File size: 12,577 Bytes
5286113 f8dffc3 5286113 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 |
import os, io, base64, json, tempfile
from pathlib import Path
from typing import Any, Dict, List, Optional, Literal
from PIL import Image
import google.generativeai as genai
from langchain_core.tools import tool
# ======================== CONFIG & CORE ========================
def _configure() -> str:
api_key = os.getenv("GOOGLE_API_KEY") or os.getenv("GENAI_API_KEY")
if not api_key:
raise RuntimeError("Missing GOOGLE_API_KEY (or GENAI_API_KEY) in environment")
genai.configure(api_key=api_key)
return api_key
def _clean_json_text(s: str) -> str:
s = s.strip()
if s.startswith("```"):
s = s.strip("`").replace("json", "", 1).strip()
start = s.find("{")
end = s.rfind("}")
if start != -1 and end != -1 and end > start:
return s[start:end+1]
return s
def _call_model(parts: List[Any], temperature: float, model_name: Optional[str] = None) -> Dict[str, Any]:
"""
Единая точка вызова модели. Возвращает dict с ключом "answer".
"""
MODEL_NAME = model_name or os.getenv("GEMMA_MODEL", "gemma-3-27b-it")
model = genai.GenerativeModel(MODEL_NAME)
resp = model.generate_content(parts, generation_config={"temperature": temperature})
text = (getattr(resp, "text", None) or "").strip()
try:
return json.loads(_clean_json_text(text))
except Exception:
fixer = genai.GenerativeModel(MODEL_NAME)
fix_prompt = (
"Convert the following text into STRICT valid JSON matching schema {\"answer\": string}. "
"Return ONLY JSON, no extra text:\n" + text
)
fix_resp = fixer.generate_content([{"text": fix_prompt}])
return json.loads(_clean_json_text((getattr(fix_resp, "text", "") or "").strip()))
# ======================== VIDEO HELPERS (OpenCV-only) ========================
_VIDEO_QA_PROMPT = (
"You will be given ONE video and a question about its visual content.\n"
"Answer STRICTLY and CONCISELY based only on what is visible/audible in the provided video.\n"
"If the video does not contain enough information, reply 'not enough information'.\n"
"Return ONLY valid JSON with the schema:\n"
"{\"answer\": string}\n"
)
def _uniform_sample_paths(paths: List[Path], k: int) -> List[Path]:
n = len(paths)
if n <= k:
return paths
idxs = [round(i*(n-1)/(k-1)) for i in range(k)]
return [paths[i] for i in idxs]
def _ensure_png_bytes(img: Image.Image, max_pixels: int = 25_000_000) -> bytes:
w, h = img.size
if w * h > max_pixels:
scale = (max_pixels / (w * h)) ** 0.5
img = img.resize((max(1, int(w*scale)), max(1, int(h*scale))), Image.LANCZOS)
buf = io.BytesIO()
img.save(buf, format="PNG", optimize=True)
return buf.getvalue()
def _image_bytes_to_part(img_bytes: bytes, mime: str = "image/png") -> Dict[str, Any]:
return {"mime_type": mime, "data": base64.b64encode(img_bytes).decode("utf-8")}
def _extract_frames_cv2(video_path: str, out_dir: Path, fps: float, start_s: float, duration_s: Optional[float]) -> List[Path]:
"""
Извлекаем кадры через OpenCV (без системного ffmpeg).
