Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
|
@@ -9,6 +9,11 @@ from threading import Thread
|
|
| 9 |
import json
|
| 10 |
import requests
|
| 11 |
import cv2
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 12 |
import gradio as gr
|
| 13 |
import spaces
|
| 14 |
import torch
|
|
@@ -16,220 +21,227 @@ from loguru import logger
|
|
| 16 |
from PIL import Image
|
| 17 |
from transformers import AutoProcessor, Gemma3ForConditionalGeneration, TextIteratorStreamer
|
| 18 |
|
| 19 |
-
# CSV/TXT ๋ถ์
|
| 20 |
import pandas as pd
|
| 21 |
-
# PDF ํ
์คํธ ์ถ์ถ
|
| 22 |
import PyPDF2
|
| 23 |
|
| 24 |
-
|
| 25 |
-
#
|
| 26 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 27 |
def clear_cuda_cache():
|
| 28 |
"""CUDA ์บ์๋ฅผ ๋ช
์์ ์ผ๋ก ๋น์๋๋ค."""
|
| 29 |
if torch.cuda.is_available():
|
| 30 |
torch.cuda.empty_cache()
|
| 31 |
gc.collect()
|
| 32 |
|
| 33 |
-
|
| 34 |
-
#
|
| 35 |
-
|
| 36 |
SERPHOUSE_API_KEY = os.getenv("SERPHOUSE_API_KEY", "")
|
| 37 |
|
| 38 |
-
##############################################################################
|
| 39 |
-
# ๊ฐ๋จํ ํค์๋ ์ถ์ถ ํจ์ (ํ๊ธ + ์ํ๋ฒณ + ์ซ์ + ๊ณต๋ฐฑ ๋ณด์กด)
|
| 40 |
-
##############################################################################
|
| 41 |
def extract_keywords(text: str, top_k: int = 5) -> str:
|
| 42 |
-
"""
|
| 43 |
-
1) ํ๊ธ(๊ฐ-ํฃ), ์์ด(a-zA-Z), ์ซ์(0-9), ๊ณต๋ฐฑ๋ง ๋จ๊น
|
| 44 |
-
2) ๊ณต๋ฐฑ ๊ธฐ์ค ํ ํฐ ๋ถ๋ฆฌ
|
| 45 |
-
3) ์ต๋ top_k๊ฐ๋ง
|
| 46 |
-
"""
|
| 47 |
text = re.sub(r"[^a-zA-Z0-9๊ฐ-ํฃ\s]", "", text)
|
| 48 |
tokens = text.split()
|
| 49 |
-
|
| 50 |
-
return " ".join(key_tokens)
|
| 51 |
|
| 52 |
-
##############################################################################
|
| 53 |
-
# SerpHouse Live endpoint ํธ์ถ
|
| 54 |
-
# - ์์ 20๊ฐ ๊ฒฐ๊ณผ JSON์ LLM์ ๋๊ธธ ๋ link, snippet ๋ฑ ๋ชจ๋ ํฌํจ
|
| 55 |
-
##############################################################################
|
| 56 |
def do_web_search(query: str) -> str:
|
| 57 |
-
"""
|
| 58 |
-
์์ 20๊ฐ 'organic' ๊ฒฐ๊ณผ item ์ ์ฒด(์ ๋ชฉ, link, snippet ๋ฑ)๋ฅผ
|
| 59 |
-
JSON ๋ฌธ์์ด ํํ๋ก ๋ฐํ
|
| 60 |
-
"""
|
| 61 |
try:
|
| 62 |
url = "https://api.serphouse.com/serp/live"
|
| 63 |
-
|
| 64 |
-
# ๊ธฐ๋ณธ GET ๋ฐฉ์์ผ๋ก ํ๋ผ๋ฏธํฐ ๊ฐ์ํํ๊ณ ๊ฒฐ๊ณผ ์๋ฅผ 20๊ฐ๋ก ์ ํ
|
| 65 |
params = {
|
| 66 |
"q": query,
|
| 67 |
"domain": "google.com",
|
| 68 |
-
"serp_type": "web",
|
| 69 |
"device": "desktop",
|
| 70 |
"lang": "en",
|
| 71 |
-
"num": "20"
|
| 72 |
}
|
| 73 |
-
|
| 74 |
-
headers = {
|
| 75 |
-
"Authorization": f"Bearer {SERPHOUSE_API_KEY}"
|
| 76 |
-
}
|
| 77 |
-
|
| 78 |
logger.info(f"SerpHouse API ํธ์ถ ์ค... ๊ฒ์์ด: {query}")
|
| 79 |
-
logger.info(f"์์ฒญ URL: {url} - ํ๋ผ๋ฏธํฐ: {params}")
|
| 80 |
-
|
| 81 |
-
# GET ์์ฒญ ์ํ
|
| 82 |
response = requests.get(url, headers=headers, params=params, timeout=60)
|
| 83 |
response.raise_for_status()
|
| 84 |
-
|
| 85 |
-
logger.info(f"SerpHouse API ์๋ต ์ํ ์ฝ๋: {response.status_code}")
|
| 86 |
data = response.json()
|
| 87 |
-
|
| 88 |
-
# ๋ค์ํ ์๋ต ๊ตฌ์กฐ ์ฒ๋ฆฌ
|
| 89 |
results = data.get("results", {})
|
| 90 |
organic = None
|
| 91 |
-
|
| 92 |
-
# ๊ฐ๋ฅํ ์๋ต ๊ตฌ์กฐ 1
|
| 93 |
if isinstance(results, dict) and "organic" in results:
|
| 94 |
organic = results["organic"]
|
| 95 |
-
|
| 96 |
-
# ๊ฐ๋ฅํ ์๋ต ๊ตฌ์กฐ 2 (์ค์ฒฉ๋ results)
|
| 97 |
elif isinstance(results, dict) and "results" in results:
|
| 98 |
if isinstance(results["results"], dict) and "organic" in results["results"]:
|
| 99 |
organic = results["results"]["organic"]
|
| 100 |
-
|
| 101 |
-
# ๊ฐ๋ฅํ ์๋ต ๊ตฌ์กฐ 3 (์ต์์ organic)
|
| 102 |
elif "organic" in data:
|
| 103 |
organic = data["organic"]
|
| 104 |
-
|
| 105 |
if not organic:
|
| 106 |
logger.warning("์๋ต์์ organic ๊ฒฐ๊ณผ๋ฅผ ์ฐพ์ ์ ์์ต๋๋ค.")
|
| 107 |
-
|
| 108 |
-
if isinstance(results, dict):
|
| 109 |
-
logger.debug(f"results ๊ตฌ์กฐ: {list(results.keys())}")
|
| 110 |
-
return "No web search results found or unexpected API response structure."
|
| 111 |
-
|
| 112 |
-
# ๊ฒฐ๊ณผ ์ ์ ํ ๋ฐ ์ปจํ
์คํธ ๊ธธ์ด ์ต์ ํ
|
| 113 |
max_results = min(20, len(organic))
|
| 114 |
limited_organic = organic[:max_results]
|
| 115 |
-
|
| 116 |
-
# ๊ฒฐ๊ณผ ํ์ ๊ฐ์ - ๋งํฌ๋ค์ด ํ์์ผ๋ก ์ถ๋ ฅํ์ฌ ๊ฐ๋
์ฑ ํฅ์
|
| 117 |
summary_lines = []
|
| 118 |
for idx, item in enumerate(limited_organic, start=1):
|
| 119 |
-
title = item.get("title", "
|
| 120 |
link = item.get("link", "#")
|
| 121 |
-
snippet = item.get("snippet", "
|
| 122 |
displayed_link = item.get("displayed_link", link)
|
| 123 |
-
|
| 124 |
-
# ๋งํฌ๋ค์ด ํ์ (๋งํฌ ํด๋ฆญ ๊ฐ๋ฅ)
|
| 125 |
summary_lines.append(
|
| 126 |
-
f"###
|
| 127 |
f"{snippet}\n\n"
|
| 128 |
f"**์ถ์ฒ**: [{displayed_link}]({link})\n\n"
|
| 129 |
f"---\n"
|
| 130 |
)
|
| 131 |
-
|
| 132 |
-
# ๋ชจ๋ธ์๊ฒ ๋ช
ํํ ์ง์นจ ์ถ๊ฐ
|
| 133 |
instructions = """
|
| 134 |
# ์น ๊ฒ์ ๊ฒฐ๊ณผ
|
| 135 |
์๋๋ ๊ฒ์ ๊ฒฐ๊ณผ์
๋๋ค. ์ง๋ฌธ์ ๋ต๋ณํ ๋ ์ด ์ ๋ณด๋ฅผ ํ์ฉํ์ธ์:
|
| 136 |
-
1. ๊ฐ ๊ฒฐ๊ณผ์ ์ ๋ชฉ, ๋ด์ฉ, ์ถ์ฒ ๋งํฌ๋ฅผ
|
| 137 |
-
2. ๋ต๋ณ์ ๊ด๋ จ ์ ๋ณด์ ์ถ์ฒ๋ฅผ ๋ช
์์ ์ผ๋ก ์ธ์ฉํ์ธ์ (์: "
|
| 138 |
-
3. ์๋ต์ ์ค์ ์ถ์ฒ ๋งํฌ๋ฅผ
|
| 139 |
-
4. ์ฌ๋ฌ ์ถ์ฒ์ ์ ๋ณด๋ฅผ ์ข
ํฉํ์ฌ
|
|
|
|
| 140 |
"""
|
| 141 |
-
|
| 142 |
-
search_results = instructions + "\n".join(summary_lines)
|
| 143 |
-
logger.info(f"๊ฒ์ ๊ฒฐ๊ณผ {len(limited_organic)}๊ฐ ์ฒ๋ฆฌ ์๋ฃ")
|
| 144 |
-
return search_results
|
| 145 |
-
|
| 146 |
except Exception as e:
|
| 147 |
-
logger.error(f"
|
| 148 |
-
return f"
|
| 149 |
-
|
| 150 |
|
| 151 |
-
|
| 152 |
-
#
|
| 153 |
-
|
| 154 |
MAX_CONTENT_CHARS = 2000
|
| 155 |
-
MAX_INPUT_LENGTH = 2096
|
| 156 |
model_id = os.getenv("MODEL_ID", "VIDraft/Gemma-3-R1984-4B")
|
| 157 |
-
|
| 158 |
processor = AutoProcessor.from_pretrained(model_id, padding_side="left")
|
| 159 |
model = Gemma3ForConditionalGeneration.from_pretrained(
|
| 160 |
model_id,
|
| 161 |
device_map="auto",
|
| 162 |
torch_dtype=torch.bfloat16,
|
| 163 |
-
attn_implementation="eager"
|
| 164 |
)
|
| 165 |
MAX_NUM_IMAGES = int(os.getenv("MAX_NUM_IMAGES", "5"))
|
| 166 |
|
| 167 |
-
|
| 168 |
-
|
| 169 |
-
#
|
| 170 |
-
##############################################################################
|
| 171 |
def analyze_csv_file(path: str) -> str:
|
| 172 |
-
"""
|
| 173 |
-
CSV ํ์ผ์ ์ ์ฒด ๋ฌธ์์ด๋ก ๋ณํ. ๋๋ฌด ๊ธธ ๊ฒฝ์ฐ ์ผ๋ถ๋ง ํ์.
|
| 174 |
-
"""
|
| 175 |
try:
|
| 176 |
df = pd.read_csv(path)
|
| 177 |
if df.shape[0] > 50 or df.shape[1] > 10:
|
| 178 |
df = df.iloc[:50, :10]
|
| 179 |
df_str = df.to_string()
|
| 180 |
if len(df_str) > MAX_CONTENT_CHARS:
|
| 181 |
-
df_str = df_str[:MAX_CONTENT_CHARS] + "\n...(
|
| 182 |
-
return f"**[CSV
|
| 183 |
except Exception as e:
|
| 184 |
-
return f"
|
| 185 |
-
|
| 186 |
|
| 187 |
def analyze_txt_file(path: str) -> str:
|
| 188 |
-
"""
|
| 189 |
-
TXT ํ์ผ ์ ๋ฌธ ์ฝ๊ธฐ. ๋๋ฌด ๊ธธ๋ฉด ์ผ๋ถ๋ง ํ์.
|
| 190 |
-
"""
|
| 191 |
try:
|
| 192 |
with open(path, "r", encoding="utf-8") as f:
|
| 193 |
text = f.read()
|
| 194 |
if len(text) > MAX_CONTENT_CHARS:
|
| 195 |
-
text = text[:MAX_CONTENT_CHARS] + "\n...(
|
| 196 |
-
return f"**[TXT
|
| 197 |
except Exception as e:
|
| 198 |
-
return f"
|
| 199 |
-
|
| 200 |
|
| 201 |
def pdf_to_markdown(pdf_path: str) -> str:
|
| 202 |
-
"""
|
| 203 |
-
PDF ํ
์คํธ๋ฅผ Markdown์ผ๋ก ๋ณํ. ํ์ด์ง๋ณ๋ก ๊ฐ๋จํ ํ
์คํธ ์ถ์ถ.
