File size: 112,190 Bytes
120e73b | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 1497 1498 1499 1500 1501 1502 1503 1504 1505 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552 1553 1554 1555 1556 1557 1558 1559 1560 1561 1562 1563 1564 1565 1566 1567 1568 1569 1570 1571 1572 1573 1574 1575 1576 1577 1578 1579 1580 1581 1582 1583 1584 1585 1586 1587 1588 1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599 1600 1601 1602 1603 1604 1605 1606 1607 1608 1609 1610 1611 1612 1613 1614 1615 1616 1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633 1634 1635 1636 1637 1638 1639 1640 1641 1642 1643 1644 1645 1646 1647 1648 1649 1650 1651 1652 1653 1654 1655 1656 1657 1658 1659 1660 1661 1662 1663 1664 1665 1666 1667 1668 1669 1670 1671 1672 1673 1674 1675 1676 1677 1678 1679 1680 1681 1682 1683 1684 1685 1686 1687 1688 1689 1690 1691 1692 1693 1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706 1707 1708 1709 1710 1711 1712 1713 1714 1715 1716 1717 1718 1719 1720 1721 1722 1723 1724 1725 1726 1727 1728 1729 1730 1731 1732 1733 1734 1735 1736 1737 1738 1739 1740 1741 1742 1743 1744 1745 1746 1747 1748 1749 1750 1751 1752 1753 1754 1755 1756 1757 1758 1759 1760 1761 1762 1763 1764 1765 1766 1767 1768 1769 1770 1771 1772 1773 1774 1775 1776 1777 1778 1779 1780 1781 1782 1783 1784 1785 1786 1787 1788 1789 1790 1791 1792 1793 1794 1795 1796 1797 1798 1799 1800 1801 1802 1803 1804 1805 1806 1807 1808 1809 1810 1811 1812 1813 1814 1815 1816 1817 1818 1819 1820 1821 1822 1823 1824 1825 1826 1827 1828 1829 1830 1831 1832 1833 1834 1835 1836 1837 1838 1839 1840 1841 1842 1843 1844 1845 1846 1847 1848 1849 1850 1851 1852 1853 1854 1855 1856 1857 1858 1859 1860 1861 1862 1863 1864 1865 1866 1867 1868 1869 1870 1871 1872 1873 1874 1875 1876 1877 1878 1879 1880 1881 1882 1883 1884 1885 1886 1887 1888 1889 1890 1891 1892 1893 1894 1895 1896 1897 1898 1899 1900 1901 1902 1903 1904 1905 1906 1907 1908 1909 1910 1911 1912 1913 1914 1915 1916 1917 1918 1919 1920 1921 1922 1923 1924 1925 1926 1927 1928 1929 1930 1931 1932 1933 1934 1935 1936 1937 1938 1939 1940 1941 1942 1943 1944 1945 1946 1947 1948 1949 1950 1951 1952 1953 1954 1955 1956 1957 1958 1959 1960 1961 1962 1963 1964 1965 1966 1967 1968 1969 1970 1971 1972 1973 1974 1975 1976 1977 1978 1979 1980 1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021 2022 2023 2024 2025 2026 2027 2028 2029 2030 2031 2032 2033 2034 2035 2036 2037 2038 2039 2040 2041 2042 2043 2044 2045 2046 2047 2048 2049 2050 2051 2052 2053 2054 2055 2056 2057 2058 2059 2060 2061 2062 2063 2064 2065 2066 2067 2068 2069 2070 2071 2072 2073 2074 2075 2076 2077 2078 2079 2080 2081 2082 2083 2084 2085 2086 2087 2088 2089 2090 2091 2092 2093 2094 2095 2096 2097 2098 2099 2100 2101 2102 2103 2104 2105 2106 2107 2108 2109 2110 2111 2112 2113 2114 2115 2116 2117 2118 2119 2120 2121 2122 2123 2124 2125 2126 2127 2128 2129 2130 2131 2132 2133 2134 2135 2136 2137 2138 2139 2140 2141 2142 2143 2144 2145 2146 2147 2148 2149 2150 2151 2152 2153 2154 2155 2156 2157 2158 2159 2160 2161 2162 2163 2164 2165 2166 2167 2168 2169 2170 2171 2172 2173 2174 2175 2176 2177 2178 2179 2180 2181 2182 2183 2184 2185 2186 2187 2188 2189 2190 2191 2192 2193 2194 2195 2196 2197 2198 2199 2200 2201 2202 2203 2204 2205 2206 2207 2208 2209 2210 2211 2212 2213 2214 2215 2216 2217 2218 2219 2220 2221 2222 2223 2224 2225 2226 2227 2228 2229 2230 2231 2232 2233 2234 2235 2236 2237 2238 2239 2240 2241 2242 2243 2244 2245 2246 2247 2248 2249 2250 2251 2252 2253 2254 2255 2256 2257 2258 2259 2260 2261 2262 2263 2264 2265 2266 2267 2268 2269 2270 2271 2272 2273 2274 2275 2276 2277 2278 2279 2280 2281 2282 2283 2284 2285 2286 2287 2288 2289 2290 2291 2292 2293 2294 2295 2296 2297 2298 2299 2300 2301 2302 2303 2304 2305 2306 2307 2308 2309 2310 2311 2312 2313 2314 2315 2316 2317 2318 2319 2320 2321 2322 2323 2324 2325 2326 2327 2328 2329 2330 2331 2332 2333 2334 2335 2336 2337 2338 2339 2340 2341 2342 2343 2344 2345 2346 2347 2348 2349 2350 2351 2352 2353 2354 2355 