Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
|
@@ -26,6 +26,23 @@ import PyPDF2
|
|
| 26 |
##############################################################################
|
| 27 |
SERPHOUSE_API_KEY = "V38CNn4HXpLtynJQyOeoUensTEYoFy8PBUxKpDqAW1pawT1vfJ2BWtPQ98h6"
|
| 28 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 29 |
##############################################################################
|
| 30 |
# Simple function to call the SERPHouse Live endpoint
|
| 31 |
# https://api.serphouse.com/serp/live
|
|
@@ -33,7 +50,7 @@ SERPHOUSE_API_KEY = "V38CNn4HXpLtynJQyOeoUensTEYoFy8PBUxKpDqAW1pawT1vfJ2BWtPQ98h
|
|
| 33 |
def do_web_search(query: str) -> str:
|
| 34 |
"""
|
| 35 |
Calls SERPHouse live endpoint with the given query (q).
|
| 36 |
-
Returns a
|
| 37 |
"""
|
| 38 |
try:
|
| 39 |
url = "https://api.serphouse.com/serp/live"
|
|
@@ -43,27 +60,26 @@ def do_web_search(query: str) -> str:
|
|
| 43 |
"lang": "en",
|
| 44 |
"device": "desktop",
|
| 45 |
"serp_type": "web",
|
|
|
|
| 46 |
"api_token": SERPHOUSE_API_KEY,
|
| 47 |
}
|
| 48 |
resp = requests.get(url, params=params, timeout=30)
|
| 49 |
resp.raise_for_status() # Raise an exception for 4xx/5xx errors
|
| 50 |
data = resp.json()
|
| 51 |
|
| 52 |
-
# For demonstration, let's extract top 3 organic results:
|
| 53 |
results = data.get("results", {})
|
| 54 |
organic = results.get("results", {}).get("organic", [])
|
| 55 |
if not organic:
|
| 56 |
return "No web search results found."
|
| 57 |
|
|
|
|
| 58 |
summary_lines = []
|
| 59 |
-
for item in organic[:
|
| 60 |
-
rank = item.get("position", "-")
|
| 61 |
title = item.get("title", "No Title")
|
| 62 |
-
|
| 63 |
-
snippet = item.get("snippet", "(No snippet)")
|
| 64 |
-
summary_lines.append(f"**Rank {rank}:** [{title}]({link})\n\n> {snippet}")
|
| 65 |
|
| 66 |
-
|
|
|
|
| 67 |
except Exception as e:
|
| 68 |
logger.error(f"Web search failed: {e}")
|
| 69 |
return f"Web search failed: {str(e)}"
|
|
@@ -87,15 +103,10 @@ MAX_NUM_IMAGES = int(os.getenv("MAX_NUM_IMAGES", "5"))
|
|
| 87 |
# CSV, TXT, PDF ๋ถ์ ํจ์
|
| 88 |
##################################################
|
| 89 |
def analyze_csv_file(path: str) -> str:
|
| 90 |
-
"""
|
| 91 |
-
CSV ํ์ผ์ ์ ์ฒด ๋ฌธ์์ด๋ก ๋ณํ. ๋๋ฌด ๊ธธ ๊ฒฝ์ฐ ์ผ๋ถ๋ง ํ์.
|
| 92 |
-
"""
|
| 93 |
try:
|
| 94 |
df = pd.read_csv(path)
|
| 95 |
-
# ๋ฐ์ดํฐ ํ๋ ์ ํฌ๊ธฐ ์ ํ (ํ/์ด ์๊ฐ ๋ง์ ๊ฒฝ์ฐ)
|
| 96 |
if df.shape[0] > 50 or df.shape[1] > 10:
|
| 97 |
df = df.iloc[:50, :10]
|
| 98 |
-
|
| 99 |
df_str = df.to_string()
|
| 100 |
if len(df_str) > MAX_CONTENT_CHARS:
|
| 101 |
df_str = df_str[:MAX_CONTENT_CHARS] + "\n...(truncated)..."
|
|
@@ -105,9 +116,6 @@ def analyze_csv_file(path: str) -> str:
|
|
| 105 |
|
| 106 |
|
| 107 |
def analyze_txt_file(path: str) -> str:
|
| 108 |
-
"""
|
| 109 |
-
TXT ํ์ผ ์ ๋ฌธ ์ฝ๊ธฐ. ๋๋ฌด ๊ธธ๋ฉด ์ผ๋ถ๋ง ํ์.
|
| 110 |
-
"""
|
| 111 |
try:
|
| 112 |
with open(path, "r", encoding="utf-8") as f:
|
| 113 |
text = f.read()
|
|
@@ -119,25 +127,19 @@ def analyze_txt_file(path: str) -> str:
|
|
| 119 |
|
| 120 |
|
| 121 |
def pdf_to_markdown(pdf_path: str) -> str:
|
| 122 |
-
"""
|
| 123 |
-
PDF โ Markdown. ํ์ด์ง๋ณ๋ก ๊ฐ๋จํ ํ
์คํธ ์ถ์ถ.
|
| 124 |
-
"""
|
| 125 |
text_chunks = []
|
| 126 |
try:
|
| 127 |
with open(pdf_path, "rb") as f:
|
| 128 |
reader = PyPDF2.PdfReader(f)
|
| 129 |
-
# ์ต๋ 5ํ์ด์ง๋ง ์ฒ๋ฆฌ
|
| 130 |
max_pages = min(5, len(reader.pages))
|
| 131 |
for page_num in range(max_pages):
|
| 132 |
page = reader.pages[page_num]
|
| 133 |
page_text = page.extract_text() or ""
|
| 134 |
page_text = page_text.strip()
|
| 135 |
if page_text:
|
| 136 |
-
# ํ์ด์ง๋ณ ํ
์คํธ๋ ์ ํ
|
| 137 |
if len(page_text) > MAX_CONTENT_CHARS // max_pages:
|
| 138 |
page_text = page_text[:MAX_CONTENT_CHARS // max_pages] + "...(truncated)"
|
| 139 |
text_chunks.append(f"## Page {page_num+1}\n\n{page_text}\n")
|
| 140 |
-
|
| 141 |
if len(reader.pages) > max_pages:
|
| 142 |
text_chunks.append(f"\n...(Showing {max_pages} of {len(reader.pages)} pages)...")