|
| 204 |
-
"""
|
| 205 |
text_chunks = []
|
| 206 |
try:
|
| 207 |
with open(pdf_path, "rb") as f:
|
| 208 |
reader = PyPDF2.PdfReader(f)
|
| 209 |
max_pages = min(5, len(reader.pages))
|
| 210 |
for page_num in range(max_pages):
|
| 211 |
-
|
| 212 |
-
page_text = page.extract_text() or ""
|
| 213 |
page_text = page_text.strip()
|
| 214 |
if page_text:
|
| 215 |
if len(page_text) > MAX_CONTENT_CHARS // max_pages:
|
| 216 |
-
page_text = page_text[:MAX_CONTENT_CHARS // max_pages] + "...(
|
| 217 |
-
text_chunks.append(f"##
|
| 218 |
if len(reader.pages) > max_pages:
|
| 219 |
-
text_chunks.append(f"\n...(
|
| 220 |
except Exception as e:
|
| 221 |
-
return f"
|
| 222 |
-
|
| 223 |
full_text = "\n".join(text_chunks)
|
| 224 |
if len(full_text) > MAX_CONTENT_CHARS:
|
| 225 |
-
full_text = full_text[:MAX_CONTENT_CHARS] + "\n...(
|
|
|
|
| 226 |
|
| 227 |
-
|
| 228 |
-
|
| 229 |
-
|
| 230 |
-
##############################################################################
|
| 231 |
-
# ์ด๋ฏธ์ง/๋น๋์ค ์
๋ก๋ ์ ํ ๊ฒ์ฌ
|
| 232 |
-
##############################################################################
|
| 233 |
def count_files_in_new_message(paths: list[str]) -> tuple[int, int]:
|
| 234 |
image_count = 0
|
| 235 |
video_count = 0
|
|
@@ -240,7 +252,6 @@ def count_files_in_new_message(paths: list[str]) -> tuple[int, int]:
|
|
| 240 |
image_count += 1
|
| 241 |
return image_count, video_count
|
| 242 |
|
| 243 |
-
|
| 244 |
def count_files_in_history(history: list[dict]) -> tuple[int, int]:
|
| 245 |
image_count = 0
|
| 246 |
video_count = 0
|
|
@@ -256,95 +267,76 @@ def count_files_in_history(history: list[dict]) -> tuple[int, int]:
|
|
| 256 |
image_count += 1
|
| 257 |
return image_count, video_count
|
| 258 |
|
| 259 |
-
|
| 260 |
def validate_media_constraints(message: dict, history: list[dict]) -> bool:
|
| 261 |
-
media_files = []
|
| 262 |
-
for f in message["files"]:
|
| 263 |
-
if re.search(r"\.(png|jpg|jpeg|gif|webp)$", f, re.IGNORECASE) or f.endswith(".mp4"):
|
| 264 |
-
media_files.append(f)
|
| 265 |
-
|
| 266 |
new_image_count, new_video_count = count_files_in_new_message(media_files)
|
| 267 |
history_image_count, history_video_count = count_files_in_history(history)
|
| 268 |
image_count = history_image_count + new_image_count
|
| 269 |
video_count = history_video_count + new_video_count
|
| 270 |
-
|
| 271 |
if video_count > 1:
|
| 272 |
-
gr.Warning("
|
| 273 |
return False
|
| 274 |
if video_count == 1:
|
| 275 |
if image_count > 0:
|
| 276 |
-
gr.Warning("
|
| 277 |
return False
|
| 278 |
if "<image>" in message["text"]:
|
| 279 |
-
gr.Warning("
|
| 280 |
return False
|
| 281 |
if video_count == 0 and image_count > MAX_NUM_IMAGES:
|
| 282 |
-
gr.Warning(f"
|
| 283 |
return False
|
| 284 |
-
|
| 285 |
if "<image>" in message["text"]:
|
| 286 |
image_files = [f for f in message["files"] if re.search(r"\.(png|jpg|jpeg|gif|webp)$", f, re.IGNORECASE)]
|
| 287 |
image_tag_count = message["text"].count("<image>")
|
| 288 |
if image_tag_count != len(image_files):
|
| 289 |
-
gr.Warning("
|
| 290 |
return False
|
| 291 |
-
|
| 292 |
return True
|
| 293 |
|
| 294 |
-
|
| 295 |
-
|
| 296 |
-
#
|
| 297 |
-
##############################################################################
|
| 298 |
def downsample_video(video_path: str) -> list[tuple[Image.Image, float]]:
|
| 299 |
vidcap = cv2.VideoCapture(video_path)
|
| 300 |
fps = vidcap.get(cv2.CAP_PROP_FPS)
|
| 301 |
total_frames = int(vidcap.get(cv2.CAP_PROP_FRAME_COUNT))
|
| 302 |
frame_interval = max(int(fps), int(total_frames / 10))
|
| 303 |
frames = []
|
| 304 |
-
|
| 305 |
for i in range(0, total_frames, frame_interval):
|
| 306 |
vidcap.set(cv2.CAP_PROP_POS_FRAMES, i)
|
| 307 |
success, image = vidcap.read()
|
| 308 |
if success:
|
| 309 |
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
|
| 310 |
-
# ์ด๋ฏธ์ง ํฌ๊ธฐ ์ค์ด๊ธฐ ์ถ๊ฐ
|
| 311 |
image = cv2.resize(image, (0, 0), fx=0.5, fy=0.5)
|
| 312 |
pil_image = Image.fromarray(image)
|
| 313 |
timestamp = round(i / fps, 2)
|
| 314 |
frames.append((pil_image, timestamp))
|
| 315 |
if len(frames) >= 5:
|
| 316 |
break
|
| 317 |
-
|
| 318 |
vidcap.release()
|
| 319 |
return frames
|
| 320 |
|
| 321 |
-
|
| 322 |
def process_video(video_path: str) -> tuple[list[dict], list[str]]:
|
| 323 |
content = []
|
| 324 |
-
temp_files = []
|
| 325 |
-
|
| 326 |
frames = downsample_video(video_path)
|
| 327 |
-
for
|
| 328 |
-
pil_image, timestamp = frame
|
| 329 |
with tempfile.NamedTemporaryFile(delete=False, suffix=".png") as temp_file:
|
| 330 |
pil_image.save(temp_file.name)
|
| 331 |
-
temp_files.append(temp_file.name)
|
| 332 |
-
content.append({"type": "text", "text": f"
|
| 333 |
content.append({"type": "image", "url": temp_file.name})
|
| 334 |
-
|
| 335 |
return content, temp_files
|
| 336 |
|
| 337 |
-
|
| 338 |
-
|
| 339 |
-
#
|
| 340 |
-
##############################################################################
|
| 341 |
def process_interleaved_images(message: dict) -> list[dict]:
|
| 342 |
parts = re.split(r"(<image>)", message["text"])
|
| 343 |
content = []
|
| 344 |
-
image_index = 0
|
| 345 |
-
|
| 346 |
image_files = [f for f in message["files"] if re.search(r"\.(png|jpg|jpeg|gif|webp)$", f, re.IGNORECASE)]
|
| 347 |
-
|
| 348 |
for part in parts:
|
| 349 |
if part == "<image>" and image_index < len(image_files):
|
| 350 |
content.append({"type": "image", "url": image_files[image_index]})
|
|
@@ -356,10 +348,9 @@ def process_interleaved_images(message: dict) -> list[dict]:
|
|
| 356 |
content.append({"type": "text", "text": part})
|
| 357 |
return content
|
| 358 |
|
| 359 |
-
|
| 360 |
-
|
| 361 |
-
#
|
| 362 |
-
##############################################################################
|
| 363 |
def is_image_file(file_path: str) -> bool:
|
| 364 |
return bool(re.search(r"\.(png|jpg|jpeg|gif|webp)$", file_path, re.IGNORECASE))
|
| 365 |
|
|
@@ -367,45 +358,29 @@ def is_video_file(file_path: str) -> bool:
|
|
| 367 |
return file_path.endswith(".mp4")
|
| 368 |
|
| 369 |
def is_document_file(file_path: str) -> bool:
|
| 370 |
-
return (
|
| 371 |
-
file_path.lower().endswith(".pdf")
|
| 372 |
-
or file_path.lower().endswith(".csv")
|
| 373 |
-
or file_path.lower().endswith(".txt")
|
| 374 |
-
)
|
| 375 |
-
|
| 376 |
|
| 377 |
def process_new_user_message(message: dict) -> tuple[list[dict], list[str]]:
|
| 378 |
-
temp_files = []
|
| 379 |
-
|
| 380 |
if not message["files"]:
|
| 381 |
return [{"type": "text", "text": message["text"]}], temp_files
|
| 382 |
-
|
| 383 |
video_files = [f for f in message["files"] if is_video_file(f)]
|
| 384 |
image_files = [f for f in message["files"] if is_image_file(f)]
|
| 385 |
csv_files = [f for f in message["files"] if f.lower().endswith(".csv")]
|
| 386 |
txt_files = [f for f in message["files"] if f.lower().endswith(".txt")]
|
| 387 |
pdf_files = [f for f in message["files"] if f.lower().endswith(".pdf")]
|
| 388 |
-
|
| 389 |
content_list = [{"type": "text", "text": message["text"]}]
|
| 390 |
-
|
| 391 |
for csv_path in csv_files:
|
| 392 |
-
|
| 393 |
-
content_list.append({"type": "text", "text": csv_analysis})
|
| 394 |
-
|
| 395 |
for txt_path in txt_files:
|
| 396 |
-
|
| 397 |
-
content_list.append({"type": "text", "text": txt_analysis})
|
| 398 |
-
|
| 399 |
for pdf_path in pdf_files:
|
| 400 |
-
|
| 401 |
-
content_list.append({"type": "text", "text": pdf_markdown})
|
| 402 |
-
|
| 403 |
if video_files:
|
| 404 |
video_content, video_temp_files = process_video(video_files[0])
|
| 405 |
content_list += video_content
|
| 406 |
temp_files.extend(video_temp_files)
|
| 407 |
return content_list, temp_files
|
| 408 |
-
|
| 409 |
if "<image>" in message["text"] and image_files:
|
| 410 |
interleaved_content = process_interleaved_images({"text": message["text"], "files": image_files})
|
| 411 |
if content_list and content_list[0]["type"] == "text":
|
|
@@ -414,16 +389,14 @@ def process_new_user_message(message: dict) -> tuple[list[dict], list[str]]:
|
|
| 414 |
else:
|
| 415 |
for img_path in image_files:
|
| 416 |
content_list.append({"type": "image", "url": img_path})
|
| 417 |
-
|
| 418 |
return content_list, temp_files
|
| 419 |
|
| 420 |
-
|
| 421 |
-
##############################################################################
|
| 422 |
# history -> LLM ๋ฉ์์ง ๋ณํ
|
| 423 |
-
|
| 424 |
def process_history(history: list[dict]) -> list[dict]:
|
| 425 |
messages = []
|
| 426 |
-
current_user_content
|
| 427 |
for item in history:
|
| 428 |
if item["role"] == "assistant":
|
| 429 |
if current_user_content:
|
|
@@ -439,36 +412,25 @@ def process_history(history: list[dict]) -> list[dict]:
|
|
| 439 |
if is_image_file(file_path):
|
| 440 |
current_user_content.append({"type": "image", "url": file_path})
|
| 441 |
else:
|
| 442 |
-
current_user_content.append({"type": "text", "text": f"[
|
| 443 |
-
|
| 444 |
if current_user_content:
|
| 445 |
messages.append({"role": "user", "content": current_user_content})
|
| 446 |
-
|
| 447 |
return messages
|
| 448 |
|
| 449 |
-
|
| 450 |
-
|
| 451 |
-
#
|
| 452 |
-
##############################################################################
|
| 453 |
def _model_gen_with_oom_catch(**kwargs):
|
| 454 |
-
"""
|
| 455 |
-
๋ณ๋ ์ค๋ ๋์์ OutOfMemoryError๋ฅผ ์ก์์ฃผ๊ธฐ ์ํด
|
| 456 |
-
"""
|
| 457 |
try:
|
| 458 |
model.generate(**kwargs)
|
| 459 |
except torch.cuda.OutOfMemoryError:
|
| 460 |
-
raise RuntimeError(
|
| 461 |
-
"[OutOfMemoryError] GPU ๋ฉ๋ชจ๋ฆฌ๊ฐ ๋ถ์กฑํฉ๋๋ค. "
|
| 462 |
-
"Max New Tokens์ ์ค์ด๊ฑฐ๋, ํ๋กฌํํธ ๊ธธ์ด๋ฅผ ์ค์ฌ์ฃผ์ธ์."