2356 2357 2358 2359 2360 2361 2362 2363 2364 2365 2366 2367 2368 2369 2370 2371 2372 2373 2374 2375 2376 2377 2378 2379 2380 2381 2382 2383 2384 2385 2386 2387 2388 2389 2390 2391 2392 2393 2394 2395 2396 2397 2398 2399 2400 2401 2402 2403 2404 2405 2406 2407 2408 2409 2410 2411 2412 2413 2414 2415 2416 2417 2418 2419 2420 2421 2422 2423 2424 2425 2426 2427 2428 2429 2430 2431 2432 2433 2434 2435 2436 2437 2438 2439 2440 2441 2442 2443 2444 2445 2446 2447 2448 2449 2450 2451 2452 2453 2454 2455 2456 2457 2458 2459 2460 2461 2462 2463 2464 2465 2466 2467 2468 2469 2470 2471 2472 2473 2474 2475 2476 2477 2478 2479 2480 2481 2482 2483 2484 2485 2486 2487 2488 2489 2490 2491 2492 2493 2494 2495 2496 2497 2498 2499 2500 2501 2502 2503 2504 2505 2506 2507 2508 2509 2510 2511 2512 2513 2514 2515 2516 2517 2518 2519 2520 2521 2522 2523 2524 2525 2526 2527 2528 2529 2530 2531 2532 2533 2534 2535 2536 2537 2538 2539 2540 2541 2542 2543 2544 2545 2546 2547 2548 2549 2550 2551 2552 2553 2554 2555 2556 2557 2558 2559 2560 2561 2562 2563 2564 2565 2566 2567 2568 2569 2570 2571 2572 2573 2574 2575 2576 2577 2578 2579 2580 2581 2582 2583 2584 2585 2586 2587 2588 2589 2590 2591 2592 2593 2594 2595 2596 2597 2598 2599 2600 2601 2602 2603 2604 2605 2606 2607 2608 2609 2610 2611 2612 2613 2614 2615 2616 2617 2618 2619 2620 2621 2622 2623 2624 2625 2626 2627 2628 2629 2630 2631 2632 2633 2634 2635 2636 2637 2638 2639 2640 2641 2642 2643 2644 2645 2646 2647 2648 2649 2650 2651 2652 | #!/usr/bin/env python3
"""
Video DeepResearch ๅ
ฌๅ
ฑๅทฅๅ
ทๆจกๅใ
่พๅ
ฅ:
- `config.py` ๅฏผๅบ็ๆจกๅ็ซฏ็นใVertex ้
็ฝฎใๆฌๅฐๆฃ็ดขๆๅกๅฐๅไธไปทๆ ผๅๆฐใ
- ่ง้ขๅธงใ่ฃๅชๅพใLLM messagesใๆ็ดข่ฏทๆฑไธๅ้ถๆฎตไธญ้ด็ปๆใ
ๅค็:
- ๆไพๅธง้ๆ ทใๅพ็็ผ็ ใๆ็ดข่ฎฟ้ฎใ่พๅบๆธ
ๆดใtoken ็ป่ฎก็ญ้็จ่ฝๅใ
- ๅจ `LLMClient` ไธญๅฎ็ฐๅค endpoint ่ฝฎ่ฏขใ429 ๅทๅดๅๆขใVertex ๅ็่ฐ็จไธ OAuth2 ่ฎค่ฏใ
- ๆฏๆโไธไธช project ๅฏนๅบไธไธช service account jsonโ็ Vertex ๅญ่ฏๆฑ ๏ผๅนถๆ URL ้ๆฉๅฏนๅบๅญ่ฏใ
่พๅบ:
- ไธบ Phase1/Phase2 ๆไพ็ปไธ็ๅทฅๅ
ทๅฝๆฐใๆ็ดขๅฝๆฐใtoken ็ป่ฎก็ปๆๅ `LLMClient`ใ
- ่ฟๅ่ง่ๅ็ๆจกๅๅๅบใๆ็ดข็ปๆใๅพ็่ทฏๅพไธ token ไฝฟ็จไฟกๆฏใ
"""
import asyncio
import aiohttp
import base64
import hashlib
import itertools
import json
import math
import os
import random
import re
import subprocess
import time
import numpy as np
from PIL import Image
from pathlib import Path
from typing import Dict, List, Tuple, Optional, Any, Set
from dataclasses import dataclass, field
from config import (
API_ENDPOINTS, DEFAULT_MODEL, API_KEY,
WEB_SEARCH_ADDRESS, WEB_SEARCH_CONFIG,
SERPER_API_KEY, MOCK_SEARCH, IMAGE_SEARCH_CACHE_FILE,
DEFAULT_TEMPERATURE, DEFAULT_TOP_P, DEFAULT_MAX_TOKENS,
DEFAULT_TIMEOUT, DEFAULT_MAX_RETRIES,
VERTEX_MIN_REQUEST_INTERVAL_SECONDS,
VERTEX_RATE_LIMIT_COOLDOWN_SECONDS,
VERTEX_REQUEST_JITTER_SECONDS,
TOKEN_PRICING, THINKING_TOKEN_PRICING,
DEFAULT_INPUT_PRICE_PER_1M, DEFAULT_OUTPUT_PRICE_PER_1M,
OSS_ACCESS_KEY_ID, OSS_ACCESS_KEY_SECRET,
OSS_ENDPOINT, OSS_BUCKET_NAME, OSS_UPLOAD_PREFIX,
SEARCH_CROP_MAX_SIZE, SEARCH_CROP_JPEG_QUALITY,
IMAGE_SEARCH_SUMMARIZE_SERPER, IMAGE_SEARCH_SUMMARIZER_ADDRESS,
IMAGE_SEARCH_SUMMARIZER_MODEL, IMAGE_SEARCH_SUMMARIZER_MAX_RESULTS,
IMAGE_SEARCH_SUMMARIZER_MAX_TOKENS,
IMAGE_SEARCH_MODE, GATEWAY_URL, GATEWAY_USERNAME, GATEWAY_USERID, GATEWAY_TOKEN,
IMAGE_SEARCH_ALLOW_BASE64_FALLBACK,
)
from config import BBOX_CONFIGS
import config
# โโ MARS-style web search config โโ
MARS_RETRIEVAL_ADDRESS = getattr(config, 'MARS_RETRIEVAL_ADDRESS', '')
MARS_SUMMARIZER_ADDRESS = getattr(config, 'MARS_SUMMARIZER_ADDRESS', '')
MARS_RETRIEVAL_TOPK = getattr(config, 'MARS_RETRIEVAL_TOPK', 3)
MARS_SUMMARIZER_MODEL = getattr(config, 'MARS_SUMMARIZER_MODEL', '')
MARS_WEB_SEARCH_MODE = getattr(config, 'MARS_WEB_SEARCH_MODE', 'serper')
MARS_RETRIEVAL_TIMEOUT = getattr(config, 'MARS_RETRIEVAL_TIMEOUT', 120)
MARS_RETRIEVAL_CONCURRENCY = getattr(config, 'MARS_RETRIEVAL_CONCURRENCY', 0)
IMAGE_SEARCH_SUMMARIZE_SERPER = getattr(config, 'IMAGE_SEARCH_SUMMARIZE_SERPER', True)
IMAGE_SEARCH_SUMMARIZER_ADDRESS = getattr(config, 'IMAGE_SEARCH_SUMMARIZER_ADDRESS', MARS_SUMMARIZER_ADDRESS)