|
| 143 |
except Exception as e:
|
|
@@ -181,14 +183,6 @@ def count_files_in_history(history: list[dict]) -> tuple[int, int]:
|
|
| 181 |
|
| 182 |
|
| 183 |
def validate_media_constraints(message: dict, history: list[dict]) -> bool:
|
| 184 |
-
"""
|
| 185 |
-
- ๋น๋์ค 1๊ฐ ์ด๊ณผ ๋ถ๊ฐ
|
| 186 |
-
- ๋น๋์ค์ ์ด๋ฏธ์ง ํผํฉ ๋ถ๊ฐ
|
| 187 |
-
- ์ด๋ฏธ์ง ๊ฐ์ MAX_NUM_IMAGES ์ด๊ณผ ๋ถ๊ฐ
|
| 188 |
-
- <image> ํ๊ทธ๊ฐ ์์ผ๋ฉด ํ๊ทธ ์์ ์ค์ ์ด๋ฏธ์ง ์ ์ผ์น
|
| 189 |
-
- CSV, TXT, PDF ๋ฑ์ ์ฌ๊ธฐ์ ์ ํํ์ง ์์
|
| 190 |
-
"""
|
| 191 |
-
# ์ด๋ฏธ์ง์ ๋น๋์ค ํ์ผ๋ง ํํฐ๋ง
|
| 192 |
media_files = []
|
| 193 |
for f in message["files"]:
|
| 194 |
if re.search(r"\.(png|jpg|jpeg|gif|webp)$", f, re.IGNORECASE) or f.endswith(".mp4"):
|
|
@@ -213,9 +207,7 @@ def validate_media_constraints(message: dict, history: list[dict]) -> bool:
|
|
| 213 |
gr.Warning(f"You can upload up to {MAX_NUM_IMAGES} images.")
|
| 214 |
return False
|
| 215 |
|
| 216 |
-
# ์ด๋ฏธ์ง ํ๊ทธ ๊ฒ์ฆ (์ค์ ์ด๋ฏธ์ง ํ์ผ๋ง ๊ณ์ฐ)
|
| 217 |
if "<image>" in message["text"]:
|
| 218 |
-
# ์ด๋ฏธ์ง ํ์ผ๋ง ํํฐ๋ง
|
| 219 |
image_files = [f for f in message["files"] if re.search(r"\.(png|jpg|jpeg|gif|webp)$", f, re.IGNORECASE)]
|
| 220 |
image_tag_count = message["text"].count("<image>")
|
| 221 |
if image_tag_count != len(image_files):
|
|
@@ -232,9 +224,7 @@ def downsample_video(video_path: str) -> list[tuple[Image.Image, float]]:
|
|
| 232 |
vidcap = cv2.VideoCapture(video_path)
|
| 233 |
fps = vidcap.get(cv2.CAP_PROP_FPS)
|
| 234 |
total_frames = int(vidcap.get(cv2.CAP_PROP_FRAME_COUNT))
|
| 235 |
-
|
| 236 |
-
# ๋ ์ ์ ํ๋ ์์ ์ถ์ถํ๋๋ก ์กฐ์
|
| 237 |
-
frame_interval = max(int(fps), int(total_frames / 10)) # ์ด๋น 1ํ๋ ์ ๋๋ ์ต๋ 10ํ๋ ์
|
| 238 |
frames = []
|
| 239 |
|
| 240 |
for i in range(0, total_frames, frame_interval):
|
|
@@ -245,8 +235,6 @@ def downsample_video(video_path: str) -> list[tuple[Image.Image, float]]:
|
|
| 245 |
pil_image = Image.fromarray(image)
|
| 246 |
timestamp = round(i / fps, 2)
|
| 247 |
frames.append((pil_image, timestamp))
|
| 248 |
-
|
| 249 |
-
# ์ต๋ 5ํ๋ ์๋ง ์ฌ์ฉ
|
| 250 |
if len(frames) >= 5:
|
| 251 |
break
|
| 252 |
|
|
@@ -275,7 +263,6 @@ def process_interleaved_images(message: dict) -> list[dict]:
|
|
| 275 |
content = []
|
| 276 |
image_index = 0
|
| 277 |
|
| 278 |
-
# ์ด๋ฏธ์ง ํ์ผ๋ง ํํฐ๋ง
|
| 279 |
image_files = [f for f in message["files"] if re.search(r"\.(png|jpg|jpeg|gif|webp)$", f, re.IGNORECASE)]
|
| 280 |
|
| 281 |
for part in parts:
|
|
@@ -285,7 +272,6 @@ def process_interleaved_images(message: dict) -> list[dict]:
|
|
| 285 |
elif part.strip():
|
| 286 |
content.append({"type": "text", "text": part.strip()})
|
| 287 |
else:
|
| 288 |
-
# ๊ณต๋ฐฑ์ด๊ฑฐ๋ \n ๊ฐ์ ๊ฒฝ์ฐ
|
| 289 |
if isinstance(part, str) and part != "<image>":
|
| 290 |
content.append({"type": "text", "text": part})
|
| 291 |
return content
|
|
@@ -295,66 +281,50 @@ def process_interleaved_images(message: dict) -> list[dict]:
|
|
| 295 |
# PDF + CSV + TXT + ์ด๋ฏธ์ง/๋น๋์ค
|
| 296 |
##################################################
|
| 297 |
def is_image_file(file_path: str) -> bool:
|
| 298 |
-
"""์ด๋ฏธ์ง ํ์ผ์ธ์ง ํ์ธ"""
|
| 299 |
return bool(re.search(r"\.(png|jpg|jpeg|gif|webp)$", file_path, re.IGNORECASE))
|
| 300 |
|
| 301 |
-
|
| 302 |
def is_video_file(file_path: str) -> bool:
|
| 303 |
-
"""๋น๋์ค ํ์ผ์ธ์ง ํ์ธ"""
|
| 304 |
return file_path.endswith(".mp4")
|
| 305 |
|
| 306 |
-
|
| 307 |
def is_document_file(file_path: str) -> bool:
|
| 308 |
-
"""๋ฌธ์ ํ์ผ์ธ์ง ํ์ธ (PDF, CSV, TXT)"""
|
| 309 |
return (file_path.lower().endswith(".pdf") or
|
| 310 |
file_path.lower().endswith(".csv") or
|
| 311 |
file_path.lower().endswith(".txt"))
|
| 312 |
|
| 313 |
-
|
| 314 |
def process_new_user_message(message: dict) -> list[dict]:
|
| 315 |
if not message["files"]:
|
| 316 |
return [{"type": "text", "text": message["text"]}]
|
| 317 |
|
| 318 |
-
# 1) ํ์ผ ๋ถ๋ฅ
|
| 319 |
video_files = [f for f in message["files"] if is_video_file(f)]
|
| 320 |
image_files = [f for f in message["files"] if is_image_file(f)]
|
| 321 |
csv_files = [f for f in message["files"] if f.lower().endswith(".csv")]
|
| 322 |
txt_files = [f for f in message["files"] if f.lower().endswith(".txt")]
|
| 323 |
pdf_files = [f for f in message["files"] if f.lower().endswith(".pdf")]
|
| 324 |
|
| 325 |
-
# 2) ์ฌ์ฉ์ ์๋ณธ text ์ถ๊ฐ
|
| 326 |
content_list = [{"type": "text", "text": message["text"]}]
|
| 327 |
|
| 328 |
-
# 3) CSV
|
| 329 |
for csv_path in csv_files:
|
| 330 |
csv_analysis = analyze_csv_file(csv_path)
|
| 331 |
content_list.append({"type": "text", "text": csv_analysis})
|
| 332 |
|
| 333 |
-
# 4) TXT
|
| 334 |
for txt_path in txt_files:
|
| 335 |
txt_analysis = analyze_txt_file(txt_path)
|
| 336 |
content_list.append({"type": "text", "text": txt_analysis})
|
| 337 |
|
| 338 |
-
# 5) PDF
|
| 339 |
for pdf_path in pdf_files:
|
| 340 |
pdf_markdown = pdf_to_markdown(pdf_path)
|
| 341 |
content_list.append({"type": "text", "text": pdf_markdown})
|
| 342 |
|
| 343 |
-
# 6) ๋น๋์ค (ํ ๊ฐ๋ง ํ์ฉ)
|
| 344 |
if video_files:
|
| 345 |
content_list += process_video(video_files[0])
|
| 346 |
return content_list
|
| 347 |
|
| 348 |
-
# 7) ์ด๋ฏธ์ง ์ฒ๋ฆฌ
|
| 349 |
if "<image>" in message["text"] and image_files:
|
| 350 |
-
# interleaved
|
| 351 |
interleaved_content = process_interleaved_images({"text": message["text"], "files": image_files})
|
| 352 |
-
# ์๋ณธ content_list ์๋ถ๋ถ(ํ
์คํธ)์ ์ ๊ฑฐํ๊ณ interleaved๋ก ๋์ฒด
|
| 353 |
if content_list[0]["type"] == "text":
|
| 354 |
-
content_list = content_list[1:]
|
| 355 |
-
return interleaved_content + content_list
|
| 356 |
else:
|
| 357 |
-
# ์ผ๋ฐ ์ฌ๋ฌ ์ฅ
|
| 358 |
for img_path in image_files:
|
| 359 |
content_list.append({"type": "image", "url": img_path})
|
| 360 |
|
|
@@ -369,14 +339,11 @@ def process_history(history: list[dict]) -> list[dict]:
|
|
| 369 |
current_user_content: list[dict] = []
|
| 370 |
for item in history:
|
| 371 |
if item["role"] == "assistant":
|
| 372 |
-
# user_content๊ฐ ์์ฌ์๋ค๋ฉด user ๋ฉ์์ง๋ก ์ ์ฅ
|
| 373 |
if current_user_content:
|
| 374 |
messages.append({"role": "user", "content": current_user_content})
|
| 375 |
current_user_content = []
|
| 376 |
-
# ๊ทธ ๋ค item์ assistant
|
| 377 |
messages.append({"role": "assistant", "content": [{"type": "text", "text": item["content"]}]})
|
| 378 |
else:
|
| 379 |
-
# user
|
| 380 |
content = item["content"]
|
| 381 |
if isinstance(content, str):
|
| 382 |
current_user_content.append({"type": "text", "text": content})
|
|
@@ -385,10 +352,8 @@ def process_history(history: list[dict]) -> list[dict]:
|
|
| 385 |
if is_image_file(file_path):
|
| 386 |
current_user_content.append({"type": "image", "url": file_path})
|
| 387 |
else:
|
| 388 |
-
# ๋น์ด๋ฏธ์ง ํ์ผ์ ํ
์คํธ๋ก ์ฒ๋ฆฌ
|
| 389 |
current_user_content.append({"type": "text", "text": f"[File: {os.path.basename(file_path)}]"})
|
| 390 |
-
|
| 391 |
-
# ๋ง์ง๋ง ์ฌ์ฉ์ ๋ฉ์์ง๊ฐ ์ฒ๋ฆฌ๋์ง ์์ ๊ฒฝ์ฐ ์ถ๊ฐ
|
| 392 |
if current_user_content:
|
| 393 |
messages.append({"role": "user", "content": current_user_content})
|
| 394 |
|
|
@@ -407,60 +372,54 @@ def run(
|
|
| 407 |
use_web_search: bool = False,
|
| 408 |
web_search_query: str = "",
|
| 409 |
) -> Iterator[str]:
|
| 410 |
-
|
| 411 |
-
The main inference function. Now extended with optional web_search arguments:
|
| 412 |
-
- use_web_search: bool
|
| 413 |
-
- web_search_query: str
|
| 414 |
-
If `use_web_search` is True, calls SERPHouse for the given `web_search_query`.
