File size: 13,819 Bytes
4c30587 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 | from concurrent.futures import ThreadPoolExecutor
import os
from queue import Queue
import threading
from typing import Annotated as A
from tqdm import tqdm
from .func_call_shema import doc as D
from .helper import image_to_base64, timesec_hms
from concurrent.futures import ThreadPoolExecutor, as_completed
from collections import defaultdict
import threading
SUMMARIZE_WORKERS = 8
SEARCH_WORKERS = 8
VIDEO_GROUNDING_PROMPT = """
You are given a video.
Your task is to identify the time ranges in the video that are relevant to the user query.
Instructions:
- Carefully watch the video and consider both visual and audio information.
- Identify video segments where the content directly matches the user query.
- Each selected segment should be a continuous time range.
- If multiple relevant segments exist, list all of them.
- For each segment, provide a brief reason or description for why this segment was selected.
- If no relevant segment is found, output NOT FOUND.
- Do NOT describe unrelated parts of the video.
User query:
"{query}"
Output format:
HH:MM:SS - HH:MM:SS, <reason or description>
(one line per segment; if nothing is found, output only NOT FOUND)
"""
def video_inspect_tool(
_video_db,
_memory,
_openai_client,
question: A[str, D("The specific detailed question to ask about the video content during the specified time ranges. No need to add time ranges in the question.")],
time_ranges_hhmmss: A[list[tuple], D("A list of tuples containing start and end times in HH:MM:SS format, at most 10 intervals.")],
) -> str:
"""
Crop the video based on the time ranges and ask the model a detailed question about the cropped video clips.
Returns:
str: The model's response to the question. If no relevant content is found within the time range,
returns an error message: "Error: Cannot find corresponding result in the given time range."
"""
clip_queue = Queue(maxsize=16)
stop_signal = object()
results = []
results_lock = threading.Lock()
def producer():
for s_hms, e_hms in time_ranges_hhmmss:
start_time_sec,end_time_sec = timesec_hms(s_hms,out="float"),min(timesec_hms(e_hms,out="float"),_video_db.duration)
# audio_b64 = _video_db.cut_segment_to_audio(start_time_sec, end_time_sec)
frames_list, timestamps, actual_indices = _video_db.sample_frames_by_fps(start_time_sec,end_time_sec,2,300 if "gpt" not in _openai_client.model_name else 50)
video_context = []
def _proc_frame(args):
frame,timestamp = args
img64 = image_to_base64(frame, (640, 360), 75)
return [
{"type": "text", "text": f"[{timesec_hms(timestamp)}]"},
{"type": "image_url", "image_url": {"url": img64}}
]
print("Encoding frames to Base64...")
with ThreadPoolExecutor(max_workers=16) as pool:
# 使用 map 保持顺序,或者使用 as_completed 后再排序
results = list(tqdm(pool.map(_proc_frame, zip(frames_list,timestamps)), total=len(frames_list)))
for res in results:
video_context.extend(res)
clip_queue.put((start_time_sec, end_time_sec, video_context, None))
for _ in range(8):
clip_queue.put(stop_signal)
def consumer():
while True:
item = clip_queue.get()
if item is stop_signal:
break
start_time_sec, end_time_sec, video_context,audio_b64 = item
caption = _openai_client.chat(messages=[{
"role": "user",
"content":
video_context+[
# {
# "type": "input_audio",
# "input_audio": {
# "data": audio_b64,
# "format": "wav",
# },
# },
{
"type": "text",
"text": (
"Carefully watch the video. Pay close attention to the cause and sequence of events, "
"the details and movements of objects, and the actions and poses of people.\n\n"
"Based on your observations, answer the question using only information that can be "
"directly verified from the video.\n\n"
"When relevant, you MAY insert time anchors from the video into your answer "
"to support your reasoning. Time anchors must be in the format [HH:MM:SS] and should "
"correspond exactly to the moment shown in the video.\n\n"
"Do NOT invent timestamps. If you are uncertain about the exact time, omit the time anchor.\n\n"
"If no relevant content is found within the given time range, return exactly:\n"
"`Error: Cannot find corresponding result in the given time range.`\n\n"
f"Question: {question}\n"
)
},
]
}])
caption = _openai_client.chat(messages=[{
"role": "user",
"content": [
{"type": "text", "text": f"""
You are given:
1) A block of text that may contain multiple timestamps in the format [HH:MM:SS]
2) A time offset in the format HH:MM:SS
Task:
- Shift EVERY timestamp in the text by the given offset.
- A timestamp [HH:MM:SS] represents a time duration, not a clock time.
- The offset should be ADDED to each timestamp.
- Properly handle carry-over for seconds and minutes.
- Preserve the original [HH:MM:SS] format (always two digits per field).
- Do NOT modify any part of the text other than the timestamps.
- Do NOT add, remove, or rephrase any text.
If the text contains no timestamps, return the original text unchanged.
Text:
{caption}
Time offset:
{timesec_hms(start_time_sec)}
Output only the modified text. Do not include any other content.