IMAGE_SEARCH_SUMMARIZER_MODEL = getattr(config, 'IMAGE_SEARCH_SUMMARIZER_MODEL', MARS_SUMMARIZER_MODEL)
IMAGE_SEARCH_SUMMARIZER_MAX_RESULTS = getattr(config, 'IMAGE_SEARCH_SUMMARIZER_MAX_RESULTS', 5)
IMAGE_SEARCH_SUMMARIZER_MAX_TOKENS = getattr(config, 'IMAGE_SEARCH_SUMMARIZER_MAX_TOKENS', 512)
GCP_PROJECT_ID = getattr(config, 'GCP_PROJECT_ID', '')
GCP_LOCATION = getattr(config, 'GCP_LOCATION', '')
GCP_SERVICE_ACCOUNT_KEY = getattr(config, 'GCP_SERVICE_ACCOUNT_KEY', '')
VERTEX_CREDENTIALS_POOL = getattr(config, 'VERTEX_CREDENTIALS_POOL', [])
# Import phase-specific system prompts from prompts.py
from prompts import PHASE1_SYSTEM_PROMPT, PHASE2_SYSTEM_PROMPT, SYSTEM_PROMPT_BASE
# Legacy alias โ kept for backward compatibility if any external code references it
SYSTEM_PROMPT = PHASE1_SYSTEM_PROMPT
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
# Token Tracking & Cost Estimation
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
def _get_pricing(model: str) -> Tuple[float, float, float]:
"""Get (input_price, output_price, thinking_price) per 1M tokens for a model.
Matches model name by substring (e.g., 'gemini-2.5-flash' matches
'gemini-2.5-flash-preview-05-20').
Returns prices in USD per 1M tokens.
"""
input_price = DEFAULT_INPUT_PRICE_PER_1M
output_price = DEFAULT_OUTPUT_PRICE_PER_1M
thinking_price = None # None means use output_price
model_lower = model.lower()
# Find best match (longest matching key wins)
best_match = ""
for pattern, (inp, outp) in TOKEN_PRICING.items():
if pattern.lower() in model_lower and len(pattern) > len(best_match):
best_match = pattern
input_price = inp
output_price = outp
# Check thinking token pricing
for pattern, tp in THINKING_TOKEN_PRICING.items():
if pattern.lower() in model_lower:
thinking_price = tp
break
if thinking_price is None:
thinking_price = output_price
return input_price, output_price, thinking_price
@dataclass
class TokenUsage:
"""Token usage for a single LLM call."""
prompt_tokens: int = 0
completion_tokens: int = 0
total_tokens: int = 0
# Some APIs (e.g., Gemini thinking mode) separate thinking tokens
thinking_tokens: int = 0
# Cache-related (some APIs report cached token counts)
cached_tokens: int = 0
def to_dict(self) -> Dict[str, int]:
return {
"prompt_tokens": self.prompt_tokens,
"completion_tokens": self.completion_tokens,
"total_tokens": self.total_tokens,
"thinking_tokens": self.thinking_tokens,
"cached_tokens": self.cached_tokens,
}
@staticmethod
def from_api_response(usage_data: Dict[str, Any]) -> "TokenUsage":
"""Parse token usage from OpenAI-compatible API response.
Handles various API formats:
- Standard: {prompt_tokens, completion_tokens, total_tokens}
- Gemini extended: {prompt_tokens, completion_tokens, total_tokens,
completion_tokens_details: {reasoning_tokens: N}}
- Some APIs: {input_tokens, output_tokens}
"""
if not usage_data:
return TokenUsage()
prompt = usage_data.get("prompt_tokens", 0) or usage_data.get("input_tokens", 0) or 0
completion = usage_data.get("completion_tokens", 0) or usage_data.get("output_tokens", 0) or 0
total = usage_data.get("total_tokens", 0) or (prompt + completion)
# Extract thinking/reasoning tokens if available
thinking = 0
details = usage_data.get("completion_tokens_details", {})
if isinstance(details, dict):
thinking = details.get("reasoning_tokens", 0) or details.get("thinking_tokens", 0) or 0
# Some Gemini APIs put it at top level
if not thinking:
thinking = usage_data.get("reasoning_tokens", 0) or usage_data.get("thinking_tokens", 0) or 0
# Cached tokens
cached = 0
prompt_details = usage_data.get("prompt_tokens_details", {})
if isinstance(prompt_details, dict):
cached = prompt_details.get("cached_tokens", 0) or 0
if not cached:
cached = usage_data.get("cached_tokens", 0) or 0
return TokenUsage(
prompt_tokens=prompt,
completion_tokens=completion,
total_tokens=total,
thinking_tokens=thinking,
cached_tokens=cached,
)
@dataclass
class TokenTracker:
"""Tracks token usage across multiple LLM calls for a single data entry.