|
| 415 |
-
"""
|
| 416 |
-
# Validate media constraints first
|
| 417 |
if not validate_media_constraints(message, history):
|
| 418 |
yield ""
|
| 419 |
return
|
| 420 |
|
| 421 |
try:
|
| 422 |
-
#
|
| 423 |
-
|
| 424 |
-
|
| 425 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 426 |
|
| 427 |
messages = []
|
| 428 |
if system_prompt:
|
| 429 |
messages.append({"role": "system", "content": [{"type": "text", "text": system_prompt}]})
|
|
|
|
|
|
|
|
|
|
|
|
|
| 430 |
messages.extend(process_history(history))
|
| 431 |
|
| 432 |
-
# ์ฌ์ฉ์ ๋ฉ์์ง ์ฒ๋ฆฌ
|
| 433 |
user_content = process_new_user_message(message)
|
| 434 |
-
|
| 435 |
-
# ํ ํฐ ์๋ฅผ ์ค์ด๊ธฐ ์ํด ๋๋ฌด ๊ธด ํ
์คํธ๋ ์๋ผ๋ด๊ธฐ
|
| 436 |
for item in user_content:
|
| 437 |
if item["type"] == "text" and len(item["text"]) > MAX_CONTENT_CHARS:
|
| 438 |
item["text"] = item["text"][:MAX_CONTENT_CHARS] + "\n...(truncated)..."
|
| 439 |
|
| 440 |
messages.append({"role": "user", "content": user_content})
|
| 441 |
|
| 442 |
-
# ๋ชจ๋ธ ์
๋ ฅ ์์ฑ ์ ์ต์ข
ํ์ธ
|
| 443 |
-
for msg in messages:
|
| 444 |
-
if msg["role"] != "user":
|
| 445 |
-
continue
|
| 446 |
-
|
| 447 |
-
filtered_content = []
|
| 448 |
-
for item in msg["content"]:
|
| 449 |
-
if item["type"] == "image":
|
| 450 |
-
if is_image_file(item["url"]):
|
| 451 |
-
filtered_content.append(item)
|
| 452 |
-
else:
|
| 453 |
-
# ์ด๋ฏธ์ง ํ์ผ์ด ์๋ ๊ฒฝ์ฐ ํ
์คํธ๋ก ๋ณํ
|
| 454 |
-
filtered_content.append({
|
| 455 |
-
"type": "text",
|
| 456 |
-
"text": f"[Non-image file: {os.path.basename(item['url'])}]"
|
| 457 |
-
})
|
| 458 |
-
else:
|
| 459 |
-
filtered_content.append(item)
|
| 460 |
-
|
| 461 |
-
msg["content"] = filtered_content
|
| 462 |
-
|
| 463 |
-
# ๋ชจ๋ธ ์
๋ ฅ ์์ฑ
|
| 464 |
inputs = processor.apply_chat_template(
|
| 465 |
messages,
|
| 466 |
add_generation_prompt=True,
|
|
@@ -469,7 +428,6 @@ def run(
|
|
| 469 |
return_tensors="pt",
|
| 470 |
).to(device=model.device, dtype=torch.bfloat16)
|
| 471 |
|
| 472 |
-
# ํ
์คํธ ์์ฑ ์คํธ๋ฆฌ๋จธ ์ค์
|
| 473 |
streamer = TextIteratorStreamer(processor, timeout=30.0, skip_prompt=True, skip_special_tokens=True)
|
| 474 |
gen_kwargs = dict(
|
| 475 |
inputs,
|
|
@@ -477,11 +435,9 @@ def run(
|
|
| 477 |
max_new_tokens=max_new_tokens,
|
| 478 |
)
|
| 479 |
|
| 480 |
-
# ๋ณ๋ ์ค๋ ๋์์ ํ
์คํธ ์์ฑ
|
| 481 |
t = Thread(target=model.generate, kwargs=gen_kwargs)
|
| 482 |
t.start()
|
| 483 |
|
| 484 |
-
# ๊ฒฐ๊ณผ ์คํธ๋ฆฌ๋ฐ
|
| 485 |
output = ""
|
| 486 |
for new_text in streamer:
|
| 487 |
output += new_text
|
|
@@ -493,9 +449,6 @@ def run(
|
|
| 493 |
|
| 494 |
|
| 495 |
|
| 496 |
-
##################################################
|
| 497 |
-
# ์์๋ค (ํ๊ธํ ๋ฒ์ )
|
| 498 |
-
##################################################
|
| 499 |
examples = [
|
| 500 |
|
| 501 |
[
|
|
@@ -601,12 +554,6 @@ examples = [
|
|
| 601 |
]