|
| 463 |
-
)
|
| 464 |
finally:
|
| 465 |
-
# ์์ฑ ์๋ฃ ํ ํ๋ฒ ๋ ์บ์ ๋น์ฐ๊ธฐ
|
| 466 |
clear_cuda_cache()
|
| 467 |
|
| 468 |
-
|
| 469 |
-
|
| 470 |
-
#
|
| 471 |
-
##############################################################################
|
| 472 |
@spaces.GPU(duration=120)
|
| 473 |
def run(
|
| 474 |
message: dict,
|
|
@@ -477,57 +439,53 @@ def run(
|
|
| 477 |
max_new_tokens: int = 512,
|
| 478 |
use_web_search: bool = False,
|
| 479 |
web_search_query: str = "",
|
|
|
|
|
|
|
|
|
|
|
|
|
| 480 |
) -> Iterator[str]:
|
| 481 |
-
|
| 482 |
if not validate_media_constraints(message, history):
|
| 483 |
yield ""
|
| 484 |
return
|
| 485 |
-
|
| 486 |
-
temp_files = [] # ์์ ํ์ผ ์ถ์ ์ฉ
|
| 487 |
-
|
| 488 |
try:
|
| 489 |
-
|
| 490 |
-
|
| 491 |
-
|
| 492 |
-
|
| 493 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 494 |
|
| 495 |
if use_web_search:
|
| 496 |
user_text = message["text"]
|
| 497 |
-
ws_query = extract_keywords(user_text
|
| 498 |
if ws_query.strip():
|
| 499 |
-
logger.info(f"[
|
| 500 |
ws_result = do_web_search(ws_query)
|
| 501 |
-
combined_system_msg += f"[
|
| 502 |
-
|
| 503 |
-
|
| 504 |
-
|
| 505 |
-
|
| 506 |
-
|
| 507 |
-
|
| 508 |
-
|
| 509 |
-
|
| 510 |
-
"""
|
| 511 |
else:
|
| 512 |
-
combined_system_msg += "[
|
| 513 |
-
|
| 514 |
messages = []
|
| 515 |
if combined_system_msg.strip():
|
| 516 |
-
messages.append({
|
| 517 |
-
"role": "system",
|
| 518 |
-
"content": [{"type": "text", "text": combined_system_msg.strip()}],
|
| 519 |
-
})
|
| 520 |
-
|
| 521 |
messages.extend(process_history(history))
|
| 522 |
-
|
| 523 |
user_content, user_temp_files = process_new_user_message(message)
|
| 524 |
-
temp_files.extend(user_temp_files)
|
| 525 |
-
|
| 526 |
for item in user_content:
|
| 527 |
if item["type"] == "text" and len(item["text"]) > MAX_CONTENT_CHARS:
|
| 528 |
-
item["text"] = item["text"][:MAX_CONTENT_CHARS] + "\n...(
|
| 529 |
messages.append({"role": "user", "content": user_content})
|
| 530 |
-
|
| 531 |
inputs = processor.apply_chat_template(
|
| 532 |
messages,
|
| 533 |
add_generation_prompt=True,
|
|
@@ -535,59 +493,198 @@ def run(
|
|
| 535 |
return_dict=True,
|
| 536 |
return_tensors="pt",
|
| 537 |
).to(device=model.device, dtype=torch.bfloat16)
|
| 538 |
-
|
| 539 |
-
# ์
๋ ฅ ํ ํฐ ์ ์ ํ ์ถ๊ฐ
|
| 540 |
if inputs.input_ids.shape[1] > MAX_INPUT_LENGTH:
|
| 541 |
inputs.input_ids = inputs.input_ids[:, -MAX_INPUT_LENGTH:]
|
| 542 |
if 'attention_mask' in inputs:
|
| 543 |
inputs.attention_mask = inputs.attention_mask[:, -MAX_INPUT_LENGTH:]
|
| 544 |
-
|
| 545 |
streamer = TextIteratorStreamer(processor, timeout=30.0, skip_prompt=True, skip_special_tokens=True)
|
| 546 |
-
gen_kwargs = dict(
|
| 547 |
-
inputs,
|
| 548 |
-
streamer=streamer,
|
| 549 |
-
max_new_tokens=max_new_tokens,
|
| 550 |
-
)
|
| 551 |
-
|
| 552 |
t = Thread(target=_model_gen_with_oom_catch, kwargs=gen_kwargs)
|
| 553 |
t.start()
|
| 554 |
-
|
| 555 |
-
output = ""
|
| 556 |
for new_text in streamer:
|
| 557 |
-
|
| 558 |
-
yield
|
| 559 |
|
| 560 |
except Exception as e:
|
| 561 |
-
logger.error(f"
|
| 562 |
yield f"์ฃ์กํฉ๋๋ค. ์ค๋ฅ๊ฐ ๋ฐ์ํ์ต๋๋ค: {str(e)}"
|
| 563 |
-
|
| 564 |
finally:
|
| 565 |
-
|
| 566 |
-
for temp_file in temp_files:
|
| 567 |
try:
|
| 568 |
-
if os.path.exists(
|
| 569 |
-
os.unlink(
|
| 570 |
-
logger.info(f"
|
| 571 |
-
except Exception as
|
| 572 |
-
logger.warning(f"
|
| 573 |
-
|
| 574 |
-
# ๋ช
์์ ๋ฉ๋ชจ๋ฆฌ ์ ๋ฆฌ
|
| 575 |
try:
|
| 576 |
del inputs, streamer
|
| 577 |
-
except:
|
| 578 |
pass
|
| 579 |
-
|
| 580 |
clear_cuda_cache()
|
| 581 |
|
| 582 |
-
|
| 583 |
-
|
| 584 |
-
|
| 585 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 586 |
examples = [
|
| 587 |
-
# ----- ๊ธฐ์กด ์ด๋ฏธ์ง/๋น๋์ค ์์ 12๊ฐ -----
|
| 588 |
[
|
| 589 |
{
|
| 590 |
-
"text": "
|
| 591 |
"files": [
|
| 592 |
"assets/additional-examples/before.pdf",
|
| 593 |
"assets/additional-examples/after.pdf",
|
|
@@ -596,43 +693,46 @@ examples = [
|
|
| 596 |
],
|
| 597 |
[
|
| 598 |
{
|
| 599 |
-
"text": "
|
| 600 |
"files": ["assets/additional-examples/sample-csv.csv"],
|
| 601 |
}
|
| 602 |
],
|
| 603 |
[
|
| 604 |
{
|
| 605 |
-
"text": "
|
| 606 |
"files": ["assets/additional-examples/tmp.mp4"],
|
| 607 |
}
|
| 608 |
],
|
| 609 |
[
|
| 610 |
{
|
| 611 |
-
"text": "
|
| 612 |
"files": ["assets/additional-examples/maz.jpg"],
|
| 613 |
}
|
| 614 |
],
|
| 615 |
[
|
| 616 |
{
|
| 617 |
-
"text": "
|
| 618 |
-
"files": [
|
|
|
|
|
|
|
|
|
|
| 619 |
}
|
| 620 |
],
|
| 621 |
[
|
| 622 |
{
|
| 623 |
-
"text": "
|
| 624 |
"files": ["assets/additional-examples/4.png"],
|
| 625 |
}
|
| 626 |
],
|
| 627 |
[
|
| 628 |
{
|
| 629 |
-
"text": "
|
| 630 |
"files": ["assets/additional-examples/2.png"],
|
| 631 |
}
|
| 632 |
],
|
| 633 |
[
|
| 634 |
{
|
| 635 |
-
"text": "
|
| 636 |
"files": [
|
| 637 |
"assets/sample-images/09-1.png",
|
| 638 |
"assets/sample-images/09-2.png",
|
|
@@ -644,224 +744,176 @@ examples = [
|
|
| 644 |
],
|
| 645 |
[
|
| 646 |
{
|
| 647 |
-
"text": "
|
| 648 |
"files": ["assets/additional-examples/barchart.png"],
|
| 649 |
}
|
| 650 |
],
|
| 651 |
[
|
| 652 |
{
|
| 653 |
-
"text": "
|
| 654 |
"files": ["assets/additional-examples/3.png"],
|
| 655 |
}
|
| 656 |
],
|
| 657 |
[
|
| 658 |
{
|
| 659 |
-
"text": "
|
| 660 |
"files": ["assets/sample-images/02.png"],
|
| 661 |
}
|
| 662 |
],
|
| 663 |
[
|
| 664 |
{
|
| 665 |
-
"text": "
|
| 666 |
"files": ["assets/sample-images/03.png"],
|
| 667 |
}
|
| 668 |
],
|
| 669 |
-
# ----- ์๋กญ๊ฒ ์ถ๊ฐํ AI ๋ฐ์ดํ
์๋๋ฆฌ์ค ์์ 6๊ฐ -----
|
| 670 |
[
|
| 671 |
{
|
| 672 |
-
"text": "
|
| 673 |
}
|
| 674 |
],
|
| 675 |
[
|
| 676 |
{
|
| 677 |
-
"text": "
|
| 678 |
}
|
| 679 |
],
|
| 680 |
[
|
| 681 |
{
|
| 682 |
-
"text": "
|
| 683 |
}
|
| 684 |
],
|
| 685 |
[
|
| 686 |
{
|
| 687 |
-
"text": "
|
| 688 |
}
|
| 689 |
],
|
| 690 |
[
|
| 691 |
{
|
| 692 |
-
"text": "
|
| 693 |
}
|
| 694 |
],
|
| 695 |
[
|
| 696 |
{
|
| 697 |
-
"text": "
|
| 698 |
}
|
| 699 |
],
|
| 700 |
]
|
| 701 |
|
| 702 |
-
|
| 703 |
-
# Gradio UI (Blocks) ๊ตฌ์ฑ
|
| 704 |
-
|
|
|
|
|
|
|
| 705 |
css = """
|
| 706 |
-
/* 1) UI๋ฅผ ์ฒ์๋ถํฐ ๊ฐ์ฅ ๋๊ฒ (width 100%) ๊ณ ์ ํ์ฌ ํ์ */
|
| 707 |
.gradio-container {
|
| 708 |
-
background: rgba(255, 255, 255, 0.7);
|
| 709 |
padding: 30px 40px;
|
| 710 |
-
margin: 20px auto;
|
| 711 |
width: 100% !important;
|
| 712 |
-
max-width: none !important;
|
| 713 |
-
}
|
| 714 |
-
.fillable {
|
| 715 |
-
width: 100% !important;
|
| 716 |
-
max-width: 100% !important;
|
| 717 |
-
}
|
| 718 |
-
/* 2) ๋ฐฐ๊ฒฝ์ ์์ ํ ํฌ๋ช
ํ๊ฒ ๋ณ๊ฒฝ */
|
| 719 |
-
body {
|
| 720 |
-
background: transparent; /* ์์ ํฌ๋ช
๋ฐฐ๊ฒฝ */
|
| 721 |
-
margin: 0;
|
| 722 |
-
padding: 0;
|
| 723 |
-
font-family: 'Helvetica Neue', Helvetica, Arial, sans-serif;
|
| 724 |
-
color: #333;
|
| 725 |
-
}
|
| 726 |
-
/* ๋ฒํผ ์์ ์์ ํ ์ ๊ฑฐํ๊ณ ํฌ๋ช
ํ๊ฒ */
|
| 727 |
-
button, .btn {
|
| 728 |
-
background: transparent !important; /* ์์ ์์ ํ ์ ๊ฑฐ */
|
| 729 |
-
border: 1px solid #ddd; /* ๊ฒฝ๊ณ์ ๋ง ์ด์ง ์ถ๊ฐ */
|
| 730 |
-
color: #333;
|
| 731 |
-
padding: 12px 24px;
|
| 732 |
-
text-transform: uppercase;
|
| 733 |
-
font-weight: bold;
|
| 734 |
-
letter-spacing: 1px;
|
| 735 |
-
cursor: pointer;
|
| 736 |
-
}
|
| 737 |
-
button:hover, .btn:hover {
|
| 738 |
-
background: rgba(0, 0, 0, 0.05) !important; /* ํธ๋ฒ ์ ์์ฃผ ์ด์ง ์ด๋ก๊ฒ๋ง */
|
| 739 |
-
}
|
| 740 |
-
|
| 741 |
-
/* examples ๊ด๋ จ ๋ชจ๋ ์์ ์ ๊ฑฐ */
|
| 742 |
-
#examples_container, .examples-container {
|
| 743 |
-
margin: auto;
|
| 744 |
-
width: 90%;
|
| 745 |
-
background: transparent !important;
|
| 746 |
-
}
|
| 747 |
-
#examples_row, .examples-row {
|
| 748 |
-
justify-content: center;
|
| 749 |
-
background: transparent !important;
|
| 750 |
-
}
|
| 751 |
-
|
| 752 |
-
/* examples ๋ฒํผ ๋ด๋ถ์ ๋ชจ๋ ์์ ๏ฟฝ๏ฟฝ๏ฟฝ๊ฑฐ */
|
| 753 |
-
.gr-samples-table button,
|
| 754 |
-
.gr-samples-table .gr-button,
|
| 755 |
-
.gr-samples-table .gr-sample-btn,
|
| 756 |
-
.gr-examples button,
|
| 757 |
-
.gr-examples .gr-button,
|
| 758 |
-
.gr-examples .gr-sample-btn,
|
| 759 |
-
.examples button,
|
| 760 |
-
.examples .gr-button,
|
| 761 |
-
.examples .gr-sample-btn {
|
| 762 |
-
background: transparent !important;
|
| 763 |
-
border: 1px solid #ddd;
|
| 764 |
-
color: #333;
|
| 765 |
-
}
|
| 766 |
-
|
| 767 |
-
/* examples ๋ฒํผ ํธ๋ฒ ์์๋ ์์ ์๊ฒ */
|
| 768 |
-
.gr-samples-table button:hover,
|
| 769 |
-
.gr-samples-table .gr-button:hover,
|
| 770 |
-
.gr-samples-table .gr-sample-btn:hover,
|
| 771 |
-
.gr-examples button:hover,
|
| 772 |
-
.gr-examples .gr-button:hover,
|
| 773 |
-
.gr-examples .gr-sample-btn:hover,
|
| 774 |
-
.examples button:hover,
|
| 775 |
-
.examples .gr-button:hover,
|
| 776 |
-
.examples .gr-sample-btn:hover {
|
| 777 |
-
background: rgba(0, 0, 0, 0.05) !important;
|
| 778 |
-
}
|
| 779 |
-
|
| 780 |
-
/* ์ฑํ
์ธํฐํ์ด์ค ์์๋ค๋ ํฌ๋ช
ํ๊ฒ */
|
| 781 |
-
.chatbox, .chatbot, .message {