Accumulates prompt_tokens, completion_tokens, thinking_tokens and
computes estimated cost based on model pricing.
"""
model: str = ""
total_prompt_tokens: int = 0
total_completion_tokens: int = 0
total_thinking_tokens: int = 0
total_cached_tokens: int = 0
num_calls: int = 0
call_details: List[Dict[str, Any]] = field(default_factory=list)
def add(self, usage: TokenUsage, call_label: str = ""):
"""Add token usage from one LLM call."""
self.total_prompt_tokens += usage.prompt_tokens
self.total_completion_tokens += usage.completion_tokens
self.total_thinking_tokens += usage.thinking_tokens
self.total_cached_tokens += usage.cached_tokens
self.num_calls += 1
self.call_details.append({
"label": call_label,
**usage.to_dict(),
})
@property
def total_tokens(self) -> int:
return self.total_prompt_tokens + self.total_completion_tokens
def estimate_cost(self, model: str = "") -> Dict[str, float]:
"""Estimate cost in USD based on model pricing.
Returns dict with input_cost, output_cost, thinking_cost, total_cost.
"""
m = model or self.model
input_price, output_price, thinking_price = _get_pricing(m)
input_cost = self.total_prompt_tokens * input_price / 1_000_000
# completion_tokens includes thinking_tokens for some APIs,
# so we separate them for pricing
non_thinking_completion = max(0, self.total_completion_tokens - self.total_thinking_tokens)
output_cost = non_thinking_completion * output_price / 1_000_000
thinking_cost = self.total_thinking_tokens * thinking_price / 1_000_000
return {
"input_cost_usd": round(input_cost, 6),
"output_cost_usd": round(output_cost, 6),
"thinking_cost_usd": round(thinking_cost, 6),
"total_cost_usd": round(input_cost + output_cost + thinking_cost, 6),
}
def to_dict(self, model: str = "") -> Dict[str, Any]:
"""Export full tracking info as a dict."""
cost = self.estimate_cost(model)
return {
"num_llm_calls": self.num_calls,
"total_prompt_tokens": self.total_prompt_tokens,
"total_completion_tokens": self.total_completion_tokens,
"total_thinking_tokens": self.total_thinking_tokens,
"total_cached_tokens": self.total_cached_tokens,
"total_tokens": self.total_tokens,
"estimated_cost": cost,
"call_details": self.call_details,
}
def summary_str(self, model: str = "") -> str:
"""Human-readable one-line summary."""
cost = self.estimate_cost(model)
return (
f"calls={self.num_calls} "
f"prompt={self.total_prompt_tokens:,} "
f"completion={self.total_completion_tokens:,} "
f"thinking={self.total_thinking_tokens:,} "
f"total={self.total_tokens:,} "
f"cost=${cost['total_cost_usd']:.4f}"
)
class GlobalTokenStats:
"""Thread-safe aggregator for token stats across all entries."""
def __init__(self, model: str = ""):
self.model = model
self.total_prompt_tokens = 0
self.total_completion_tokens = 0
self.total_thinking_tokens = 0
self.total_cached_tokens = 0
self.total_calls = 0
self.total_entries = 0
self._lock = asyncio.Lock()
async def add(self, tracker: TokenTracker):
async with self._lock:
self.total_prompt_tokens += tracker.total_prompt_tokens
self.total_completion_tokens += tracker.total_completion_tokens
self.total_thinking_tokens += tracker.total_thinking_tokens
self.total_cached_tokens += tracker.total_cached_tokens
self.total_calls += tracker.num_calls
self.total_entries += 1
def estimate_cost(self) -> Dict[str, float]:
input_price, output_price, thinking_price = _get_pricing(self.model)
input_cost = self.total_prompt_tokens * input_price / 1_000_000
non_thinking = max(0, self.total_completion_tokens - self.total_thinking_tokens)
output_cost = non_thinking * output_price / 1_000_000
thinking_cost = self.total_thinking_tokens * thinking_price / 1_000_000
return {
"input_cost_usd": round(input_cost, 4),
"output_cost_usd": round(output_cost, 4),
"thinking_cost_usd": round(thinking_cost, 4),
"total_cost_usd": round(input_cost + output_cost + thinking_cost, 4),
}
def summary_str(self) -> str:
cost = self.estimate_cost()
avg_tokens = self.total_prompt_tokens + self.total_completion_tokens
avg_per_entry = avg_tokens / max(1, self.total_entries)
avg_cost = cost["total_cost_usd"] / max(1, self.total_entries)
return (
f"\n{'=' * 72}\n"
f" Token Usage Summary\n"
f"{'=' * 72}\n"
f" Model : {self.model}\n"
f" Total entries : {self.total_entries}\n"
f" Total LLM calls : {self.total_calls}\n"
f" Total prompt tok : {self.total_prompt_tokens:,}\n"
f" Total completion : {self.total_completion_tokens:,}\n"
f" Total thinking : {self.total_thinking_tokens:,}\n"
f" Total cached : {self.total_cached_tokens:,}\n"
f" Total tokens : {self.total_prompt_tokens + self.total_completion_tokens:,}\n"
f" โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ\n"
f" Avg tokens/entry : {avg_per_entry:,.0f}\n"
f" Avg cost/entry : ${avg_cost:.4f}\n"
f" โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ\n"
f" Input cost : ${cost['input_cost_usd']:.4f}\n"
f" Output cost : ${cost['output_cost_usd']:.4f}\n"
f" Thinking cost : ${cost['thinking_cost_usd']:.4f}\n"
f" TOTAL COST : ${cost['total_cost_usd']:.4f}\n"
f"{'=' * 72}"
)
def to_dict(self) -> Dict[str, Any]:
cost = self.estimate_cost()
avg_tokens = self.total_prompt_tokens + self.total_completion_tokens
return {
"model": self.model,
"total_entries": self.total_entries,
"total_llm_calls": self.total_calls,
"total_prompt_tokens": self.total_prompt_tokens,
"total_completion_tokens": self.total_completion_tokens,
"total_thinking_tokens": self.total_thinking_tokens,
"total_cached_tokens": self.total_cached_tokens,
"total_tokens": avg_tokens,
"avg_tokens_per_entry": round(avg_tokens / max(1, self.total_entries)),
"avg_cost_per_entry_usd": round(cost["total_cost_usd"] / max(1, self.total_entries), 6),
"estimated_cost": cost,
}
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
# Think Validation Utilities
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
def think_is_nonempty(text: str) -> bool:
"""Check if a think block contains basic meaningful content.