|
| 602 |
|
| 603 |
|
| 604 |
-
|
| 605 |
-
|
| 606 |
-
|
| 607 |
-
##############################################################################
|
| 608 |
-
# Custom CSS similar to second example (colorful background, panel, etc.)
|
| 609 |
-
##############################################################################
|
| 610 |
css = """
|
| 611 |
body {
|
| 612 |
background: linear-gradient(135deg, #667eea, #764ba2);
|
|
@@ -668,11 +615,6 @@ title_html = """
|
|
| 668 |
</p>
|
| 669 |
"""
|
| 670 |
|
| 671 |
-
##############################################################################
|
| 672 |
-
# Build a Blocks layout that includes:
|
| 673 |
-
# - A left sidebar with "Web Search" controls
|
| 674 |
-
# - The main ChatInterface in the center or right
|
| 675 |
-
##############################################################################
|
| 676 |
with gr.Blocks(css=css, title="Vidraft-Gemma-3-27B") as demo:
|
| 677 |
gr.Markdown(title_html)
|
| 678 |
|
|
@@ -686,10 +628,11 @@ with gr.Blocks(css=css, title="Vidraft-Gemma-3-27B") as demo:
|
|
| 686 |
value=False,
|
| 687 |
info="Check to enable a SERPHouse web search before the chat reply"
|
| 688 |
)
|
|
|
|
| 689 |
web_search_text = gr.Textbox(
|
| 690 |
lines=1,
|
| 691 |
-
label="Web Search Query",
|
| 692 |
-
placeholder="
|
| 693 |
)
|
| 694 |
|
| 695 |
gr.Markdown("---")
|
|
@@ -710,9 +653,8 @@ with gr.Blocks(css=css, title="Vidraft-Gemma-3-27B") as demo:
|
|
| 710 |
value=2000,
|
| 711 |
)
|
| 712 |
|
| 713 |
-
gr.Markdown("<br><br>")
|
| 714 |
|
| 715 |
-
# Main ChatInterface to the right
|
| 716 |
with gr.Column(scale=7):
|
| 717 |
chat = gr.ChatInterface(
|
| 718 |
fn=run,
|
|
@@ -731,7 +673,7 @@ with gr.Blocks(css=css, title="Vidraft-Gemma-3-27B") as demo:
|
|
| 731 |
system_prompt_box,
|
| 732 |
max_tokens_slider,
|
| 733 |
web_search_checkbox,
|
| 734 |
-
web_search_text,
|
| 735 |
],
|
| 736 |
stop_btn=False,
|
| 737 |
title="Vidraft-Gemma-3-27B",
|
|
@@ -745,10 +687,9 @@ with gr.Blocks(css=css, title="Vidraft-Gemma-3-27B") as demo:
|
|
| 745 |
with gr.Row(elem_id="examples_row"):
|
| 746 |
with gr.Column(scale=12, elem_id="examples_container"):
|
| 747 |
gr.Markdown("### Example Inputs (click to load)")
|
| 748 |
-
# The fix: pass an empty list to avoid the "None" error, so we keep the code structure.
|
| 749 |
gr.Examples(
|
| 750 |
examples=examples,
|
| 751 |
-
inputs=[], #
|
| 752 |
cache_examples=False
|
| 753 |
)
|
| 754 |
|
|
|
|
| 26 |
##############################################################################
|
| 27 |
SERPHOUSE_API_KEY = "V38CNn4HXpLtynJQyOeoUensTEYoFy8PBUxKpDqAW1pawT1vfJ2BWtPQ98h6"
|
| 28 |
|
| 29 |
+
##############################################################################
|
| 30 |
+
# [์๋ก ์ถ๊ฐ] ์ฌ์ฉ์ ๋ฉ์์ง๋ก๋ถํฐ ๊ฐ๋จํ ํค์๋ ์ถ์ถํ๋ ํจ์ ์์
|
| 31 |
+
# - ์ค์ ํ๊ฒฝ์ ๋ง๊ฒ stopwords, ํํ์ ๋ถ์ ๋ฑ ๊ณ ๋ํ ๊ฐ๋ฅ
|
| 32 |
+
##############################################################################
|
| 33 |
+
def extract_keywords(text: str, top_k: int = 5) -> str:
|
| 34 |
+
# 1) ์๋ฌธ์๋ก
|
| 35 |
+
text = text.lower()
|
| 36 |
+
# 2) ์ํ๋ฒณ/์ซ์/๊ณต๋ฐฑ ์ ์ธ ๋ฌธ์ ์ ๊ฑฐ
|
| 37 |
+
text = re.sub(r"[^a-z0-9\s]", "", text)
|
| 38 |
+
# 3) ๊ณต๋ฐฑ๋จ์ ํ ํฐ
|
| 39 |
+
tokens = text.split()
|
| 40 |
+
# 4) ์ฐ์ ์ ์์์ ๋ช ๊ฐ ํ ํฐ๋ง ์ฌ์ฉ (top_k=5)
|
| 41 |
+
# - ํ์์ stopword ์ ๊ฑฐ๋ ๋น๋์ ๊ณ์ฐ ํ ์์ k๊ฐ ์ถ์ถํ๋๋ก ๋ณ๊ฒฝ ๊ฐ๋ฅ
|
| 42 |
+
key_tokens = tokens[:top_k]
|
| 43 |
+
# 5) ๊ณต๋ฐฑ์ผ๋ก join
|
| 44 |
+
return " ".join(key_tokens)
|
| 45 |
+
|
| 46 |
##############################################################################
|
| 47 |
# Simple function to call the SERPHouse Live endpoint
|
| 48 |
# https://api.serphouse.com/serp/live
|
|
|
|
| 50 |
def do_web_search(query: str) -> str:
|
| 51 |
"""
|
| 52 |
Calls SERPHouse live endpoint with the given query (q).
|
| 53 |
+
Returns top-20 results' titles as a bullet list, or an error message.
|
| 54 |
"""
|
| 55 |
try:
|
| 56 |
url = "https://api.serphouse.com/serp/live"
|
|
|
|
| 60 |
"lang": "en",
|
| 61 |
"device": "desktop",
|
| 62 |
"serp_type": "web",
|
| 63 |
+
"num_result": "20", # [์๋ก ์ถ๊ฐ] ์์ 20๊ฐ ๊ฒฐ๊ณผ
|
| 64 |
"api_token": SERPHOUSE_API_KEY,
|
| 65 |
}
|
| 66 |
resp = requests.get(url, params=params, timeout=30)
|
| 67 |
resp.raise_for_status() # Raise an exception for 4xx/5xx errors
|
| 68 |
data = resp.json()
|
| 69 |
|
|
|
|
| 70 |
results = data.get("results", {})
|
| 71 |
organic = results.get("results", {}).get("organic", [])
|
| 72 |
if not organic:
|
| 73 |
return "No web search results found."