|
| 782 |
-
background: transparent !important;
|
| 783 |
-
}
|
| 784 |
-
|
| 785 |
-
/* ์
๋ ฅ์ฐฝ ํฌ๋ช
๋ ์กฐ์ */
|
| 786 |
-
.multimodal-textbox, textarea, input {
|
| 787 |
-
background: rgba(255, 255, 255, 0.5) !important;
|
| 788 |
-
}
|
| 789 |
-
|
| 790 |
-
/* ๋ชจ๋ ์ปจํ
์ด๋ ์์์ ๋ฐฐ๊ฒฝ์ ์ ๊ฑฐ */
|
| 791 |
-
.container, .wrap, .box, .panel, .gr-panel {
|
| 792 |
-
background: transparent !important;
|
| 793 |
-
}
|
| 794 |
-
|
| 795 |
-
/* ์์ ์น์
์ ๋ชจ๋ ์์์์ ๋ฐฐ๊ฒฝ์ ์ ๊ฑฐ */
|
| 796 |
-
.gr-examples-container, .gr-examples, .gr-sample, .gr-sample-row, .gr-sample-cell {
|
| 797 |
-
background: transparent !important;
|
| 798 |
}
|
| 799 |
"""
|
| 800 |
-
|
| 801 |
title_html = """
|
| 802 |
<h1 align="center" style="margin-bottom: 0.2em; font-size: 1.6em;"> ๐ HeartSync ๐ </h1>
|
| 803 |
<p align="center" style="font-size:1.1em; color:#555;">
|
| 804 |
-
โ
|
| 805 |
</p>
|
| 806 |
"""
|
| 807 |
|
| 808 |
with gr.Blocks(css=css, title="HeartSync") as demo:
|
| 809 |
gr.Markdown(title_html)
|
| 810 |
-
|
| 811 |
-
#
|
| 812 |
-
|
| 813 |
-
label="
|
| 814 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 815 |
)
|
| 816 |
-
|
| 817 |
-
|
| 818 |
-
|
|
|
|
|
|
|
|
|
|
| 819 |
lines=3,
|
| 820 |
-
value=
|
| 821 |
-
|
| 822 |
-
|
| 823 |
-
|
| 824 |
-
|
| 825 |
-
|
| 826 |
-
|
| 827 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 828 |
)
|
| 829 |
-
|
| 830 |
max_tokens_slider = gr.Slider(
|
| 831 |
-
label="
|
| 832 |
-
minimum=100,
|
| 833 |
-
|
| 834 |
-
step=50,
|
| 835 |
-
value=1000,
|
| 836 |
-
visible=False # ์จ๊น
|
| 837 |
)
|
| 838 |
-
|
| 839 |
web_search_text = gr.Textbox(
|
| 840 |
lines=1,
|
| 841 |
-
label="
|
| 842 |
-
placeholder="
|
| 843 |
-
visible=False
|
| 844 |
)
|
| 845 |
-
|
| 846 |
-
#
|
| 847 |
chat = gr.ChatInterface(
|
| 848 |
-
fn=
|
| 849 |
type="messages",
|
| 850 |
chatbot=gr.Chatbot(type="messages", scale=1, allow_tags=["image"]),
|
| 851 |
textbox=gr.MultimodalTextbox(
|
| 852 |
-
file_types=[
|
| 853 |
-
".webp", ".png", ".jpg", ".jpeg", ".gif",
|
| 854 |
-
".mp4", ".csv", ".txt", ".pdf"
|
| 855 |
-
],
|
| 856 |
file_count="multiple",
|
| 857 |
autofocus=True
|
| 858 |
),
|
| 859 |
multimodal=True,
|
| 860 |
additional_inputs=[
|
| 861 |
-
|
| 862 |
max_tokens_slider,
|
| 863 |
web_search_checkbox,
|
| 864 |
web_search_text,
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 865 |
],
|
| 866 |
stop_btn=False,
|
| 867 |
title='<a href="https://discord.gg/openfreeai" target="_blank">https://discord.gg/openfreeai</a>',
|
|
@@ -872,11 +924,10 @@ with gr.Blocks(css=css, title="HeartSync") as demo:
|
|
| 872 |
delete_cache=(1800, 1800),
|
| 873 |
)
|
| 874 |
|
| 875 |
-
|
| 876 |
with gr.Row(elem_id="examples_row"):
|
| 877 |
with gr.Column(scale=12, elem_id="examples_container"):
|
| 878 |
-
gr.Markdown("###
|
| 879 |
-
|
| 880 |
if __name__ == "__main__":
|
| 881 |
-
|
| 882 |
-
demo.launch()
|
|
|
|
| 9 |
import json
|
| 10 |
import requests
|
| 11 |
import cv2
|
| 12 |
+
import base64
|
| 13 |
+
import logging
|
| 14 |
+
import time
|
| 15 |
+
from urllib.parse import quote # URL ์ธ์ฝ๋ฉ์ ์ํด ์ถ๊ฐ
|
| 16 |
+
|
| 17 |
import gradio as gr
|
| 18 |
import spaces
|
| 19 |
import torch
|
|
|
|
| 21 |
from PIL import Image
|
| 22 |
from transformers import AutoProcessor, Gemma3ForConditionalGeneration, TextIteratorStreamer
|
| 23 |
|
| 24 |
+
# CSV/TXT/PDF ๋ถ์
|
| 25 |
import pandas as pd
|
|
|
|
| 26 |
import PyPDF2
|
| 27 |
|
| 28 |
+
# =============================================================================
|
| 29 |
+
# (์ ๊ท) ์ด๋ฏธ์ง API ๊ด๋ จ ํจ์๋ค
|
| 30 |
+
# =============================================================================
|
| 31 |
+
from gradio_client import Client
|
| 32 |
+
|
| 33 |
+
API_URL = "http://211.233.58.201:7896"
|
| 34 |
+
|
| 35 |
+
logging.basicConfig(
|
| 36 |
+
level=logging.DEBUG,
|
| 37 |
+
format='%(asctime)s - %(levelname)s - %(message)s'
|
| 38 |
+
)
|
| 39 |
+
|
| 40 |
+
def test_api_connection() -> str:
|
| 41 |
+
"""API ์๋ฒ ์ฐ๊ฒฐ ํ
์คํธ"""
|
| 42 |
+
try:
|
| 43 |
+
client = Client(API_URL)
|
| 44 |
+
return "API ์ฐ๊ฒฐ ์ฑ๊ณต: ์ ์ ์๋ ์ค"
|
| 45 |
+
except Exception as e:
|
| 46 |
+
logging.error(f"API ์ฐ๊ฒฐ ํ
์คํธ ์คํจ: {e}")
|
| 47 |
+
return f"API ์ฐ๊ฒฐ ์คํจ: {e}"
|
| 48 |
+
|
| 49 |
+
def generate_image(prompt: str, width: float, height: float, guidance: float, inference_steps: float, seed: float):
|
| 50 |
+
"""์ด๋ฏธ์ง ์์ฑ ํจ์ (๋ฐํ ํ์์ ์ ์ฐํ๊ฒ ๋์)"""
|
| 51 |
+
if not prompt:
|
| 52 |
+
return None, "์ค๋ฅ: ํ๋กฌํํธ๊ฐ ํ์ํฉ๋๋ค."
|
| 53 |
+
try:
|
| 54 |
+
logging.info(f"ํ๋กฌํํธ๋ฅผ ์ฌ์ฉํ์ฌ ์ด๋ฏธ์ง ์์ฑ API ํธ์ถ: {prompt}")
|
| 55 |
+
|
| 56 |
+
client = Client(API_URL)
|
| 57 |
+
result = client.predict(
|
| 58 |
+
prompt=prompt,
|
| 59 |
+
width=int(width),
|
| 60 |
+
height=int(height),
|
| 61 |
+
guidance=float(guidance),
|
| 62 |
+
inference_steps=int(inference_steps),
|
| 63 |
+
seed=int(seed),
|
| 64 |
+
do_img2img=False,
|
| 65 |
+
init_image=None,
|
| 66 |
+
image2image_strength=0.8,
|
| 67 |
+
resize_img=True,
|
| 68 |
+
api_name="/generate_image"
|
| 69 |
+
)
|
| 70 |
+
|
| 71 |
+
logging.info(f"์ด๋ฏธ์ง ์์ฑ ๊ฒฐ๊ณผ: {type(result)}, ๊ธธ์ด: {len(result) if isinstance(result, (list, tuple)) else '์ ์ ์์'}")
|
| 72 |
+
|
| 73 |
+
# ๊ฒฐ๊ณผ๊ฐ ํํ์ด๋ ๋ฆฌ์คํธ ํํ๋ก ๋ฐํ๋๋ ๊ฒฝ์ฐ ์ฒ๋ฆฌ
|
| 74 |
+
if isinstance(result, (list, tuple)) and len(result) > 0:
|
| 75 |
+
image_data = result[0] # ์ฒซ ๋ฒ์งธ ์์๊ฐ ์ด๋ฏธ์ง ๋ฐ์ดํฐ
|
| 76 |
+
seed_info = result[1] if len(result) > 1 else "์ ์ ์๋ ์๋"
|
| 77 |
+
return image_data, seed_info
|
| 78 |
+
else:
|
| 79 |
+
# ๋ค๋ฅธ ํํ๋ก ๋ฐํ๋ ๊ฒฝ์ฐ (๋จ์ผ ๊ฐ์ธ ๊ฒฝ์ฐ)
|
| 80 |
+
return result, "์ ์ ์๋ ์๋"
|
| 81 |
+
|
| 82 |
+
except Exception as e:
|
| 83 |
+
logging.error(f"์ด๋ฏธ์ง ์์ฑ ์คํจ: {str(e)}")
|
| 84 |
+
return None, f"์ค๋ฅ: {str(e)}"
|
| 85 |
+
|
| 86 |
+
# Base64 ํจ๋ฉ ์์ ํจ์
|
| 87 |
+
def fix_base64_padding(data):
|
| 88 |
+
"""Base64 ๋ฌธ์์ด์ ํจ๋ฉ์ ์์ ํฉ๋๋ค."""
|
| 89 |
+
if isinstance(data, bytes):
|
| 90 |
+
data = data.decode('utf-8')
|
| 91 |
+
|
| 92 |
+
# base64,๋ก ์์ํ๋ ๋ถ๋ถ ์ ๊ฑฐ
|
| 93 |
+
if "base64," in data:
|
| 94 |
+
data = data.split("base64,", 1)[1]
|
| 95 |
+
|
| 96 |
+
# ํจ๋ฉ ๋ฌธ์ ์ถ๊ฐ (4์ ๋ฐฐ์ ๊ธธ์ด๊ฐ ๋๋๋ก)
|
| 97 |
+
missing_padding = len(data) % 4
|
| 98 |
+
if missing_padding:
|
| 99 |
+
data += '=' * (4 - missing_padding)
|
| 100 |
+
|
| 101 |
+
return data
|
| 102 |
+
|
| 103 |
+
# =============================================================================
|
| 104 |
+
# ๋ฉ๋ชจ๋ฆฌ ์ ๋ฆฌ ํจ์
|
| 105 |
+
# =============================================================================
|
| 106 |
def clear_cuda_cache():
|
| 107 |
"""CUDA ์บ์๋ฅผ ๋ช
์์ ์ผ๋ก ๋น์๋๋ค."""
|
| 108 |
if torch.cuda.is_available():
|
| 109 |
torch.cuda.empty_cache()
|
| 110 |
gc.collect()
|
| 111 |
|
| 112 |
+
# =============================================================================
|
| 113 |
+
# SerpHouse ๊ด๋ จ ํจ์
|
| 114 |
+
# =============================================================================
|
| 115 |
SERPHOUSE_API_KEY = os.getenv("SERPHOUSE_API_KEY", "")
|
| 116 |
|
|
|
|
|
|
|
|
|
|
| 117 |
def extract_keywords(text: str, top_k: int = 5) -> str:
|
| 118 |
+
"""๋จ์ ํค์๋ ์ถ์ถ: ํ๊ธ, ์์ด, ์ซ์, ๊ณต๋ฐฑ๋ง ๋จ๊น"""
|
|
|
|
|
|
|
|
|
|
|
|
|
| 119 |
text = re.sub(r"[^a-zA-Z0-9๊ฐ-ํฃ\s]", "", text)
|
| 120 |
tokens = text.split()
|
| 121 |
+
return " ".join(tokens[:top_k])
|
|
|
|
| 122 |
|
|
|
|
|
|
|
|
|
|
|
|
|
| 123 |
def do_web_search(query: str) -> str:
|
| 124 |
+
"""SerpHouse LIVE API ํธ์ถํ์ฌ ๊ฒ์ ๊ฒฐ๊ณผ ๋งํฌ๋ค์ด ๋ฐํ"""
|
|
|
|
|
|
|
|
|
|
| 125 |
try:
|
| 126 |
url = "https://api.serphouse.com/serp/live"
|
|
|
|
|
|
|
| 127 |
params = {
|
| 128 |
"q": query,
|
| 129 |
"domain": "google.com",
|
| 130 |
+
"serp_type": "web",
|
| 131 |
"device": "desktop",
|
| 132 |
"lang": "en",
|
| 133 |
+
"num": "20"
|
| 134 |
}
|
| 135 |
+
headers = {"Authorization": f"Bearer {SERPHOUSE_API_KEY}"}
|
|
|
|
|
|
|
|
|
|
|
|
|
| 136 |
logger.info(f"SerpHouse API ํธ์ถ ์ค... ๊ฒ์์ด: {query}")
|
|
|
|
|
|
|
|
|
|
| 137 |
response = requests.get(url, headers=headers, params=params, timeout=60)
|
| 138 |
response.raise_for_status()
|
|
|
|
|
|
|
| 139 |
data = response.json()
|
|
|
|
|
|
|
| 140 |
results = data.get("results", {})
|
| 141 |
organic = None
|
|
|
|
|
|
|
| 142 |
if isinstance(results, dict) and "organic" in results:
|
| 143 |
organic = results["organic"]
|
|
|
|
|
|
|
| 144 |
elif isinstance(results, dict) and "results" in results:
|
| 145 |
if isinstance(results["results"], dict) and "organic" in results["results"]:
|
| 146 |
organic = results["results"]["organic"]
|
|
|
|
|
|
|
| 147 |
elif "organic" in data:
|
| 148 |
organic = data["organic"]
|
|
|
|
| 149 |
if not organic:
|
| 150 |
logger.warning("์๋ต์์ organic ๊ฒฐ๊ณผ๋ฅผ ์ฐพ์ ์ ์์ต๋๋ค.")
|
| 151 |
+
return "์น ๊ฒ์ ๊ฒฐ๊ณผ๊ฐ ์๊ฑฐ๋ API ์๋ต ๊ตฌ์กฐ๊ฐ ์์๊ณผ ๋ค๋ฆ
๋๋ค."