Relaxed version: Removed all arbitrary length/word count limits.
Only ensures the model didn't return a completely blank string."""
if not text:
return False
# Strip tags if present
clean = re.sub(r'</?think>', '', text).strip()
# Must have just a tiny bit of alphanumeric content to prove it's not just spaces/punctuation
alnum = re.sub(r'[^A-Za-z0-9]', '', clean)
if len(alnum) < 5:
return False
return True
def extract_think_text(text: str) -> str:
"""Extract text content from a <think> block."""
m = re.search(r'<think>(.*?)</think>', text, re.DOTALL)
return (m.group(1) or '').strip() if m else ''
def _dedup_think_content(text: str) -> str:
"""Remove paragraph-level repetition inside a <think> block.
LLMs sometimes repeat entire paragraphs verbatim within a single turn
(decoding-level repetition / "repetition hallucination").
This function detects and removes such duplicates while preserving
unique content and ordering.
Strategy: split by double-newline into paragraphs, keep only the first
occurrence of each paragraph (compared after whitespace normalisation).
Also detects the case where the entire content is duplicated as one
contiguous block (no blank-line separator between the copies).
"""
if not text or not text.strip():
return text
# --- Case 1: paragraph-level dedup (split on blank lines) ---
paragraphs = re.split(r'\n\s*\n', text.strip())
if len(paragraphs) >= 2:
seen = set()
unique = []
for p in paragraphs:
key = ' '.join(p.split()) # normalise whitespace for comparison
if key and key not in seen:
seen.add(key)
unique.append(p)
if len(unique) < len(paragraphs):
return '\n\n'.join(unique)
# --- Case 2: whole-block duplication without blank-line separator ---
# e.g. "ABC\nABC" where ABC is a multi-sentence chunk
stripped = text.strip()
length = len(stripped)
if length >= 80: # only bother for non-trivial blocks
# try splitting at every \n boundary near the midpoint
mid = length // 2
for offset in range(0, min(40, mid)):
for pos in (mid + offset, mid - offset):
if pos <= 0 or pos >= length:
continue
if stripped[pos] != '\n':
continue
first_half = stripped[:pos].strip()
second_half = stripped[pos:].strip()
if first_half == second_half:
return first_half
return text
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
# Hallucination Detection & Sanitization
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
def sanitize_llm_output(text: str) -> str:
"""Truncate LLM output after the FIRST valid action block.
Detects and removes hallucinated content where the model generates:
- Multiple tool_calls in one turn
- Fake <tool_response> blocks
- "MODERATION:" blocks
- Fake search results
- Both <tool_call> AND <answer> in the same turn
Returns cleaned text containing at most: <think>...</think> + one action.
"""
if not text or not text.strip():
return text
text = text.strip()
# Remove any <tool_response> blocks the model hallucinated
# (tool_response should ONLY come from the system)
if '<tool_response>' in text:
# Truncate at the first <tool_response>
tr_start = text.index('<tool_response>')
text = text[:tr_start].strip()
# Remove any "---" separator + MODERATION blocks
moderation_pattern = re.compile(r'\n*---+\s*\n*MODERATION:.*', re.DOTALL | re.IGNORECASE)
text = moderation_pattern.sub('', text).strip()
# Find ALL action blocks (tool_call and answer) with their positions
tc_matches = list(re.finditer(r'<tool_call>.*?</tool_call>', text, re.DOTALL))
ans_matches = list(re.finditer(r'<answer>.*?</answer>', text, re.DOTALL))
all_actions = []
for m in tc_matches:
all_actions.append(('tool_call', m.start(), m.end()))
for m in ans_matches:
all_actions.append(('answer', m.start(), m.end()))
if not all_actions:
# No action found โ return as-is (will be handled by normalize)
return text
# Sort by position โ keep only the FIRST action
all_actions.sort(key=lambda x: x[1])
first_type, first_start, first_end = all_actions[0]
# Truncate: keep everything up to and including the first action
text = text[:first_end].strip()
return text
def is_hallucinated_output(text: str) -> bool:
"""Check if the LLM output contains hallucination markers.