|
| 74 |
|
| 75 |
+
# ์์ 20๊ฐ ์ ๋ชฉ๋ง ๋ฝ์์ ์ ๋ฆฌ
|
| 76 |
summary_lines = []
|
| 77 |
+
for idx, item in enumerate(organic[:20], start=1):
|
|
|
|
| 78 |
title = item.get("title", "No Title")
|
| 79 |
+
summary_lines.append(f"{idx}. {title}")
|
|
|
|
|
|
|
| 80 |
|
| 81 |
+
# 20๊ฐ๋ฅผ \n ์ผ๋ก ์ฐ๊ฒฐ
|
| 82 |
+
return "\n".join(summary_lines)
|
| 83 |
except Exception as e:
|
| 84 |
logger.error(f"Web search failed: {e}")
|
| 85 |
return f"Web search failed: {str(e)}"
|
|
|
|
| 103 |
# CSV, TXT, PDF ๋ถ์ ํจ์
|
| 104 |
##################################################
|
| 105 |
def analyze_csv_file(path: str) -> str:
|
|
|
|
|
|
|
|
|
|
| 106 |
try:
|
| 107 |
df = pd.read_csv(path)
|
|
|
|
| 108 |
if df.shape[0] > 50 or df.shape[1] > 10:
|
| 109 |
df = df.iloc[:50, :10]
|
|
|
|
| 110 |
df_str = df.to_string()
|
| 111 |
if len(df_str) > MAX_CONTENT_CHARS:
|
| 112 |
df_str = df_str[:MAX_CONTENT_CHARS] + "\n...(truncated)..."
|
|
|
|
| 116 |
|
| 117 |
|
| 118 |
def analyze_txt_file(path: str) -> str:
|
|
|
|
|
|
|
|
|
|
| 119 |
try:
|
| 120 |
with open(path, "r", encoding="utf-8") as f:
|
| 121 |
text = f.read()
|
|
|
|
| 127 |
|
| 128 |
|
| 129 |
def pdf_to_markdown(pdf_path: str) -> str:
|
|
|
|
|
|
|
|
|
|
| 130 |
text_chunks = []
|
| 131 |
try:
|
| 132 |
with open(pdf_path, "rb") as f:
|
| 133 |
reader = PyPDF2.PdfReader(f)
|
|
|
|
| 134 |
max_pages = min(5, len(reader.pages))
|
| 135 |
for page_num in range(max_pages):
|
| 136 |
page = reader.pages[page_num]
|
| 137 |
page_text = page.extract_text() or ""
|
| 138 |
page_text = page_text.strip()
|
| 139 |
if page_text:
|
|
|
|
| 140 |
if len(page_text) > MAX_CONTENT_CHARS // max_pages:
|
| 141 |
page_text = page_text[:MAX_CONTENT_CHARS // max_pages] + "...(truncated)"
|
| 142 |
text_chunks.append(f"## Page {page_num+1}\n\n{page_text}\n")
|
|
|
|
| 143 |
if len(reader.pages) > max_pages:
|
| 144 |
text_chunks.append(f"\n...(Showing {max_pages} of {len(reader.pages)} pages)...")
|
| 145 |
except Exception as e:
|
|
|
|
| 183 |
|
| 184 |
|
| 185 |
def validate_media_constraints(message: dict, history: list[dict]) -> bool:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 186 |
media_files = []
|
| 187 |
for f in message["files"]:
|
| 188 |
if re.search(r"\.(png|jpg|jpeg|gif|webp)$", f, re.IGNORECASE) or f.endswith(".mp4"):
|
|
|
|
| 207 |
gr.Warning(f"You can upload up to {MAX_NUM_IMAGES} images.")
|
| 208 |
return False
|
| 209 |
|
|
|
|
| 210 |
if "<image>" in message["text"]:
|
|
|
|
| 211 |
image_files = [f for f in message["files"] if re.search(r"\.(png|jpg|jpeg|gif|webp)$", f, re.IGNORECASE)]
|
| 212 |
image_tag_count = message["text"].count("<image>")
|
| 213 |
if image_tag_count != len(image_files):
|
|
|
|
| 224 |
vidcap = cv2.VideoCapture(video_path)
|
| 225 |
fps = vidcap.get(cv2.CAP_PROP_FPS)
|
| 226 |
total_frames = int(vidcap.get(cv2.CAP_PROP_FRAME_COUNT))
|
| 227 |
+
frame_interval = max(int(fps), int(total_frames / 10))
|
|
|
|
|
|
|
| 228 |
frames = []
|
| 229 |
|
| 230 |
for i in range(0, total_frames, frame_interval):
|
|
|
|
| 235 |
pil_image = Image.fromarray(image)
|
| 236 |
timestamp = round(i / fps, 2)
|
| 237 |
frames.append((pil_image, timestamp))
|
|
|
|
|
|
|
| 238 |
if len(frames) >= 5:
|
| 239 |
break
|
| 240 |
|
|
|
|
| 263 |
content = []
|
| 264 |
image_index = 0
|
| 265 |
|
|
|
|
| 266 |
image_files = [f for f in message["files"] if re.search(r"\.(png|jpg|jpeg|gif|webp)$", f, re.IGNORECASE)]
|
| 267 |
|
| 268 |
for part in parts:
|
|
|
|
| 272 |
elif part.strip():
|
| 273 |
content.append({"type": "text", "text": part.strip()})
|
| 274 |
else:
|
|
|
|
| 275 |
if isinstance(part, str) and part != "<image>":
|
| 276 |
content.append({"type": "text", "text": part})
|
| 277 |
return content
|
|
|
|
| 281 |
# PDF + CSV + TXT + ์ด๋ฏธ์ง/๋น๋์ค
|
| 282 |
##################################################
|
| 283 |