"""
},
]
}])
_memory.add(
{
"role": "user",
"content": (
f"From {timesec_hms(start_time_sec)} to {timesec_hms(end_time_sec)}, "
f"answering the question '{question}', the video clip shows: {caption}"
)
},
user_id="root",
metadata={
"start_time": start_time_sec,
"end_time": end_time_sec,
"question": question,
"mem_type": "video_info"
}
)
# 线程安全地把结果 append 到列表
with results_lock:
results.append({
"start_time": start_time_sec,
"end_time": end_time_sec,
"caption": caption
})
prod_thread = threading.Thread(target=producer)
prod_thread.start()
consumer_threads = []
for _ in range(8):
t = threading.Thread(target=consumer)
t.start()
consumer_threads.append(t)
prod_thread.join()
for t in consumer_threads:
t.join()
return "\n".join([f"From {timesec_hms(msg['start_time'])} to {timesec_hms(msg['end_time'])}, {msg['caption']}" for msg in results])
def _search_single_query(_memory, q, top_k):
result = _memory.search(
q,
user_id="root",
filters={"mem_type": "video_info"},
limit=top_k,
)
video_hits = []
contextual = []
for item in result.get("results", []):
metadata = item.get("metadata", {}) or {}
mem = item.get("memory", "").replace("\n", " ").strip()
st = metadata.get("start_time")
ed = metadata.get("end_time")
if st is not None and ed is not None and mem:
video_hits.append((st, ed, q, mem))
elif mem:
contextual.append(mem)
return video_hits, contextual
def _summarize_and_add(
_openai_client,
_memory,
start_time,
end_time,
q,
summarize_query,
texts,
):
merged = " ".join(texts)
summary = summarize_with_gpt(
_openai_client=_openai_client,
query=q,
summarize_query=summarize_query,
text=merged,
)
if not summary or "NOT FOUND" in summary.strip():
return None
line = (
f"From {timesec_hms(start_time)} to {timesec_hms(end_time)}, "
f"(query: {q}) {summary}"
)
_memory.add(
{
"role": "user",
"content": (
f"From {timesec_hms(start_time)} to {timesec_hms(end_time)}, "
f"related to query '{q}', {summary}"
),
},
user_id="root",
metadata={
"start_time": start_time,
"end_time": end_time,
"question": q,
"mem_type": "video_info",
}
)
return line
def memory_search_tool(
_video_db,
_memory,
_openai_client,
query: A[str, D("Short text queries for memory search; use ';' to separate up to 5 queries.")],
summarize_query: A[
str,
D("Query used during LLM summarization to filter and extract only useful information.")
],
top_k: A[int, D("Maximum number of relevant memory to return. Default is 10.")] = 10,
) -> str:
"""
Retrieve relevant video memory using textual search queries and perform
query-conditioned summarization to extract task-relevant evidence by LLM.
This function first retrieves candidate memory entries based on one or more
search queries. It then applies a separate summarization query to filter out
irrelevant content and generate concise, question-oriented summaries.
Returns:
str:
A newline-separated list of video evidence in the format:
"From HH:MM:SS to HH:MM:SS, <summary>".
Only summaries that are relevant to the summarization query are included.
If no relevant evidence is found, an empty string is returned.
"""
queries = [q.strip() for q in query.split(";") if q.strip()]
if not queries:
return ""
video_grouped = defaultdict(list)
contextual_memories = []
# ===== 1. parallel search =====
with ThreadPoolExecutor(
max_workers=min(SEARCH_WORKERS, len(queries))
) as executor:
futures = [
executor.submit(_search_single_query, _memory, q, top_k)
for q in queries
]
for fut in as_completed(futures):
hits, context = fut.result()
contextual_memories.extend(context)
for st, ed, q, mem in hits:
video_grouped[(st, ed, q)].append(mem)
if not video_grouped and not contextual_memories:
return ""
sections = []
if video_grouped:
sections.append("[Video Evidence]")
# ===== 2. parallel summarize + add =====
with ThreadPoolExecutor(max_workers=SUMMARIZE_WORKERS) as executor:
sum_futures = []
for (st, ed, search_q), texts in video_grouped.items():
sum_futures.append(
executor.submit(
_summarize_and_add,
_openai_client,
_memory,
st,
ed,
search_q,
summarize_query,
texts,
)
)
for fut in as_completed(sum_futures):
line = fut.result()
if line:
sections.append(line)
# ===== 3. contextual memory =====
if contextual_memories:
sections.append("\n[Contextual Memory]")
sections.extend(contextual_memories)
return "\n".join(sections)
def summarize_with_gpt(_openai_client, query: str,summarize_query, text: str) -> str:
prompt = f"""
You are summarizing retrieved video memory.
Search query (for retrieval):
{query}
Filtering / summarization query (IMPORTANT):
{summarize_query}
Below are memory snippets retrieved from the same video segment.
Only keep information that is directly useful for answering the filtering query.
Rules:
- If the content does NOT help answer the filtering query, return an empty string.
- Be concise and factual.
- Do NOT speculate.
- If useful, produce ONE concise sentence.
Memory snippets:
{text}
"""
return _openai_client.chat(messages=[{
"role": "user",
"content": [
{"type": "text", "text": prompt},
]
}])
class StopException(Exception):
"""
Stop Execution by raising this exception (Signal that the task is Finished).
"""
def finish(_video_db,
_memory,
_openai_client,answer: A[str, D("Answer to the user's question.")]) -> None:
"""Call this function after confirming the answer of the user's question, and finish the conversation."""
raise StopException(answer)
|