Returns True if the output contains:
- Multiple <tool_call> blocks
- Any <tool_response> block (should only come from system)
- "MODERATION:" blocks
- Both <tool_call> and <answer> in same turn
"""
if not text:
return False
tc_count = len(re.findall(r'<tool_call>', text))
has_tool_response = '<tool_response>' in text
has_moderation = bool(re.search(r'MODERATION:', text, re.IGNORECASE))
has_both = '<tool_call>' in text and '<answer>' in text
return tc_count > 1 or has_tool_response or has_moderation or has_both
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
# GPT Output Normalizer
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
def normalize_gpt_output(text: str) -> str:
"""Ensure every gpt turn follows strict format:
<think>...</think> followed by <tool_call>...</tool_call> or <answer>...</answer>
Pipeline:
1. Sanitize hallucinated content (truncate after first action)
2. If <think> already present, validate and fix duplicates
3. If no <think>, wrap pre-action text as think block if valid
4. ZERO FILLER: If think is missing or empty, DO NOT inject fake fallback text.
"""
if not text or not text.strip():
return text
text = text.strip()
# First pass: deduplicate </think> tags โ keep only the first occurrence
if text.count('</think>') > 1:
first_close_pos = text.index('</think>')
before_and_first = text[:first_close_pos + len('</think>')]
after_first = text[first_close_pos + len('</think>'):]
after_first = after_first.replace('</think>', '')
text = before_and_first + after_first
# Also deduplicate <think> tags โ keep only the first occurrence
if text.count('<think>') > 1:
first_open_pos = text.index('<think>')
before_and_first = text[:first_open_pos + len('<think>')]
after_first = text[first_open_pos + len('<think>'):]
after_first = after_first.replace('<think>', '')
text = before_and_first + after_first
has_think_open = '<think>' in text
has_think_close = '</think>' in text
has_tc = '<tool_call>' in text
has_answer = '<answer>' in text
if has_think_open:
# already has <think>...</think> โ validate & return
if not has_think_close:
# fix unclosed think: insert </think> before first action
if has_tc:
tc_pos = text.index('<tool_call>')
text = text[:tc_pos] + '</think>\n\n' + text[tc_pos:]
elif has_answer:
ans_pos = text.index('<answer>')
text = text[:ans_pos] + '</think>\n\n' + text[ans_pos:]
else:
text = text + '</think>'
# Dedup paragraph-level repetition inside <think> block
think_text = extract_think_text(text)
if think_text:
deduped = _dedup_think_content(think_text)
if deduped != think_text:
text = text.replace(think_text, deduped, 1)
# ไธๅๅๅผบ่ก่กฅๅ
จ๏ผๅชๆ <think> ้้ขๆฏ็ฉบ็๏ผไนๅๆ ท่ฟๅ๏ผ็ปๅฏนไธๅ ๅๅพๆฐๆฎ
return text
# No <think> tag โ check if there is valid text before the action
if has_tc:
tc_start = text.index('<tool_call>')
pre_text = text[:tc_start].strip()
tc_and_after = text[tc_start:]
if pre_text and think_is_nonempty(pre_text):
return f"<think>{pre_text}</think>\n\n{tc_and_after}"
else:
# ๆฒกๆๆๆๅ็ฝฎๆๆฌ๏ผ็ดๆฅ่ฟๅๅจไฝ๏ผไธ่กฅๅ
ๅบ่ฏ
return tc_and_after
elif has_answer:
ans_start = text.index('<answer>')
pre_text = text[:ans_start].strip()
ans_and_after = text[ans_start:]
if pre_text and think_is_nonempty(pre_text):
return f"<think>{pre_text}</think>\n\n{ans_and_after}"
else:
# ๆฒกๆๆๆๅ็ฝฎๆๆฌ๏ผ็ดๆฅ่ฟๅๅจไฝ๏ผไธ่กฅๅ
ๅบ่ฏ
return ans_and_after
else:
# No tool_call and no answer โ wrap whatever it is as think, let downstream filters drop it if needed
return f"<think>{text}</think>"
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
# Training Data Conversation Cleaning
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
def clean_conversations(conversations: List[Dict[str, str]]) -> List[Dict[str, str]]:
"""Remove error recovery turns from training conversations.
Produces a clean trajectory suitable for SFT training by:
1. Removing empty gpt turns (model returned nothing)
2. Removing corresponding error tool_response turns
3. Removing gpt turns that were error messages
4. Re-normalizing remaining gpt turns
Returns a new list of clean conversation turns.
"""
if not conversations:
return conversations
cleaned = []
skip_next_human = False
for i, turn in enumerate(conversations):
role = turn.get("from", "")
value = turn.get("value", "")
if skip_next_human:
if role == "human":
skip_next_human = False
# Check if this is an error response we should skip
if _is_error_tool_response(value):
continue
# Otherwise keep it
cleaned.append(turn)
else:
# Unexpected: another gpt turn โ don't skip
skip_next_human = False
cleaned.append(turn)
continue
if role == "gpt":
# Check if this turn is problematic
stripped = value.strip() if value else ""
# Case 1: Empty gpt turn
if not stripped:
# Skip this turn AND the next human turn (error response)
skip_next_human = True
continue
# Case 2: gpt turn with no valid action (only think, no tool_call/answer)
has_tc = '<tool_call>' in stripped
has_answer = '<answer>' in stripped
if not has_tc and not has_answer:
# No action โ skip this turn and next error response
skip_next_human = True
continue
# Case 3: Valid turn โ normalize and keep
normalized = normalize_gpt_output(stripped)
cleaned.append({"from": "gpt", "value": normalized})
elif role == "human":
# Check if this is an error response to skip
if _is_error_tool_response(value):
# Remove the preceding gpt turn too if it was just added
if cleaned and cleaned[-1].get("from") == "gpt":
# Check if the gpt turn led to this error
# Only remove if the gpt turn had no valid action
pass
continue
cleaned.append(turn)
else:
# system or other โ keep as-is
cleaned.append(turn)
return cleaned
def _is_error_tool_response(value: str) -> bool:
"""Check if a human turn is an error tool_response that should be removed."""