def is_image_file(file_path: str) -> bool:
|
|
|
|
| 284 |
return bool(re.search(r"\.(png|jpg|jpeg|gif|webp)$", file_path, re.IGNORECASE))
|
| 285 |
|
|
|
|
| 286 |
def is_video_file(file_path: str) -> bool:
|
|
|
|
| 287 |
return file_path.endswith(".mp4")
|
| 288 |
|
|
|
|
| 289 |
def is_document_file(file_path: str) -> bool:
|
|
|
|
| 290 |
return (file_path.lower().endswith(".pdf") or
|
| 291 |
file_path.lower().endswith(".csv") or
|
| 292 |
file_path.lower().endswith(".txt"))
|
| 293 |
|
|
|
|
| 294 |
def process_new_user_message(message: dict) -> list[dict]:
|
| 295 |
if not message["files"]:
|
| 296 |
return [{"type": "text", "text": message["text"]}]
|
| 297 |
|
|
|
|
| 298 |
video_files = [f for f in message["files"] if is_video_file(f)]
|
| 299 |
image_files = [f for f in message["files"] if is_image_file(f)]
|
| 300 |
csv_files = [f for f in message["files"] if f.lower().endswith(".csv")]
|
| 301 |
txt_files = [f for f in message["files"] if f.lower().endswith(".txt")]
|
| 302 |
pdf_files = [f for f in message["files"] if f.lower().endswith(".pdf")]
|
| 303 |
|
|
|
|
| 304 |
content_list = [{"type": "text", "text": message["text"]}]
|
| 305 |
|
|
|
|
| 306 |
for csv_path in csv_files:
|
| 307 |
csv_analysis = analyze_csv_file(csv_path)
|
| 308 |
content_list.append({"type": "text", "text": csv_analysis})
|
| 309 |
|
|
|
|
| 310 |
for txt_path in txt_files:
|
| 311 |
txt_analysis = analyze_txt_file(txt_path)
|
| 312 |
content_list.append({"type": "text", "text": txt_analysis})
|
| 313 |
|
|
|
|
| 314 |
for pdf_path in pdf_files:
|
| 315 |
pdf_markdown = pdf_to_markdown(pdf_path)
|
| 316 |
content_list.append({"type": "text", "text": pdf_markdown})
|
| 317 |
|
|
|
|
| 318 |
if video_files:
|
| 319 |
content_list += process_video(video_files[0])
|
| 320 |
return content_list
|
| 321 |
|
|
|
|
| 322 |
if "<image>" in message["text"] and image_files:
|
|
|
|
| 323 |
interleaved_content = process_interleaved_images({"text": message["text"], "files": image_files})
|
|
|
|
| 324 |
if content_list[0]["type"] == "text":
|
| 325 |
+
content_list = content_list[1:]
|
| 326 |
+
return interleaved_content + content_list
|
| 327 |
else:
|
|
|
|
| 328 |
for img_path in image_files:
|
| 329 |
content_list.append({"type": "image", "url": img_path})
|
| 330 |
|
|
|
|
| 339 |
current_user_content: list[dict] = []
|
| 340 |
for item in history:
|
| 341 |
if item["role"] == "assistant":
|
|
|
|
| 342 |
if current_user_content:
|
| 343 |
messages.append({"role": "user", "content": current_user_content})
|
| 344 |
current_user_content = []
|
|
|
|
| 345 |
messages.append({"role": "assistant", "content": [{"type": "text", "text": item["content"]}]})
|
| 346 |
else:
|
|
|
|
| 347 |
content = item["content"]
|
| 348 |
if isinstance(content, str):
|
| 349 |
current_user_content.append({"type": "text", "text": content})
|
|
|
|
| 352 |
if is_image_file(file_path):
|
| 353 |
current_user_content.append({"type": "image", "url": file_path})
|
| 354 |
else:
|
|
|
|
| 355 |
current_user_content.append({"type": "text", "text": f"[File: {os.path.basename(file_path)}]"})
|
| 356 |
+
|
|
|
|
| 357 |
if current_user_content:
|
| 358 |
messages.append({"role": "user", "content": current_user_content})
|
| 359 |
|
|
|
|
| 372 |
use_web_search: bool = False,
|
| 373 |
web_search_query: str = "",
|
| 374 |
) -> Iterator[str]:
|
| 375 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 376 |
if not validate_media_constraints(message, history):
|
| 377 |
yield ""
|
| 378 |
return
|
| 379 |
|
| 380 |
try:
|
| 381 |
+
# [์๋ก ์ถ๊ฐ] web search ์ฒดํฌ๋ ๊ฒฝ์ฐ, ์ฌ์ฉ์๊ฐ ์
๋ ฅํ "web_search_query" ๋์
|
| 382 |
+
# ์ฌ์ฉ์์ ๋ฉ์์ง์์ ํค์๋๋ฅผ ์ถ์ถํ์ฌ ๊ฒ์
|
| 383 |
+
if use_web_search:
|
| 384 |
+
user_text = message["text"]
|
| 385 |
+
# ํค์๋ ์ถ์ถ
|
| 386 |
+
ws_query = extract_keywords(user_text, top_k=5)
|
| 387 |
+
logger.info(f"[Auto WebSearch Keyword] {ws_query!r}")