Требует: pip install opencv-python
"""
import cv2
out_dir.mkdir(parents=True, exist_ok=True)
cap = cv2.VideoCapture(video_path)
if not cap.isOpened():
raise RuntimeError("OpenCV cannot open video")
in_fps = cap.get(cv2.CAP_PROP_FPS) or 30.0
total_frames = cap.get(cv2.CAP_PROP_FRAME_COUNT) or 0
total_ms = (total_frames / in_fps) * 1000.0 if total_frames and in_fps else None
start_ms = max(0.0, float(start_s) * 1000.0)
end_ms = start_ms + float(duration_s) * 1000.0 if duration_s is not None else (total_ms or start_ms + 30_000.0)
step_ms = 1000.0 / max(0.001, fps) # период семплинга по ms
t = start_ms
idx = 0
saved: List[Path] = []
while t <= end_ms:
cap.set(cv2.CAP_PROP_POS_MSEC, t)
ok, frame = cap.read()
if not ok:
break
fp = out_dir / f"{idx:06d}.jpg"
# JPEG сохраняем без ffmpeg
ok = cv2.imwrite(str(fp), frame)
if ok:
saved.append(fp)
idx += 1
t += step_ms
cap.release()
if not saved:
raise RuntimeError("No frames extracted (OpenCV).")
return saved
def _frames_to_image_parts(frame_paths: List[Path], max_images: int) -> List[Dict[str, Any]]:
"""
Прореживаем кадры до <= max_images и упаковываем как inline-изображения.
"""
frame_paths = _uniform_sample_paths(frame_paths, k=max_images)
out: List[Dict[str, Any]] = []
for fp in frame_paths:
img = Image.open(fp)
img_bytes = _ensure_png_bytes(img)
out.append(_image_bytes_to_part(img_bytes, "image/png"))
return out
def _download_youtube_to_mp4(youtube_url: str, out_path: str) -> str:
"""
Скачиваем YouTube через библиотеку yt_dlp (без системного ffmpeg).
Требует: pip install yt-dlp
Стараемся выбрать прогрессивный MP4 (single file), чтобы не потребовался mux.
"""
from yt_dlp import YoutubeDL
ydl_opts = {
# выбираем ЛУЧШИЙ одиночный файл, предпочитая MP4 (без mux/ffmpeg)
"format": "b[ext=mp4]/b",
"outtmpl": out_path,
"noprogress": True,
"quiet": True,
"nocheckcertificate": True,
}
with YoutubeDL(ydl_opts) as ydl:
info = ydl.extract_info(youtube_url, download=True)
# yt-dlp может игнорировать outtmpl при некоторых шаблонах — подстрахуемся
fn = ydl.prepare_filename(info)
# Если получили другой путь, перенесём
src = Path(fn)
dst = Path(out_path)
if src.resolve() != dst.resolve():
dst.parent.mkdir(parents=True, exist_ok=True)
src.replace(dst)
return str(dst)
def _get_client(api_key: Optional[str]):
"""
Опционально: новый Google GenAI SDK (google-genai) для Files API в 'auto' режиме.
Если нет — вернём None.
"""
try:
from google import genai as ggenai # новый пакет "google-genai"
return ggenai.Client(api_key=api_key)
except Exception:
return None
def _video_part_from_youtube(url: str) -> Dict[str, Any]:
"""Для mode='auto': передаём YouTube как file_data без скачивания."""
return {"file_data": {"file_uri": url}}
def _video_part_from_file(path: str, api_key: Optional[str]) -> Dict[str, Any]:
"""
Для mode='auto': загружаем локальный файл в Files API.
"""
if not os.path.exists(path):
raise FileNotFoundError(f"Video not found: {path}")
client = _get_client(api_key)
if client is not None and hasattr(client, "files"):
try:
f = client.files.upload(file=path)
return {"file_data": {"file_uri": f.uri, "mime_type": getattr(f, "mime_type", None) or "video/mp4"}}
except Exception:
pass
f = genai.upload_file(path=path)
file_uri = getattr(f, "uri", None) or getattr(f, "file_uri", None)
mime = getattr(f, "mime_type", None) or "video/mp4"
return {"file_data": {"file_uri": file_uri, "mime_type": mime}}
# ======================== VIDEO QA TOOL (OpenCV frames по умолчанию) ========================
@tool
def video_qa_gemma(
question: str,
youtube_url: Optional[str] = None,
video_path: Optional[str] = None,
temperature: float = 0.2,
model_name: Optional[str] = None,
mode: Literal["frames", "auto"] = "auto", # по умолчанию безопасный режим кадров (OpenCV) #default frames
fps: float = 0.8, # 0.8 * 30s ≈ 24 кадров
start_s: float = 0.0,
duration_s: Optional[float] = 30.0, # держим сегмент коротким
max_images: int = 24, # < 32 — жёсткая крышка
) -> str:
"""
Answer questions about the visual content of a video (YouTube URL or local file).