if not value:
return False
v = value.strip()
error_markers = [
"No <tool_call> or <answer> found",
"Error: Tool",
"Malformed tool_call JSON",
"Error: bbox must have exactly",
"Error: web_search requires a non-empty query",
]
for marker in error_markers:
if marker in v:
return True
return False
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
# Frame Utilities
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
def get_video_frame_count(video_path: str) -> int:
"""Get total frame count from a 1fps video (duration ~ frames)."""
try:
result = subprocess.run(
["ffprobe", "-v", "error", "-show_entries", "format=duration",
"-of", "csv=p=0", video_path],
capture_output=True, text=True, timeout=30)
return int(math.ceil(float(result.stdout.strip())))
except Exception:
return -1
def extract_all_frames(video_path: str, output_dir: str,
max_resolution: int = 768,
jpeg_quality: int = 85) -> Dict[int, str]:
"""Extract all frames from 1fps video. Returns {frame_index(0-based): path}.
Reuses existing frames if the directory already has them."""
os.makedirs(output_dir, exist_ok=True)
# Check if frames already exist
existing = {}
for f in sorted(os.listdir(output_dir)):
if f.startswith("frame_") and f.endswith(".jpg"):
idx = int(f.replace("frame_", "").replace(".jpg", ""))
existing[idx] = os.path.join(output_dir, f)
if existing:
return existing
vf = (f"scale='min({max_resolution},iw)':'min({max_resolution},ih)'"
f":force_original_aspect_ratio=decrease")
q = max(2, min(31, round(2 + (100 - jpeg_quality) * 29 / 99)))
cmd = [
"ffmpeg", "-y", "-i", video_path,
"-vf", vf, "-q:v", str(q),
"-start_number", "0",
os.path.join(output_dir, "frame_%06d.jpg"),
]
proc = subprocess.run(cmd, capture_output=True, text=True, timeout=600)
if proc.returncode != 0:
raise RuntimeError(f"ffmpeg failed: {proc.stderr[-500:]}")
frames = {}
for f in sorted(os.listdir(output_dir)):
if f.startswith("frame_") and f.endswith(".jpg"):
idx = int(f.replace("frame_", "").replace(".jpg", ""))
frames[idx] = os.path.join(output_dir, f)
if not frames:
raise RuntimeError(f"No frames extracted from {video_path}")
return frames
def uniform_sample_indices(total_frames: int, num_samples: int) -> List[int]:
"""Uniformly sample 0-based frame indices. total_frames is the count."""
if total_frames <= num_samples:
return list(range(total_frames))
return sorted(set(int(i) for i in np.linspace(0, total_frames - 1, num_samples)))
def sample_interval(all_frames: Dict[int, str], start: int, end: int,
num_samples: int = 8) -> List[Tuple[int, str]]:
"""Uniformly sample frames from [start, end] interval."""
available = sorted(k for k in all_frames if start <= k <= end)
if not available:
return []
if len(available) <= num_samples:
return [(k, all_frames[k]) for k in available]
positions = np.linspace(0, len(available) - 1, num_samples, dtype=int)
selected = sorted(set(available[p] for p in positions))
return [(k, all_frames[k]) for k in selected]
def get_frame(all_frames: Dict[int, str], idx: int) -> Tuple[int, str]:
"""Get exact frame or nearest available."""
if idx in all_frames:
return (idx, all_frames[idx])
nearest = min(all_frames.keys(), key=lambda k: abs(k - idx))
return (nearest, all_frames[nearest])
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
# Bbox Format Detection & Normalization
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
def get_bbox_config(model: str) -> dict:
"""ๆ นๆฎๆจกๅๅ็งฐ่ฟๅๅฏนๅบ็ bbox ๆ ผๅผ้
็ฝฎใ
ๅน้
่งๅ: model ๅ็งฐไธญๅ
ๅซ pattern ็ๆ้ฟๅน้
ไผๅ
ใ
ๅฆๆๆฒกๆๅน้
, ่ฟๅ default ้
็ฝฎใ
Returns:
{"order": "xyxy"|"yxyx", "range": "norm"|"permille"}
"""
model_lower = model.lower()
best_match = ""
best_config = BBOX_CONFIGS.get("default", {"order": "xyxy", "range": "norm"})
for pattern, config in BBOX_CONFIGS.items():