|
| 388 |
+
# ์์ 20๊ฐ ๊ฒฐ๊ณผ ๊ฐ์ ธ์ค๊ธฐ
|
| 389 |
+
ws_result = do_web_search(ws_query)
|
| 390 |
+
# ๊ฒ์๋ 20๊ฐ ์ ๋ชฉ์ system ๋ฉ์์ง์ ์ถ๊ฐ
|
| 391 |
+
system_search_content = f"[Search top-20 Titles Based on user prompt]\n{ws_result}\n"
|
| 392 |
+
# system ๋ฉ์์ง๋ก ์ถ๊ฐ
|
| 393 |
+
# (LLM์ด ์ด ์ ๋ณด๋ฅผ ์ฐธ๊ณ ํ๋๋ก)
|
| 394 |
+
if system_search_content.strip():
|
| 395 |
+
history_system_msg = {
|
| 396 |
+
"role": "system",
|
| 397 |
+
"content": [{"type": "text", "text": system_search_content}]
|
| 398 |
+
}
|
| 399 |
+
else:
|
| 400 |
+
history_system_msg = {
|
| 401 |
+
"role": "system",
|
| 402 |
+
"content": [{"type": "text", "text": "No web search results"}]
|
| 403 |
+
}
|
| 404 |
+
else:
|
| 405 |
+
history_system_msg = None
|
| 406 |
|
| 407 |
messages = []
|
| 408 |
if system_prompt:
|
| 409 |
messages.append({"role": "system", "content": [{"type": "text", "text": system_prompt}]})
|
| 410 |
+
# ๋ง์ฝ web search๊ฐ ์์๋ค๋ฉด, ๊ทธ ๊ฒฐ๊ณผ๋ฅผ ์ถ๊ฐ system ๋ฉ์์ง๋ก ์ฝ์
|
| 411 |
+
if history_system_msg:
|
| 412 |
+
messages.append(history_system_msg)
|
| 413 |
+
|
| 414 |
messages.extend(process_history(history))
|
| 415 |
|
|
|
|
| 416 |
user_content = process_new_user_message(message)
|
|
|
|
|
|
|
| 417 |
for item in user_content:
|
| 418 |
if item["type"] == "text" and len(item["text"]) > MAX_CONTENT_CHARS:
|
| 419 |
item["text"] = item["text"][:MAX_CONTENT_CHARS] + "\n...(truncated)..."
|
| 420 |
|
| 421 |
messages.append({"role": "user", "content": user_content})
|
| 422 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 423 |
inputs = processor.apply_chat_template(
|
| 424 |
messages,
|
| 425 |
add_generation_prompt=True,
|
|
|
|
| 428 |
return_tensors="pt",
|
| 429 |
).to(device=model.device, dtype=torch.bfloat16)
|
| 430 |
|
|
|
|
| 431 |
streamer = TextIteratorStreamer(processor, timeout=30.0, skip_prompt=True, skip_special_tokens=True)
|
| 432 |
gen_kwargs = dict(
|
| 433 |
inputs,
|
|
|
|
| 435 |
max_new_tokens=max_new_tokens,
|
| 436 |
)
|
| 437 |
|
|
|
|
| 438 |
t = Thread(target=model.generate, kwargs=gen_kwargs)
|
| 439 |
t.start()
|
| 440 |
|
|
|
|
| 441 |
output = ""
|
| 442 |
for new_text in streamer:
|
| 443 |
output += new_text
|
|
|
|
| 449 |
|
| 450 |
|
| 451 |
|
|
|
|
|
|
|
|
|
|
| 452 |
examples = [
|
| 453 |
|
| 454 |
[
|
|
|
|
| 554 |
]
|
| 555 |
|
| 556 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 557 |
css = """
|
| 558 |
body {
|
| 559 |
background: linear-gradient(135deg, #667eea, #764ba2);
|
|
|
|
| 615 |
</p>
|
| 616 |
"""
|
| 617 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 618 |
with gr.Blocks(css=css, title="Vidraft-Gemma-3-27B") as demo:
|
| 619 |
gr.Markdown(title_html)
|
| 620 |
|
|
|
|
| 628 |
value=False,
|
| 629 |
info="Check to enable a SERPHouse web search before the chat reply"
|
| 630 |
)
|
| 631 |
+
# [์ค์] web_search_text๋ ์ฌ์ค์ ์ฌ์ฉ ์ ํจ (์๋์ถ์ถ๋ก ๊ฒ์)
|
| 632 |
web_search_text = gr.Textbox(
|
| 633 |
lines=1,
|
| 634 |
+
label="(Unused) Web Search Query",
|
| 635 |
+
placeholder="No direct input needed"
|
| 636 |
)
|
| 637 |
|
| 638 |
gr.Markdown("---")
|
|
|
|
| 653 |
value=2000,
|
| 654 |
)
|
| 655 |
|
| 656 |
+
gr.Markdown("<br><br>")
|
| 657 |
|
|
|
|
| 658 |
with gr.Column(scale=7):
|
| 659 |
chat = gr.ChatInterface(
|
| 660 |
fn=run,
|
|
|
|
| 673 |
system_prompt_box,
|
| 674 |
max_tokens_slider,
|
| 675 |
web_search_checkbox,
|
| 676 |
+
web_search_text, # ์ค์ ๋ก๋ ์ฌ์ฉ ์ํจ
|
| 677 |
],
|
| 678 |
stop_btn=False,
|
| 679 |
title="Vidraft-Gemma-3-27B",
|
|
|
|
| 687 |
with gr.Row(elem_id="examples_row"):
|
| 688 |
with gr.Column(scale=12, elem_id="examples_container"):
|
| 689 |
gr.Markdown("### Example Inputs (click to load)")
|
|
|
|
| 690 |
gr.Examples(
|
| 691 |
examples=examples,
|
| 692 |
+
inputs=[], # Gradio๊ฐ dataset์ ์ฐ๊ฒฐํ inputs๊ฐ ์์ผ๋ฏ๋ก ๋น ๋ฆฌ์คํธ
|
| 693 |
cache_examples=False
|
| 694 |
)
|
| 695 |
|