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 152 |
max_results = min(20, len(organic))
|
| 153 |
limited_organic = organic[:max_results]
|
|
|
|
|
|
|
| 154 |
summary_lines = []
|
| 155 |
for idx, item in enumerate(limited_organic, start=1):
|
| 156 |
+
title = item.get("title", "์ ๋ชฉ ์์")
|
| 157 |
link = item.get("link", "#")
|
| 158 |
+
snippet = item.get("snippet", "์ค๋ช
์์")
|
| 159 |
displayed_link = item.get("displayed_link", link)
|
|
|
|
|
|
|
| 160 |
summary_lines.append(
|
| 161 |
+
f"### ๊ฒฐ๊ณผ {idx}: {title}\n\n"
|
| 162 |
f"{snippet}\n\n"
|
| 163 |
f"**์ถ์ฒ**: [{displayed_link}]({link})\n\n"
|
| 164 |
f"---\n"
|
| 165 |
)
|
|
|
|
|
|
|
| 166 |
instructions = """
|
| 167 |
# ์น ๊ฒ์ ๊ฒฐ๊ณผ
|
| 168 |
์๋๋ ๊ฒ์ ๊ฒฐ๊ณผ์
๋๋ค. ์ง๋ฌธ์ ๋ต๋ณํ ๋ ์ด ์ ๋ณด๋ฅผ ํ์ฉํ์ธ์:
|
| 169 |
+
1. ๊ฐ ๊ฒฐ๊ณผ์ ์ ๋ชฉ, ๋ด์ฉ, ์ถ์ฒ ๋งํฌ๋ฅผ ์ฐธ๊ณ ํ์ธ์.
|
| 170 |
+
2. ๋ต๋ณ์ ๊ด๋ จ ์ ๋ณด์ ์ถ์ฒ๋ฅผ ๋ช
์์ ์ผ๋ก ์ธ์ฉํ์ธ์ (์: "[์ถ์ฒ ์ ๋ชฉ](๋งํฌ)").
|
| 171 |
+
3. ์๋ต์ ์ค์ ์ถ์ฒ ๋งํฌ๋ฅผ ํฌํจํ์ธ์.
|
| 172 |
+
4. ์ฌ๋ฌ ์ถ์ฒ์ ์ ๋ณด๋ฅผ ์ข
ํฉํ์ฌ ๋ต๋ณํ์ธ์.
|
| 173 |
+
5. ๋ง์ง๋ง์ "์ฐธ๊ณ ์๋ฃ:" ์น์
์ ์ถ๊ฐํ๊ณ ์ฃผ์ ์ถ์ฒ ๋งํฌ๋ฅผ ๋์ดํ์ธ์.
|
| 174 |
"""
|
| 175 |
+
return instructions + "\n".join(summary_lines)
|
|
|
|
|
|
|
|
|
|
|
|
|
| 176 |
except Exception as e:
|
| 177 |
+
logger.error(f"์น ๊ฒ์ ์คํจ: {e}")
|
| 178 |
+
return f"์น ๊ฒ์ ์คํจ: {str(e)}"
|
|
|
|
| 179 |
|
| 180 |
+
# =============================================================================
|
| 181 |
+
# ๋ชจ๋ธ ๋ฐ ํ๋ก์ธ์ ๋ก๋ฉ
|
| 182 |
+
# =============================================================================
|
| 183 |
MAX_CONTENT_CHARS = 2000
|
| 184 |
+
MAX_INPUT_LENGTH = 2096
|
| 185 |
model_id = os.getenv("MODEL_ID", "VIDraft/Gemma-3-R1984-4B")
|
|
|
|
| 186 |
processor = AutoProcessor.from_pretrained(model_id, padding_side="left")
|
| 187 |
model = Gemma3ForConditionalGeneration.from_pretrained(
|
| 188 |
model_id,
|
| 189 |
device_map="auto",
|
| 190 |
torch_dtype=torch.bfloat16,
|
| 191 |
+
attn_implementation="eager"
|
| 192 |
)
|
| 193 |
MAX_NUM_IMAGES = int(os.getenv("MAX_NUM_IMAGES", "5"))
|
| 194 |
|
| 195 |
+
# =============================================================================
|
| 196 |
+
# CSV, TXT, PDF ๋ถ์ ํจ์๋ค
|
| 197 |
+
# =============================================================================
|
|
|
|
| 198 |
def analyze_csv_file(path: str) -> str:
|
|
|
|
|
|
|
|
|
|
| 199 |
try:
|
| 200 |
df = pd.read_csv(path)
|
| 201 |
if df.shape[0] > 50 or df.shape[1] > 10:
|
| 202 |
df = df.iloc[:50, :10]
|
| 203 |
df_str = df.to_string()
|
| 204 |
if len(df_str) > MAX_CONTENT_CHARS:
|
| 205 |
+
df_str = df_str[:MAX_CONTENT_CHARS] + "\n...(์ผ๋ถ ์๋ต)..."
|
| 206 |
+
return f"**[CSV ํ์ผ: {os.path.basename(path)}]**\n\n{df_str}"
|
| 207 |
except Exception as e:
|
| 208 |
+
return f"CSV ํ์ผ ์ฝ๊ธฐ ์คํจ ({os.path.basename(path)}): {str(e)}"
|
|
|
|
| 209 |
|
| 210 |
def analyze_txt_file(path: str) -> str:
|
|
|
|
|
|
|
|
|
|
| 211 |
try:
|
| 212 |
with open(path, "r", encoding="utf-8") as f:
|
| 213 |
text = f.read()
|
| 214 |
if len(text) > MAX_CONTENT_CHARS:
|
| 215 |
+
text = text[:MAX_CONTENT_CHARS] + "\n...(์ผ๋ถ ์๋ต)..."
|
| 216 |
+
return f"**[TXT ํ์ผ: {os.path.basename(path)}]**\n\n{text}"
|
| 217 |
except Exception as e:
|
| 218 |
+
return f"TXT ํ์ผ ์ฝ๊ธฐ ์คํจ ({os.path.basename(path)}): {str(e)}"
|
|
|
|
| 219 |
|
| 220 |
def pdf_to_markdown(pdf_path: str) -> str:
|
|
|
|
|
|
|
|
|
|
| 221 |
text_chunks = []
|
| 222 |
try:
|
| 223 |
with open(pdf_path, "rb") as f:
|
| 224 |
reader = PyPDF2.PdfReader(f)
|
| 225 |
max_pages = min(5, len(reader.pages))
|
| 226 |
for page_num in range(max_pages):
|
| 227 |
+
page_text = reader.pages[page_num].extract_text() or ""
|
|
|
|
| 228 |
page_text = page_text.strip()
|
| 229 |
if page_text:
|
| 230 |
if len(page_text) > MAX_CONTENT_CHARS // max_pages:
|
| 231 |
+
page_text = page_text[:MAX_CONTENT_CHARS // max_pages] + "...(์ผ๋ถ ์๋ต)"
|
| 232 |
+
text_chunks.append(f"## ํ์ด์ง {page_num+1}\n\n{page_text}\n")
|
| 233 |
if len(reader.pages) > max_pages:
|
| 234 |
+
text_chunks.append(f"\n...(์ ์ฒด {len(reader.pages)}ํ์ด์ง ์ค {max_pages}ํ์ด์ง๋ง ํ์)...")
|
| 235 |
except Exception as e:
|
| 236 |
+
return f"PDF ํ์ผ ์ฝ๊ธฐ ์คํจ ({os.path.basename(pdf_path)}): {str(e)}"
|
|
|
|
| 237 |
full_text = "\n".join(text_chunks)
|
| 238 |
if len(full_text) > MAX_CONTENT_CHARS:
|
| 239 |
+
full_text = full_text[:MAX_CONTENT_CHARS] + "\n...(์ผ๋ถ ์๋ต)..."
|
| 240 |
+
return f"**[PDF ํ์ผ: {os.path.basename(pdf_path)}]**\n\n{full_text}"
|
| 241 |
|
| 242 |
+
# =============================================================================
|
| 243 |
+
# ์ด๋ฏธ์ง/๋น๋์ค ํ์ผ ์ ํ ๊ฒ์ฌ
|
| 244 |
+
# =============================================================================
|
|
|
|
|
|
|
|
|
|
| 245 |
def count_files_in_new_message(paths: list[str]) -> tuple[int, int]:
|
| 246 |
image_count = 0
|
| 247 |
video_count = 0
|
|
|
|
| 252 |
image_count += 1
|
| 253 |
return image_count, video_count
|
| 254 |
|
|
|
|
| 255 |
def count_files_in_history(history: list[dict]) -> tuple[int, int]:
|
| 256 |
image_count = 0
|
| 257 |
video_count = 0
|
|
|
|
| 267 |
image_count += 1
|
| 268 |
return image_count, video_count
|
| 269 |
|
|
|
|
| 270 |
def validate_media_constraints(message: dict, history: list[dict]) -> bool:
|
| 271 |
+
media_files = [f for f in message["files"] if re.search(r"\.(png|jpg|jpeg|gif|webp)$", f, re.IGNORECASE) or f.endswith(".mp4")]
|
|
|
|
|
|
|
|
|
|
|
|
|
| 272 |
new_image_count, new_video_count = count_files_in_new_message(media_files)
|
| 273 |
history_image_count, history_video_count = count_files_in_history(history)
|
| 274 |
image_count = history_image_count + new_image_count
|
| 275 |
video_count = history_video_count + new_video_count
|
|
|
|
| 276 |
if video_count > 1:
|
| 277 |
+
gr.Warning("๋น๋์ค ํ์ผ์ ํ๋๋ง ์ง์๋ฉ๋๋ค.")
|
| 278 |
return False
|
| 279 |
if video_count == 1:
|
| 280 |
if image_count > 0:
|
| 281 |
+
gr.Warning("์ด๋ฏธ์ง์ ๋น๋์ค๋ฅผ ํผํฉํ๋ ๊ฒ์ ํ์ฉ๋์ง ์์ต๋๋ค.")
|
| 282 |
return False
|
| 283 |
if "<image>" in message["text"]:
|
| 284 |
+
gr.Warning("<image> ํ๊ทธ์ ๋น๋์ค ํ์ผ์ ํจ๊ป ์ฌ์ฉํ ์ ์์ต๋๋ค.")
|
| 285 |
return False
|
| 286 |
if video_count == 0 and image_count > MAX_NUM_IMAGES:
|
| 287 |
+
gr.Warning(f"์ต๋ {MAX_NUM_IMAGES}์ฅ์ ์ด๋ฏธ์ง๋ฅผ ์
๋ก๋ํ ์ ์์ต๋๋ค.")
|
| 288 |
return False
|
|
|
|
| 289 |
if "<image>" in message["text"]:
|
| 290 |
image_files = [f for f in message["files"] if re.search(r"\.(png|jpg|jpeg|gif|webp)$", f, re.IGNORECASE)]
|
| 291 |
image_tag_count = message["text"].count("<image>")
|
| 292 |
if image_tag_count != len(image_files):
|
| 293 |
+
gr.Warning("ํ
์คํธ์ ์๋ <image> ํ๊ทธ์ ๊ฐ์๊ฐ ์ด๋ฏธ์ง ํ์ผ ๊ฐ์์ ์ผ์นํ์ง ์์ต๋๋ค.")
|
| 294 |
return False
|
|
|
|
| 295 |
return True
|
| 296 |
|
| 297 |
+
# =============================================================================
|
| 298 |
+
# ๋น๋์ค ์ฒ๋ฆฌ ํจ์
|
| 299 |
+
# =============================================================================
|
|
|
|
| 300 |
def downsample_video(video_path: str) -> list[tuple[Image.Image, float]]:
|
| 301 |
vidcap = cv2.VideoCapture(video_path)
|
| 302 |
fps = vidcap.get(cv2.CAP_PROP_FPS)
|
| 303 |
total_frames = int(vidcap.get(cv2.CAP_PROP_FRAME_COUNT))
|
| 304 |
frame_interval = max(int(fps), int(total_frames / 10))
|
| 305 |
frames = []
|
|
|
|
| 306 |
for i in range(0, total_frames, frame_interval):
|
| 307 |
vidcap.set(cv2.CAP_PROP_POS_FRAMES, i)
|
| 308 |
success, image = vidcap.read()
|
| 309 |
if success:
|
| 310 |
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
|
|
|
|
| 311 |
image = cv2.resize(image, (0, 0), fx=0.5, fy=0.5)
|
| 312 |
pil_image = Image.fromarray(image)
|
| 313 |
timestamp = round(i / fps, 2)
|
| 314 |
frames.append((pil_image, timestamp))
|
| 315 |
if len(frames) >= 5:
|
| 316 |
break
|
|
|
|
| 317 |
vidcap.release()
|
| 318 |
return frames
|
| 319 |
|
|
|
|
| 320 |
def process_video(video_path: str) -> tuple[list[dict], list[str]]:
|
| 321 |
content = []
|
| 322 |
+
temp_files = []
|
|
|
|
| 323 |
frames = downsample_video(video_path)
|
| 324 |
+
for pil_image, timestamp in frames:
|
|
|
|
| 325 |
with tempfile.NamedTemporaryFile(delete=False, suffix=".png") as temp_file:
|
| 326 |
pil_image.save(temp_file.name)
|
| 327 |
+
temp_files.append(temp_file.name)
|
| 328 |
+
content.append({"type": "text", "text": f"ํ๋ ์ {timestamp}:"})
|
| 329 |
content.append({"type": "image", "url": temp_file.name})
|
|
|
|
| 330 |
return content, temp_files
|
| 331 |
|
| 332 |
+
# =============================================================================
|
| 333 |
+
# interleaved <image> ์ฒ๋ฆฌ ํจ์
|
| 334 |
+
# =============================================================================
|
|
|
|
| 335 |
def process_interleaved_images(message: dict) -> list[dict]:
|
| 336 |
parts = re.split(r"(<image>)", message["text"])
|
| 337 |
content = []
|
|
|
|
|
|
|
| 338 |
image_files = [f for f in message["files"] if re.search(r"\.(png|jpg|jpeg|gif|webp)$", f, re.IGNORECASE)]
|
| 339 |
+
image_index = 0
|
| 340 |
for part in parts:
|
| 341 |
if part == "<image>" and image_index < len(image_files):
|
| 342 |
content.append({"type": "image", "url": image_files[image_index]})
|
|
|
|
| 348 |
content.append({"type": "text", "text": part})
|
| 349 |
return content
|
| 350 |
|
| 351 |
+
# =============================================================================
|
| 352 |
+
# ํ์ผ ์ฒ๋ฆฌ -> content ์์ฑ
|
| 353 |
+
# =============================================================================
|
|
|
|
| 354 |
def is_image_file(file_path: str) -> bool:
|
| 355 |
return bool(re.search(r"\.(png|jpg|jpeg|gif|webp)$", file_path, re.IGNORECASE))
|
| 356 |
|
|
|
|
| 358 |
return file_path.endswith(".mp4")
|
| 359 |
|
| 360 |
def is_document_file(file_path: str) -> bool:
|
| 361 |
+
return file_path.lower().endswith(".pdf") or file_path.lower().endswith(".csv") or file_path.lower().endswith(".txt")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 362 |
|
| 363 |
def process_new_user_message(message: dict) -> tuple[list[dict], list[str]]:
|
| 364 |
+
temp_files = []
|
|
|
|
| 365 |
if not message["files"]:
|
| 366 |
return [{"type": "text", "text": message["text"]}], temp_files
|
|
|
|
| 367 |
video_files = [f for f in message["files"] if is_video_file(f)]
|
| 368 |
image_files = [f for f in message["files"] if is_image_file(f)]
|
| 369 |
csv_files = [f for f in message["files"] if f.lower().endswith(".csv")]
|
| 370 |
txt_files = [f for f in message["files"] if f.lower().endswith(".txt")]
|
| 371 |
pdf_files = [f for f in message["files"] if f.lower().endswith(".pdf")]
|
|
|
|
| 372 |
content_list = [{"type": "text", "text": message["text"]}]
|
|
|
|
| 373 |
for csv_path in csv_files:
|
| 374 |
+
content_list.append({"type": "text", "text": analyze_csv_file(csv_path)})
|
|
|
|
|
|
|
| 375 |
for txt_path in txt_files:
|
| 376 |
+
content_list.append({"type": "text", "text": analyze_txt_file(txt_path)})
|
|
|
|
|
|
|
| 377 |
for pdf_path in pdf_files:
|
| 378 |
+
content_list.append({"type": "text", "text": pdf_to_markdown(pdf_path)})
|
|
|
|
|
|
|
| 379 |
if video_files:
|
| 380 |
video_content, video_temp_files = process_video(video_files[0])
|
| 381 |
content_list += video_content
|
| 382 |
temp_files.extend(video_temp_files)
|
| 383 |
return content_list, temp_files
|
|
|
|
| 384 |
if "<image>" in message["text"] and image_files:
|
| 385 |
interleaved_content = process_interleaved_images({"text": message["text"], "files": image_files})
|
| 386 |
if content_list and content_list[0]["type"] == "text":
|
|
|
|
| 389 |
else:
|
| 390 |
for img_path in image_files:
|
| 391 |
content_list.append({"type": "image", "url": img_path})
|
|
|
|
| 392 |
return content_list, temp_files
|
| 393 |
|
| 394 |
+
# =============================================================================
|
|
|
|
| 395 |
# history -> LLM ๋ฉ์์ง ๋ณํ
|
| 396 |
+
# =============================================================================
|
| 397 |
def process_history(history: list[dict]) -> list[dict]:
|
| 398 |
messages = []
|
| 399 |
+
current_user_content = []
|
| 400 |
for item in history:
|
| 401 |
if item["role"] == "assistant":
|
| 402 |
if current_user_content:
|
|
|
|
| 412 |
if is_image_file(file_path):
|
| 413 |
current_user_content.append({"type": "image", "url": file_path})
|
| 414 |
else:
|
| 415 |
+
current_user_content.append({"type": "text", "text": f"[ํ์ผ: {os.path.basename(file_path)}]"})
|
|
|
|
| 416 |
if current_user_content:
|
| 417 |
messages.append({"role": "user", "content": current_user_content})
|
|
|
|
| 418 |
return messages
|
| 419 |
|
| 420 |
+
# =============================================================================
|
| 421 |
+
# ๋ชจ๋ธ ์์ฑ ํจ์ (OOM ์บ์น)
|
| 422 |
+
# =============================================================================
|
|
|
|
| 423 |
def _model_gen_with_oom_catch(**kwargs):
|
|
|
|
|
|
|
|
|
|
| 424 |
try:
|
| 425 |
model.generate(**kwargs)
|
| 426 |
except torch.cuda.OutOfMemoryError:
|
| 427 |
+
raise RuntimeError("[OutOfMemoryError] GPU ๋ฉ๋ชจ๋ฆฌ๊ฐ ๋ถ์กฑํฉ๋๋ค.")