Args:
question: Natural language question about the video.
youtube_url: Link to a YouTube video (exclusive with video_path).
video_path: Local path to a video file.
mode: "frames" (default, extracts ≤max_images frames with OpenCV) or "auto" (send whole video).
fps/start_s/duration_s: Frame sampling parameters in "frames" mode.
max_images: Max number of frames (<32). Default 24.
Returns:
JSON string: {"answer": "..."} (or "not enough information").
Notes:
- Provide exactly ONE of youtube_url or video_path.
- Use "frames" mode to avoid API errors on models with image limits.
"""
import json as _json
try:
api_key = _configure()
if bool(youtube_url) == bool(video_path):
return _json.dumps({"error": "Provide exactly ONE of youtube_url or video_path"})
if mode == "auto":
# Без OpenCV: отдаём видео целиком (иногда API внутри раздувает до >32 изображений).
if youtube_url:
video_part = _video_part_from_youtube(youtube_url)
else:
video_part = _video_part_from_file(video_path, api_key)
parts = [video_part, {"text": _VIDEO_QA_PROMPT + "\nQuestion: " + question.strip()}]
data = _call_model(parts, temperature, model_name=model_name)
else:
# OpenCV: извлекаем кадры и отправляем как <= max_images изображений
tmp_video_path = None
if youtube_url and not video_path:
with tempfile.TemporaryDirectory(prefix="yt_") as td:
tmp_video_path = str(Path(td) / "video.mp4")
_download_youtube_to_mp4(youtube_url, tmp_video_path)
# внутри with мы не можем вернуть, поэтому делаем обработку ниже в том же блоке
frame_dir = Path(td) / "frames"
files = _extract_frames_cv2(tmp_video_path, frame_dir, fps=fps, start_s=start_s, duration_s=duration_s)
img_parts = _frames_to_image_parts(files, max_images=max_images)
parts = img_parts + [{"text": _VIDEO_QA_PROMPT + "\nQuestion: " + question.strip()}]
data = _call_model(parts, temperature, model_name=model_name)
# выходим из with — файлы удалятся
answer = data["answer"] if isinstance(data, dict) and "answer" in data else None
if not isinstance(answer, str):
answer = str(answer) if answer is not None else "not enough information"
return _json.dumps({"answer": answer})
# локальный файл (или если youtube уже скачали и вышли return выше)
frame_dir = Path(tempfile.mkdtemp(prefix="frames_"))
try:
src_video = video_path if video_path else tmp_video_path
files = _extract_frames_cv2(src_video, frame_dir, fps=fps, start_s=start_s, duration_s=duration_s)
img_parts = _frames_to_image_parts(files, max_images=max_images)
parts = img_parts + [{"text": _VIDEO_QA_PROMPT + "\nQuestion: " + question.strip()}]
data = _call_model(parts, temperature, model_name=model_name)
finally:
# подчистим временные файлы
for p in frame_dir.glob("*"):
try:
p.unlink()
except Exception:
pass
try:
frame_dir.rmdir()
except Exception:
pass
answer = data["answer"] if isinstance(data, dict) and "answer" in data else None
if not isinstance(answer, str):
answer = str(answer) if answer is not None else "not enough information"
return _json.dumps({"answer": answer})
except Exception as e:
return _json.dumps({"error": str(e)}) |