐Ÿ“ท ์•ฝ ๋ด‰ํˆฌ ์‚ฌ์ง„์„ ์˜ฌ๋ฆฌ๋ฉด ์ธ์‹์ด ์‹œ์ž‘๋ผ์š”.", elem_classes=["notice"])
btn = gr.Button("์ธ์‹ & ์„ค๋ช… ์ƒ์„ฑ", elem_classes=["primary-btn"])
with gr.Column(scale=6, elem_classes=["glass-panel"]):
gr.Markdown("### 2. ๊ฒฐ๊ณผ๋ฅผ ํ™•์ธํ•˜์„ธ์š”")
explain_md = gr.Markdown("์—ฌ๊ธฐ์— ์•ฝ ์„ค๋ช…์ด ํ‘œ์‹œ๋ฉ๋‹ˆ๋‹ค.", elem_classes=["output-card"])
raw_box = gr.Textbox(label="OCR ์›๋ฌธ ํ…์ŠคํŠธ", lines=5, interactive=False)
card_out = gr.Image(type="pil", label="์ผ์ • ์นด๋“œ(๋ฏธ๋ฆฌ๋ณด๊ธฐ)")
csv_box = gr.Textbox(label="CSV(์•ฝ๋ช…,1ํšŒ์šฉ๋Ÿ‰,1์ผํšŸ์ˆ˜,์‹œ๊ฐ„๋Œ€)", lines=2, elem_classes=["csv-box"])
with gr.Accordion("์„ธ๋ถ€ JSON ๊ฒฐ๊ณผ", open=False, elem_classes=["accordion"]):
json_out = gr.Code(label="์ธ์‹ ๊ฒฐ๊ณผ(JSON)")
btn.click(
run_pipeline,
inputs=img_in,
outputs=[json_out, card_out, csv_box, explain_md, warn_md, raw_box],
)
gr.Markdown(
"""
> โ„น๏ธ **์ฃผ์˜**: ์ด ์„œ๋น„์Šค๋Š” ์ฐธ๊ณ ์šฉ ๋„๊ตฌ์ด๋ฉฐ, ์‹ค์ œ ๋ณต์•ฝ์€ ๋ฐ˜๋“œ์‹œ ์˜์‚ฌยท์•ฝ์‚ฌ์˜ ์ง€์‹œ์— ๋”ฐ๋ผ ์ฃผ์„ธ์š”.
"""
)
if __name__ == "__main__":
demo.queue().launch()