|
|
|
|
|
|
|
|
|
|
| 428 |
finally:
|
|
|
|
| 429 |
clear_cuda_cache()
|
| 430 |
|
| 431 |
+
# =============================================================================
|
| 432 |
+
# ๋ฉ์ธ ์ถ๋ก ํจ์
|
| 433 |
+
# =============================================================================
|
|
|
|
| 434 |
@spaces.GPU(duration=120)
|
| 435 |
def run(
|
| 436 |
message: dict,
|
|
|
|
| 439 |
max_new_tokens: int = 512,
|
| 440 |
use_web_search: bool = False,
|
| 441 |
web_search_query: str = "",
|
| 442 |
+
age_group: str = "20๋",
|
| 443 |
+
mbti_personality: str = "INTP",
|
| 444 |
+
sexual_openness: int = 2,
|
| 445 |
+
image_gen: bool = False # "Image Gen" ์ฒดํฌ ์ฌ๋ถ
|
| 446 |
) -> Iterator[str]:
|
|
|
|
| 447 |
if not validate_media_constraints(message, history):
|
| 448 |
yield ""
|
| 449 |
return
|
| 450 |
+
temp_files = []
|
|
|
|
|
|
|
| 451 |
try:
|
| 452 |
+
# ์์คํ
ํ๋กฌํํธ์ ํ๋ฅด์๋ ์ ๋ณด ์ถ๊ฐ
|
| 453 |
+
persona = (
|
| 454 |
+
f"{system_prompt.strip()}\n\n"
|
| 455 |
+
f"์ฑ๋ณ: ์ฌ์ฑ\n"
|
| 456 |
+
f"์ฐ๋ น๋: {age_group}\n"
|
| 457 |
+
f"MBTI ํ๋ฅด์๋: {mbti_personality}\n"
|
| 458 |
+
f"์น์์ผ ๊ฐ๋ฐฉ์ฑ (1~5): {sexual_openness}\n"
|
| 459 |
+
)
|
| 460 |
+
combined_system_msg = f"[์์คํ
ํ๋กฌํํธ]\n{persona.strip()}\n\n"
|
| 461 |
|
| 462 |
if use_web_search:
|
| 463 |
user_text = message["text"]
|
| 464 |
+
ws_query = extract_keywords(user_text)
|
| 465 |
if ws_query.strip():
|
| 466 |
+
logger.info(f"[์๋ ์น ๊ฒ์ ํค์๋] {ws_query!r}")
|
| 467 |
ws_result = do_web_search(ws_query)
|
| 468 |
+
combined_system_msg += f"[๊ฒ์ ๊ฒฐ๊ณผ (์์ 20๊ฐ ํญ๋ชฉ)]\n{ws_result}\n\n"
|
| 469 |
+
combined_system_msg += (
|
| 470 |
+
"[์ฐธ๊ณ : ์ ๊ฒ์ ๊ฒฐ๊ณผ ๋งํฌ๋ฅผ ์ถ์ฒ๋ก ์ธ์ฉํ์ฌ ๋ต๋ณ]\n"
|
| 471 |
+
"[์ค์ ์ง์์ฌํญ]\n"
|
| 472 |
+
"1. ๋ต๋ณ์ ๊ฒ์ ๊ฒฐ๊ณผ์์ ์ฐพ์ ์ ๋ณด์ ์ถ์ฒ๋ฅผ ๋ฐ๋์ ์ธ์ฉํ์ธ์.\n"
|
| 473 |
+
"2. ์ถ์ฒ ์ธ์ฉ ์ \"[์ถ์ฒ ์ ๋ชฉ](๋งํฌ)\" ํ์์ ๋งํฌ๋ค์ด ๋งํฌ๋ฅผ ์ฌ์ฉํ์ธ์.\n"
|
| 474 |
+
"3. ์ฌ๋ฌ ์ถ์ฒ์ ์ ๋ณด๋ฅผ ์ข
ํฉํ์ฌ ๋ต๋ณํ์ธ์.\n"
|
| 475 |
+
"4. ๋ต๋ณ ๋ง์ง๋ง์ \"์ฐธ๊ณ ์๋ฃ:\" ์น์
์ ์ถ๊ฐํ๊ณ ์ฌ์ฉํ ์ฃผ์ ์ถ์ฒ ๋งํฌ๋ฅผ ๋์ดํ์ธ์.\n"
|
| 476 |
+
)
|
|
|
|
| 477 |
else:
|
| 478 |
+
combined_system_msg += "[์ ํจํ ํค์๋๊ฐ ์์ด ์น ๊ฒ์์ ๊ฑด๋๋๋๋ค]\n\n"
|
|
|
|
| 479 |
messages = []
|
| 480 |
if combined_system_msg.strip():
|
| 481 |
+
messages.append({"role": "system", "content": [{"type": "text", "text": combined_system_msg.strip()}]})
|
|
|
|
|
|
|
|
|
|
|
|
|
| 482 |
messages.extend(process_history(history))
|
|
|
|
| 483 |
user_content, user_temp_files = process_new_user_message(message)
|
| 484 |
+
temp_files.extend(user_temp_files)
|
|
|
|
| 485 |
for item in user_content:
|
| 486 |
if item["type"] == "text" and len(item["text"]) > MAX_CONTENT_CHARS:
|
| 487 |
+
item["text"] = item["text"][:MAX_CONTENT_CHARS] + "\n...(์ผ๋ถ ์๋ต)..."
|
| 488 |
messages.append({"role": "user", "content": user_content})
|
|
|
|
| 489 |
inputs = processor.apply_chat_template(
|
| 490 |
messages,
|
| 491 |
add_generation_prompt=True,
|
|
|
|
| 493 |
return_dict=True,
|
| 494 |
return_tensors="pt",
|
| 495 |
).to(device=model.device, dtype=torch.bfloat16)
|
|
|
|
|
|
|
| 496 |
if inputs.input_ids.shape[1] > MAX_INPUT_LENGTH:
|
| 497 |
inputs.input_ids = inputs.input_ids[:, -MAX_INPUT_LENGTH:]
|
| 498 |
if 'attention_mask' in inputs:
|
| 499 |
inputs.attention_mask = inputs.attention_mask[:, -MAX_INPUT_LENGTH:]
|
|
|
|
| 500 |
streamer = TextIteratorStreamer(processor, timeout=30.0, skip_prompt=True, skip_special_tokens=True)
|
| 501 |
+
gen_kwargs = dict(inputs, streamer=streamer, max_new_tokens=max_new_tokens)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 502 |
t = Thread(target=_model_gen_with_oom_catch, kwargs=gen_kwargs)
|
| 503 |
t.start()
|
| 504 |
+
output_so_far = ""
|
|
|
|
| 505 |
for new_text in streamer:
|
| 506 |
+
output_so_far += new_text
|
| 507 |
+
yield output_so_far
|
| 508 |
|
| 509 |
except Exception as e:
|
| 510 |
+
logger.error(f"run ํจ์ ์๋ฌ: {str(e)}")
|
| 511 |
yield f"์ฃ์กํฉ๋๋ค. ์ค๋ฅ๊ฐ ๋ฐ์ํ์ต๋๋ค: {str(e)}"
|
|
|
|
| 512 |
finally:
|
| 513 |
+
for tmp in temp_files:
|
|
|
|
| 514 |
try:
|
| 515 |
+
if os.path.exists(tmp):
|
| 516 |
+
os.unlink(tmp)
|
| 517 |
+
logger.info(f"์์ ํ์ผ ์ญ์ ๋จ: {tmp}")
|
| 518 |
+
except Exception as ee:
|
| 519 |
+
logger.warning(f"์์ ํ์ผ {tmp} ์ญ์ ์คํจ: {ee}")
|
|
|
|
|
|
|
| 520 |
try:
|
| 521 |
del inputs, streamer
|
| 522 |
+
except Exception:
|
| 523 |
pass
|
|
|
|
| 524 |
clear_cuda_cache()
|
| 525 |
|
| 526 |
+
# ์์ ๋ ๋ชจ๋ธ ์คํ ํจ์ - ์ด๋ฏธ์ง ์์ฑ ๋ฐ ๊ฐค๋ฌ๋ฆฌ ์ถ๋ ฅ ์ฒ๋ฆฌ
|
| 527 |
+
def modified_run(message, history, system_prompt, max_new_tokens, use_web_search, web_search_query,
|
| 528 |
+
age_group, mbti_personality, sexual_openness, image_gen):
|
| 529 |
+
# ๊ฐค๋ฌ๋ฆฌ ์ด๊ธฐํ ๋ฐ ์จ๊ธฐ๊ธฐ
|
| 530 |
+
output_so_far = ""
|
| 531 |
+
gallery_update = gr.Gallery(visible=False, value=[])
|
| 532 |
+
yield output_so_far, gallery_update
|
| 533 |
+
|
| 534 |
+
# ๊ธฐ์กด run ํจ์ ๋ก์ง
|
| 535 |
+
text_generator = run(message, history, system_prompt, max_new_tokens, use_web_search,
|
| 536 |
+
web_search_query, age_group, mbti_personality, sexual_openness, image_gen)
|
| 537 |
+
|
| 538 |
+
for text_chunk in text_generator:
|
| 539 |
+
output_so_far = text_chunk
|
| 540 |
+
yield output_so_far, gallery_update
|
| 541 |
+
|
| 542 |
+
# ์ด๋ฏธ์ง ์์ฑ์ด ํ์ฑํ๋ ๊ฒฝ์ฐ ๊ฐค๋ฌ๋ฆฌ ์
๋ฐ์ดํธ
|
| 543 |
+
if image_gen and message["text"].strip():
|
| 544 |
+
try:
|
| 545 |
+
width, height = 512, 512
|
| 546 |
+
guidance, steps, seed = 7.5, 30, 42
|
| 547 |
+
|
| 548 |
+
logger.info(f"๊ฐค๋ฌ๋ฆฌ์ฉ ์ด๋ฏธ์ง ์์ฑ ํธ์ถ, ํ๋กฌํํธ: {message['text']}")
|
| 549 |
+
|
| 550 |
+
# API ํธ์ถํด์ ์ด๋ฏธ์ง ์์ฑ
|
| 551 |
+
image_result, seed_info = generate_image(
|
| 552 |
+
prompt=message["text"].strip(),
|
| 553 |
+
width=width,
|
| 554 |
+
height=height,
|
| 555 |
+
guidance=guidance,
|
| 556 |
+
inference_steps=steps,
|
| 557 |
+
seed=seed
|
| 558 |
+
)
|
| 559 |
+
|
| 560 |
+
if image_result:
|
| 561 |
+
# ์ง์ ์ด๋ฏธ์ง ๋ฐ์ดํฐ ์ฒ๋ฆฌ: base64 ๋ฌธ์์ด์ธ ๊ฒฝ์ฐ
|
| 562 |
+
if isinstance(image_result, str) and (
|
| 563 |
+
image_result.startswith('data:') or
|
| 564 |
+
len(image_result) > 100 and '/' not in image_result
|
| 565 |
+
):
|
| 566 |
+
# base64 ์ด๋ฏธ์ง ๋ฌธ์์ด์ ํ์ผ๋ก ๋ณํ
|
| 567 |
+
try:
|
| 568 |
+
# data:image ์ ๋์ฌ ์ ๊ฑฐ
|
| 569 |
+
if image_result.startswith('data:'):
|
| 570 |
+
content_type, b64data = image_result.split(';base64,')
|
| 571 |
+
else:
|
| 572 |
+
b64data = image_result
|
| 573 |
+
content_type = "image/webp" # ๊ธฐ๋ณธ๊ฐ์ผ๋ก ๊ฐ์
|
| 574 |
+
|
| 575 |
+
# base64 ๋์ฝ๋ฉ
|
| 576 |
+
image_bytes = base64.b64decode(b64data)
|
| 577 |
+
|
| 578 |
+
# ์์ ํ์ผ๋ก ์ ์ฅ
|
| 579 |
+
with tempfile.NamedTemporaryFile(delete=False, suffix=".webp") as temp_file:
|
| 580 |
+
temp_file.write(image_bytes)
|
| 581 |
+
temp_path = temp_file.name
|
| 582 |
+
|
| 583 |
+
# ๊ฐค๋ฌ๋ฆฌ ํ์ ๋ฐ ์ด๋ฏธ์ง ์ถ๊ฐ
|
| 584 |
+
gallery_update = gr.Gallery(visible=True, value=[temp_path])
|
| 585 |
+
yield output_so_far + "\n\n*์ด๋ฏธ์ง๊ฐ ์์ฑ๋์ด ์๋ ๊ฐค๋ฌ๋ฆฌ์ ํ์๋ฉ๋๋ค.*", gallery_update
|
| 586 |
+
|
| 587 |
+
except Exception as e:
|
| 588 |
+
logger.error(f"Base64 ์ด๋ฏธ์ง ์ฒ๋ฆฌ ์ค๋ฅ: {e}")
|
| 589 |
+
yield output_so_far + f"\n\n(์ด๋ฏธ์ง ์ฒ๋ฆฌ ์ค ์ค๋ฅ: {e})", gallery_update
|
| 590 |
+
|
| 591 |
+
# ํ์ผ ๊ฒฝ๋ก์ธ ๊ฒฝ์ฐ
|
| 592 |
+
elif isinstance(image_result, str) and os.path.exists(image_result):
|
| 593 |
+
# ๋ก์ปฌ ํ์ผ ๊ฒฝ๋ก๋ฅผ ๊ทธ๋๋ก ์ฌ์ฉ
|
| 594 |
+
gallery_update = gr.Gallery(visible=True, value=[image_result])
|
| 595 |
+
yield output_so_far + "\n\n*์ด๋ฏธ์ง๊ฐ ์์ฑ๋์ด ์๋ ๊ฐค๋ฌ๋ฆฌ์ ํ์๋ฉ๋๋ค.*", gallery_update
|
| 596 |
+
|
| 597 |
+
# /tmp ๊ฒฝ๋ก์ธ ๊ฒฝ์ฐ (API ์๋ฒ์๋ง ์กด์ฌํ๋ ํ์ผ)
|
| 598 |
+
elif isinstance(image_result, str) and '/tmp/' in image_result:
|
| 599 |
+
# API์์ ๋ฐํ๋ ํ์ผ ๊ฒฝ๋ก์์ ์ด๋ฏธ์ง ์ ๋ณด ์ถ์ถ
|
| 600 |
+
try:
|
| 601 |
+
# API ์๋ต์ base64 ์ธ์ฝ๋ฉ๋ ๋ฌธ์์ด๋ก ์ฒ๋ฆฌ
|
| 602 |
+
client = Client(API_URL)
|
| 603 |
+
result = client.predict(
|
| 604 |
+
prompt=message["text"].strip(),
|
| 605 |
+
api_name="/generate_base64_image" # base64 ๋ฐํ API
|
| 606 |
+
)
|
| 607 |
+
|
| 608 |
+
if isinstance(result, str) and (result.startswith('data:') or len(result) > 100):
|
| 609 |
+
# base64 ์ด๋ฏธ์ง ์ฒ๋ฆฌ
|
| 610 |
+
if result.startswith('data:'):
|
| 611 |
+
content_type, b64data = result.split(';base64,')
|
| 612 |
+
else:
|
| 613 |
+
b64data = result
|
| 614 |
+
|
| 615 |
+
# base64 ๋์ฝ๋ฉ
|
| 616 |
+
image_bytes = base64.b64decode(b64data)
|
| 617 |
+
|
| 618 |
+
# ์์ ํ์ผ๋ก ์ ์ฅ
|
| 619 |
+
with tempfile.NamedTemporaryFile(delete=False, suffix=".webp") as temp_file:
|
| 620 |
+
temp_file.write(image_bytes)
|
| 621 |
+
temp_path = temp_file.name
|
| 622 |
+
|
| 623 |
+
# ๊ฐค๋ฌ๋ฆฌ ํ์ ๋ฐ ์ด๋ฏธ์ง ์ถ๊ฐ
|
| 624 |
+
gallery_update = gr.Gallery(visible=True, value=[temp_path])
|
| 625 |
+
yield output_so_far + "\n\n*์ด๋ฏธ์ง๊ฐ ์์ฑ๋์ด ์๋ ๊ฐค๋ฌ๋ฆฌ์ ํ์๋ฉ๋๋ค.*", gallery_update
|
| 626 |
+
else:
|
| 627 |
+
yield output_so_far + "\n\n(์ด๋ฏธ์ง ์์ฑ ์คํจ: ์ฌ๋ฐ๋ฅธ ํ์์ด ์๋๋๋ค)", gallery_update
|
| 628 |
+
|
| 629 |
+
except Exception as e:
|
| 630 |
+
logger.error(f"๋์ฒด API ํธ์ถ ์ค ์ค๋ฅ: {e}")
|
| 631 |
+
yield output_so_far + f"\n\n(์ด๋ฏธ์ง ์์ฑ ์คํจ: {e})", gallery_update
|
| 632 |
+
|
| 633 |
+
# URL์ธ ๊ฒฝ์ฐ
|
| 634 |
+
elif isinstance(image_result, str) and (
|
| 635 |
+
image_result.startswith('http://') or
|
| 636 |
+
image_result.startswith('https://')
|
| 637 |
+
):
|
| 638 |
+
try:
|
| 639 |
+
# URL์์ ์ด๋ฏธ์ง ๋ค์ด๋ก๋
|
| 640 |
+
response = requests.get(image_result, timeout=10)
|
| 641 |
+
response.raise_for_status()
|
| 642 |
+
|
| 643 |
+
# ์์ ํ์ผ๋ก ์ ์ฅ
|
| 644 |
+
with tempfile.NamedTemporaryFile(delete=False, suffix=".webp") as temp_file:
|
| 645 |
+
temp_file.write(response.content)
|
| 646 |
+
temp_path = temp_file.name
|
| 647 |
+
|
| 648 |
+
# ๊ฐค๋ฌ๋ฆฌ ํ์ ๋ฐ ์ด๋ฏธ์ง ์ถ๊ฐ
|
| 649 |
+
gallery_update = gr.Gallery(visible=True, value=[temp_path])
|
| 650 |
+
yield output_so_far + "\n\n*์ด๋ฏธ์ง๊ฐ ์์ฑ๋์ด ์๋ ๊ฐค๋ฌ๋ฆฌ์ ํ์๋ฉ๋๋ค.*", gallery_update
|
| 651 |
+
|
| 652 |
+
except Exception as e:
|
| 653 |
+
logger.error(f"URL ์ด๋ฏธ์ง ๋ค์ด๋ก๋ ์ค๋ฅ: {e}")
|
| 654 |
+
yield output_so_far + f"\n\n(์ด๋ฏธ์ง ๋ค์ด๋ก๋ ์ค ์ค๋ฅ: {e})", gallery_update
|
| 655 |
+
|
| 656 |
+
# ์ด๋ฏธ์ง ๊ฐ์ฒด์ธ ๊ฒฝ์ฐ (PIL Image ๋ฑ)
|
| 657 |
+
elif hasattr(image_result, 'save'):
|
| 658 |
+
try:
|
| 659 |
+
with tempfile.NamedTemporaryFile(delete=False, suffix=".webp") as temp_file:
|
| 660 |
+
image_result.save(temp_file.name)
|
| 661 |
+
temp_path = temp_file.name
|
| 662 |
+
|
| 663 |
+
# ๊ฐค๋ฌ๋ฆฌ ํ์ ๋ฐ ์ด๋ฏธ์ง ์ถ๊ฐ
|
| 664 |
+
gallery_update = gr.Gallery(visible=True, value=[temp_path])
|
| 665 |
+
yield output_so_far + "\n\n*์ด๋ฏธ์ง๊ฐ ์์ฑ๋์ด ์๋ ๊ฐค๋ฌ๋ฆฌ์ ํ์๋ฉ๋๋ค.*", gallery_update
|
| 666 |
+
|
| 667 |
+
except Exception as e:
|
| 668 |
+
logger.error(f"์ด๋ฏธ์ง ๊ฐ์ฒด ์ ์ฅ ์ค๋ฅ: {e}")
|
| 669 |
+
yield output_so_far + f"\n\n(์ด๋ฏธ์ง ๊ฐ์ฒด ์ ์ฅ ์ค ์ค๋ฅ: {e})", gallery_update
|
| 670 |
+
|
| 671 |
+
else:
|
| 672 |
+
# ๋ค๋ฅธ ํ์์ ์ด๋ฏธ์ง ๊ฒฐ๊ณผ
|
| 673 |
+
yield output_so_far + f"\n\n(์ง์๋์ง ์๋ ์ด๋ฏธ์ง ํ์: {type(image_result)})", gallery_update
|
| 674 |
+
else:
|
| 675 |
+
yield output_so_far + f"\n\n(์ด๋ฏธ์ง ์์ฑ ์คํจ: {seed_info})", gallery_update
|
| 676 |
+
|
| 677 |
+
except Exception as e:
|
| 678 |
+
logger.error(f"๊ฐค๋ฌ๋ฆฌ์ฉ ์ด๋ฏธ์ง ์์ฑ ์ค ์ค๋ฅ: {e}")
|
| 679 |
+
yield output_so_far + f"\n\n(์ด๋ฏธ์ง ์์ฑ ์ค ์ค๋ฅ: {e})", gallery_update
|
| 680 |
+
|
| 681 |
+
# =============================================================================
|
| 682 |
+
# ์์๋ค: ๊ธฐ์กด ์ด๋ฏธ์ง/๋น๋์ค ์์ 12๊ฐ + AI ๋ฐ์ดํ
์๋๋ฆฌ์ค ์์ 6๊ฐ
|
| 683 |
+
# =============================================================================
|
| 684 |
examples = [
|
|
|
|
| 685 |
[
|
| 686 |
{
|
| 687 |
+
"text": "๋ PDF ํ์ผ์ ๋ด์ฉ์ ๋น๊ตํ์ธ์.",
|
| 688 |
"files": [
|
| 689 |
"assets/additional-examples/before.pdf",
|
| 690 |
"assets/additional-examples/after.pdf",
|
|
|
|
| 693 |
],
|
| 694 |
[
|
| 695 |
{
|
| 696 |
+
"text": "CSV ํ์ผ์ ๋ด์ฉ์ ์์ฝ ๋ฐ ๋ถ์ํ์ธ์.",
|
| 697 |
"files": ["assets/additional-examples/sample-csv.csv"],
|
| 698 |
}
|
| 699 |
],
|
| 700 |
[
|
| 701 |
{
|
| 702 |
+
"text": "์น์ ํ๊ณ ์ดํด์ฌ ๋ง์ ์ฌ์์น๊ตฌ ์ญํ ์ ๋งก์ผ์ธ์. ์ด ์์์ ์ค๋ช
ํด ์ฃผ์ธ์.",
|
| 703 |
"files": ["assets/additional-examples/tmp.mp4"],
|
| 704 |
}
|
| 705 |
],
|
| 706 |
[
|
| 707 |
{
|
| 708 |
+
"text": "ํ์ง๋ฅผ ์ค๋ช
ํ๊ณ ๊ทธ ์์ ๊ธ์จ๋ฅผ ์ฝ์ด ์ฃผ์ธ์.",
|
| 709 |
"files": ["assets/additional-examples/maz.jpg"],
|
| 710 |
}
|
| 711 |
],
|
| 712 |
[
|
| 713 |
{
|
| 714 |
+
"text": "์ ๋ ์ด๋ฏธ ์ด ๋ณด์ถฉ์ ๋ฅผ ๊ฐ์ง๊ณ ์๊ณ <image> ์ด ์ ํ๋ ๊ตฌ๋งคํ ๊ณํ์
๋๋ค. ํจ๊ป ๋ณต์ฉํ ๋ ์ฃผ์ํ ์ ์ด ์๋์?",
|
| 715 |
+
"files": [
|
| 716 |
+
"assets/additional-examples/pill1.png",
|
| 717 |
+
"assets/additional-examples/pill2.png"
|
| 718 |
+
],
|
| 719 |
}
|
| 720 |
],
|
| 721 |
[
|
| 722 |
{
|
| 723 |
+
"text": "์ด ์ ๋ถ ๋ฌธ์ ๋ฅผ ํ์ด ์ฃผ์ธ์.",
|
| 724 |
"files": ["assets/additional-examples/4.png"],
|
| 725 |
}
|
| 726 |
],
|
| 727 |
[
|
| 728 |
{
|
| 729 |
+
"text": "์ด ํฐ์ผ์ ์ธ์ ๋ฐํ๋์๊ณ , ๊ฐ๊ฒฉ์ ์ผ๋ง์ธ๊ฐ์?",
|
| 730 |
"files": ["assets/additional-examples/2.png"],
|
| 731 |
}
|
| 732 |
],
|
| 733 |
[
|
| 734 |
{
|
| 735 |
+
"text": "์ด ์ด๋ฏธ์ง๋ค์ ์์๋ฅผ ๋ฐํ์ผ๋ก ์งง์ ์ด์ผ๊ธฐ๋ฅผ ๋ง๋ค์ด ์ฃผ์ธ์.",
|
| 736 |
"files": [
|
| 737 |
"assets/sample-images/09-1.png",
|
| 738 |
"assets/sample-images/09-2.png",
|
|
|
|
| 744 |
],
|
| 745 |
[
|
| 746 |
{
|
| 747 |
+
"text": "์ด ์ด๋ฏธ์ง์ ์ผ์นํ๋ ๋ง๋ ์ฐจํธ๋ฅผ ๊ทธ๋ฆฌ๊ธฐ ์ํ matplotlib๋ฅผ ์ฌ์ฉํ๋ Python ์ฝ๋๋ฅผ ์์ฑํด ์ฃผ์ธ์.",
|
| 748 |
"files": ["assets/additional-examples/barchart.png"],
|
| 749 |
}
|
| 750 |
],
|
| 751 |
[
|
| 752 |
{
|
| 753 |
+
"text": "์ด๋ฏธ์ง์ ํ
์คํธ๋ฅผ ์ฝ๊ณ Markdown ํ์์ผ๋ก ์์ฑํด ์ฃผ์ธ์.",
|
| 754 |
"files": ["assets/additional-examples/3.png"],
|
| 755 |
}
|
| 756 |
],
|
| 757 |
[
|
| 758 |
{
|
| 759 |
+
"text": "์ด ํ์งํ์ ๋ฌด์จ ๊ธ์๊ฐ ์ฐ์ฌ ์๋์?",
|
| 760 |
"files": ["assets/sample-images/02.png"],
|
| 761 |
}
|
| 762 |
],
|
| 763 |
[
|
| 764 |
{
|
| 765 |
+
"text": "๋ ์ด๋ฏธ์ง๋ฅผ ๋น๊ตํ๊ณ ์ ์ฌ์ ๊ณผ ์ฐจ์ด์ ์ ์ค๋ช
ํด ์ฃผ์ธ์.",
|
| 766 |
"files": ["assets/sample-images/03.png"],
|
| 767 |
}
|
| 768 |
],
|
|
|
|
| 769 |
[
|
| 770 |
{
|
| 771 |
+
"text": "๋กคํ๋ ์ด ํด๋ด
์๋ค. ๋น์ ์ ์ ์ ๋ ์์๊ฐ๊ณ ์ถ์ ์๋ก์ด ์จ๋ผ์ธ ๋ฐ์ดํธ ์๋์
๋๋ค. ๋ค์ ํ๊ณ ๋ฐฐ๋ ค ๊น์ ๋ฐฉ์์ผ๋ก ์๊ธฐ ์๊ฐ๋ฅผ ํด์ฃผ์ธ์!",
|
| 772 |
}
|
| 773 |
],
|
| 774 |
[
|
| 775 |
{
|
| 776 |
+
"text": "ํด๋ณ์ ๊ฑท๋ ๋ ๋ฒ์งธ ๋ฐ์ดํธ ์ค์
๋๋ค. ์ฅ๋์ค๋ฌ์ด ๋ํ์ ๋ถ๋๋ฌ์ด ํ๋ฌํ
์ผ๋ก ์ฅ๋ฉด์ ์ด์ด๋๊ฐ ์ฃผ์ธ์.",
|
| 777 |
}
|
| 778 |
],
|
| 779 |
[
|
| 780 |
{
|
| 781 |
+
"text": "์ข์ํ๋ ์ฌ๋์๊ฒ ๋ฉ์์ง๋ฅผ ๋ณด๋ด๋ ๊ฒ์ด ๋ถ์ํฉ๋๋ค. ๊ฒฉ๋ ค์ ๋ง์ด๋ ์ ๊ทผ ๋ฐฉ๋ฒ์ ๋ํ ์ ์์ ํด์ค ์ ์๋์?",
|
| 782 |
}
|
| 783 |
],
|
| 784 |
[
|
| 785 |
{
|
| 786 |
+
"text": "๊ด๊ณ์์ ์ด๋ ค์์ ๊ทน๋ณตํ ๋ ์ฌ๋์ ๋ํ ๋ก๋งจํฑํ ์ด์ผ๊ธฐ๋ฅผ ๋ค๋ ค์ฃผ์ธ์.",
|
| 787 |
}
|
| 788 |
],
|
| 789 |
[
|
| 790 |
{
|
| 791 |
+
"text": "์์ ์ธ ๋ฐฉ์์ผ๋ก ์ฌ๋์ ํํํ๊ณ ์ถ์ต๋๋ค. ์ ํํธ๋๋ฅผ ์ํ ์ง์ฌ์ด ๋ด๊ธด ์๋ฅผ ์์ฑํ๋ ๋ฐ ๋์์ ์ค ์ ์๋์?",
|
| 792 |
}
|
| 793 |
],
|
| 794 |
[
|
| 795 |
{
|
| 796 |
+
"text": "์์ ๋คํผ์ด ์์์ต๋๋ค. ์ง์ฌ์ผ๋ก ์ฌ๊ณผํ๋ฉด์ ์ ๊ฐ์ ์ ํํํ ์ ์๋ ๋ฐฉ๋ฒ์ ์ฐพ์์ฃผ์ธ์.",
|
| 797 |
}
|
| 798 |
],
|
| 799 |
]
|
| 800 |
|
| 801 |
+
# =============================================================================
|
| 802 |
+
# Gradio UI (Blocks) ๊ตฌ์ฑ
|
| 803 |
+
# =============================================================================
|
| 804 |
+
|
| 805 |
+
# 1. Gradio Blocks UI ์์ - ๊ฐค๋ฌ๋ฆฌ ์ปดํฌ๋ํธ ์ถ๊ฐ
|
| 806 |
css = """
|
|
|
|
| 807 |
.gradio-container {
|
| 808 |
+
background: rgba(255, 255, 255, 0.7);
|
| 809 |
padding: 30px 40px;
|
| 810 |
+
margin: 20px auto;
|
| 811 |
width: 100% !important;
|
| 812 |
+
max-width: none !important;
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 813 |
}
|
| 814 |
"""
|
|
|
|
| 815 |
title_html = """
|
| 816 |
<h1 align="center" style="margin-bottom: 0.2em; font-size: 1.6em;"> ๐ HeartSync ๐ </h1>
|
| 817 |
<p align="center" style="font-size:1.1em; color:#555;">
|
| 818 |
+
โ
FLUX ์ด๋ฏธ์ง ์์ฑ โ
์ถ๋ก โ
๊ฒ์ด ํด์ โ
๋ฉํฐ๋ชจ๋ฌ & VLM โ
์ค์๊ฐ ์น ๊ฒ์ โ
RAG <br>
|
| 819 |
</p>
|
| 820 |
"""
|
| 821 |
|
| 822 |
with gr.Blocks(css=css, title="HeartSync") as demo:
|
| 823 |
gr.Markdown(title_html)
|
| 824 |
+
|
| 825 |
+
# ์์ฑ๋ ์ด๋ฏธ์ง๋ฅผ ์ ์ฅํ ๊ฐค๋ฌ๋ฆฌ ์ปดํฌ๋ํธ (์ด ๋ถ๋ถ์ด ์๋ก ์ถ๊ฐ๋จ)
|
| 826 |
+
generated_images = gr.Gallery(
|
| 827 |
+
label="์์ฑ๋ ์ด๋ฏธ์ง",
|
| 828 |
+
show_label=True,
|
| 829 |
+
visible=False,
|
| 830 |
+
elem_id="generated_images",
|
| 831 |
+
columns=2,
|
| 832 |
+
height="auto",
|
| 833 |
+
object_fit="contain"
|
| 834 |
)
|
| 835 |
+
|
| 836 |
+
with gr.Row():
|
| 837 |
+
web_search_checkbox = gr.Checkbox(label="์ฌ๋ ์๋ ์ฐ๊ตฌ", value=False)
|
| 838 |
+
image_gen_checkbox = gr.Checkbox(label="์ด๋ฏธ์ง ์์ฑ", value=False)
|
| 839 |
+
|
| 840 |
+
base_system_prompt_box = gr.Textbox(
|
| 841 |
lines=3,
|
| 842 |
+
value="๋น์ ์ ๊น์ด ์ฌ๊ณ ํ๋ AI์
๋๋ค. ํญ์ ๋
ผ๋ฆฌ์ ์ด๊ณ ์ฐฝ์์ ์ผ๋ก ๋ฌธ์ ๋ฅผ ํด๊ฒฐํฉ๋๋ค.\nํ๋ฅด์๋: ๋น์ ์ ๋ค์ ํ๊ณ ์ฌ๋์ด ๋์น๋ ์ฌ์์น๊ตฌ์
๋๋ค.",
|
| 843 |
+
label="๊ธฐ๋ณธ ์์คํ
ํ๋กฌํํธ",
|
| 844 |
+
visible=False
|
| 845 |
+
)
|
| 846 |
+
with gr.Row():
|
| 847 |
+
age_group_dropdown = gr.Dropdown(
|
| 848 |
+
label="์ฐ๋ น๋ ์ ํ (๊ธฐ๋ณธ 20๋)",
|
| 849 |
+
choices=["10๋", "20๋", "30~40๋", "50~60๋", "70๋ ์ด์"],
|
| 850 |
+
value="20๋",
|
| 851 |
+
interactive=True
|
| 852 |
+
)
|
| 853 |
+
mbti_choices = [
|
| 854 |
+
"INTJ (์ฉ์์ฃผ๋ํ ์ ๋ต๊ฐ)",
|
| 855 |
+
"INTP (๋
ผ๋ฆฌ์ ์ธ ์ฌ์๊ฐ)",
|
| 856 |
+
"ENTJ (๋๋ดํ ํต์์)",
|
| 857 |
+
"ENTP (๋จ๊ฑฐ์ด ๋
ผ์๊ฐ)",
|
| 858 |
+
"INFJ (์ ์์ ์นํธ์)",
|
| 859 |
+
"INFP (์ด์ ์ ์ธ ์ค์ฌ์)",
|
| 860 |
+
"ENFJ (์ ์๋ก์ด ์ฌํ์ด๋๊ฐ)",
|
| 861 |
+
"ENFP (์ฌ๊ธฐ๋ฐ๋ํ ํ๋๊ฐ)",
|
| 862 |
+
"ISTJ (์ฒญ๋ ด๊ฒฐ๋ฐฑํ ๋
ผ๋ฆฌ์ฃผ์์)",
|
| 863 |
+
"ISFJ (์ฉ๊ฐํ ์ํธ์)",
|
| 864 |
+
"ESTJ (์๊ฒฉํ ๊ด๋ฆฌ์)",
|
| 865 |
+
"ESFJ (์ฌ๊ต์ ์ธ ์ธ๊ต๊ด)",
|
| 866 |
+
"ISTP (๋ง๋ฅ ์ฌ์ฃผ๊พผ)",
|
| 867 |
+
"ISFP (ํธ๊ธฐ์ฌ ๋ง์ ์์ ๊ฐ)",
|
| 868 |
+
"ESTP (๋ชจํ์ ์ฆ๊ธฐ๋ ์ฌ์
๊ฐ)",
|
| 869 |
+
"ESFP (์์ ๋ก์ด ์ํผ์ ์ฐ์์ธ)"
|
| 870 |
+
]
|
| 871 |
+
mbti_dropdown = gr.Dropdown(
|
| 872 |
+
label="AI ํ๋ฅด์๋ MBTI (๊ธฐ๋ณธ INTP)",
|
| 873 |
+
choices=mbti_choices,
|
| 874 |
+
value="INTP (๋
ผ๋ฆฌ์ ์ธ ์ฌ์๊ฐ)",
|
| 875 |
+
interactive=True
|
| 876 |
+
)
|
| 877 |
+
sexual_openness_slider = gr.Slider(
|
| 878 |
+
minimum=1, maximum=5, step=1, value=2,
|
| 879 |
+
label="์น์์ผ ๊ด์ฌ๋/๊ฐ๋ฐฉ์ฑ (1~5, ๊ธฐ๋ณธ=2)",
|
| 880 |
+
interactive=True
|
| 881 |
)
|
|
|
|
| 882 |
max_tokens_slider = gr.Slider(
|
| 883 |
+
label="์ต๋ ์์ฑ ํ ํฐ ์",
|
| 884 |
+
minimum=100, maximum=8000, step=50, value=1000,
|
| 885 |
+
visible=False
|
|
|
|
|
|
|
|
|
|
| 886 |
)
|
|
|
|
| 887 |
web_search_text = gr.Textbox(
|
| 888 |
lines=1,
|
| 889 |
+
label="์น ๊ฒ์ ์ฟผ๋ฆฌ (๋ฏธ์ฌ์ฉ)",
|
| 890 |
+
placeholder="์ง์ ์
๋ ฅํ ํ์ ์์",
|
| 891 |
+
visible=False
|
| 892 |
)
|
| 893 |
+
|
| 894 |
+
# ์ฑํ
์ธํฐํ์ด์ค ์์ฑ - ์์ ๋ run ํจ์ ์ฌ์ฉ
|
| 895 |
chat = gr.ChatInterface(
|
| 896 |
+
fn=modified_run, # ์ฌ๊ธฐ์ ์์ ๋ ํจ์ ์ฌ์ฉ
|
| 897 |
type="messages",
|
| 898 |
chatbot=gr.Chatbot(type="messages", scale=1, allow_tags=["image"]),
|
| 899 |
textbox=gr.MultimodalTextbox(
|
| 900 |
+
file_types=[".webp", ".png", ".jpg", ".jpeg", ".gif", ".mp4", ".csv", ".txt", ".pdf"],
|
|
|
|
|
|
|
|
|
|
| 901 |
file_count="multiple",
|
| 902 |
autofocus=True
|
| 903 |
),
|
| 904 |
multimodal=True,
|
| 905 |
additional_inputs=[
|
| 906 |
+
base_system_prompt_box,
|
| 907 |
max_tokens_slider,
|
| 908 |
web_search_checkbox,
|
| 909 |
web_search_text,
|
| 910 |
+
age_group_dropdown,
|
| 911 |
+
mbti_dropdown,
|
| 912 |
+
sexual_openness_slider,
|
| 913 |
+
image_gen_checkbox,
|
| 914 |
+
],
|
| 915 |
+
additional_outputs=[
|
| 916 |
+
generated_images, # ๊ฐค๋ฌ๋ฆฌ ์ปดํฌ๋ํธ๋ฅผ ์ถ๋ ฅ์ผ๋ก ์ถ๊ฐ
|
| 917 |
],
|
| 918 |
stop_btn=False,
|
| 919 |
title='<a href="https://discord.gg/openfreeai" target="_blank">https://discord.gg/openfreeai</a>',
|
|
|
|
| 924 |
delete_cache=(1800, 1800),
|
| 925 |
)
|
| 926 |
|
| 927 |
+
|
| 928 |
with gr.Row(elem_id="examples_row"):
|
| 929 |
with gr.Column(scale=12, elem_id="examples_container"):
|
| 930 |
+
gr.Markdown("### ์์ ์
๋ ฅ (ํด๋ฆญํ์ฌ ๋ถ๋ฌ์ค๊ธฐ)")
|
| 931 |
+
|
| 932 |
if __name__ == "__main__":
|
| 933 |
+
demo.launch(share=True)
|
|
|