if pattern == "default":
continue
if pattern.lower() in model_lower and len(pattern) > len(best_match):
best_match = pattern
best_config = config
return best_config
def normalize_bbox(raw_bbox: list, bbox_config: dict) -> list:
"""ๅฐๆจกๅ่พๅบ็ bbox ็ปไธ่ฝฌๆขไธบ [x1, y1, x2, y2] ไธๅผๅ [0.0, 1.0]ใ
ๆฏๆ็่พๅ
ฅๆ ผๅผ:
- Gemini: [y_min, x_min, y_max, x_max] in [0, 1000]
- ๆ ๅ: [x1, y1, x2, y2] in [0.0, 1.0]
- ๅ็ด : [x1, y1, x2, y2] ็ปๅฏนๅ็ด (่ชๅจ้็บง)
Args:
raw_bbox: ้ฟๅบฆไธบ 4 ็ๅ่กจ
bbox_config: get_bbox_config() ่ฟๅ็้
็ฝฎๅญๅ
ธ
Returns:
[x1, y1, x2, y2] ๅ
จ้จๅฝไธๅๅฐ [0.0, 1.0]
"""
if len(raw_bbox) != 4:
return [0.0, 0.0, 1.0, 1.0]
# Step 1: ่งฃๆๅๆ ้กบๅบ
if bbox_config.get("order") == "yxyx":
# Gemini: [y_min, x_min, y_max, x_max]
y1, x1, y2, x2 = raw_bbox
else:
# ๆ ๅ: [x1, y1, x2, y2]
x1, y1, x2, y2 = raw_bbox
# Step 2: ๅฝไธๅๅผๅๅฐ [0.0, 1.0]
if bbox_config.get("range") == "permille":
# [0, 1000] โ [0.0, 1.0]
x1, y1, x2, y2 = x1 / 1000.0, y1 / 1000.0, x2 / 1000.0, y2 / 1000.0
else:
# ่ชๅจๆฃๆต: ๅฆๆๆๅคงๅผ > 1.0 ไฝ <= 1000, ๆ permille ๅค็
max_coord = max(abs(x1), abs(y1), abs(x2), abs(y2))
if max_coord > 1.0 and max_coord <= 1000:
x1, y1, x2, y2 = x1 / 1000.0, y1 / 1000.0, x2 / 1000.0, y2 / 1000.0
elif max_coord > 1000:
# ็ปๅฏนๅ็ด ๅๆ โ ๆ ๆณๅจๆญคๅฝไธๅ, ่ฟๅๅๅงๅผ่ฎฉ crop_frame ๅค็
# (crop_frame ๅ
้จๆ auto-detect ้ป่พ)
pass
# Step 3: Clamp ๅฐ [0.0, 1.0]
x1 = max(0.0, min(1.0, float(x1)))
y1 = max(0.0, min(1.0, float(y1)))
x2 = max(0.0, min(1.0, float(x2)))
y2 = max(0.0, min(1.0, float(y2)))
# Step 4: ็กฎไฟ x1 < x2, y1 < y2
if x1 > x2:
x1, x2 = x2, x1
if y1 > y2:
y1, y2 = y2, y1
# Step 5: ไฟ่ฏๆๅฐ้ข็งฏ
if x2 - x1 < 0.001:
x2 = min(1.0, x1 + 0.01)
if y2 - y1 < 0.001:
y2 = min(1.0, y1 + 0.01)
return [x1, y1, x2, y2]
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
# Image Search Failure Handling & Padding
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
class RetrieverDownError(Exception):
"""Retriever ๆๅก(IP/็ซฏๅฃ)่ฟๆฅๅคฑ่ดฅ๏ผๅบ็ซๅณๅๆญขๆดไธช pipelineใ"""
pass
class ImageSearchFailedError(Exception):
"""Raised when image_search returns an error, timeout, or no results.
Signals the entire entry should be retried from scratch so that
failed serper results never appear in the SFT dataset."""
pass
class QuotaExhaustedError(Exception):
"""API ้ขๅบฆ็ๆญฃ็จๅฐฝ (้ไธดๆถ้ๆต)๏ผๅบ็ซๅณๅๆญขๆดไธช pipelineใ"""
pass
class ProhibitedContentError(Exception):
"""ๅ
ๅฎน่ขซ Vertex AI ๅฎๅ
จ็ญ็ฅๆฆๆช (PROHIBITED_CONTENT)๏ผไธๅบ้่ฏใ"""
pass
class ProjectDisabledError(Exception):
"""ๆไธช Vertex project ไธๅฏ็จ๏ผๅบไปๆฌ่ฝฎ project ๆฑ ไธญ็ฆ็จใ"""
pass
def _is_quota_exhausted(status_code: int, error_body: str) -> bool:
"""ๅคๆญ API ้่ฏฏๆฏๅฆไธบ้ขๅบฆ็ๆญฃ็จๅฐฝ (ๅบๅซไบไธดๆถ rate limit)ใ
Vertex AI ็ 429 / RESOURCE_EXHAUSTED ้ๅธธๆฏ็ฌๆถ้ๆต๏ผQPM/TPM ๆๆปก๏ผ๏ผ
ๅบ่ฏฅ้่ฏ่้ๅๆญขใๅชๆ 403 + ๆ็กฎ็้
้ข/่ดฆๅๅ
ณ้ฎ่ฏๆ่ฎคไธบๆฏ็ๆญฃ็้ขๅบฆ็จๅฐฝใ
ๅ่: https://docs.cloud.google.com/docs/quotas/troubleshoot
- 429 โ ไธดๆถ้ๆต๏ผ้้ฟ้่ฏ
- 403 + QUOTA_EXCEEDED / RATE_LIMIT_EXCEEDED / billing โ ็ๆญฃ้ขๅบฆ็จๅฐฝ
"""
body_lower = error_body.lower()
# 429 ไธๅพ่งไธบไธดๆถ้ๆต๏ผไธ็ฎ้ขๅบฆ็จๅฐฝ
if status_code == 429:
return False
# 403 + ๆ็กฎ้
้ข/่ดฆๅไฟกๅท โ ็ๆญฃ้ขๅบฆ็จๅฐฝ
if status_code == 403:
quota_keywords = [
"quota_exceeded",
"quota exceeded",
"rate_limit_exceeded",
"billing account",
"billing is disabled",
"insufficient quota",
"out of quota",
"daily limit",
"per-day limit",
]
if any(kw in body_lower for kw in quota_keywords):
return True
return False
def is_image_search_failed(result_text: str) -> bool:
"""Check if an image search result indicates a failure.
Returns True for any error, timeout, or empty-result response
from the Serper Lens API.
"""
if not result_text or not result_text.strip():
return True
fail_patterns = [
"Image search error:",
"No results found from reverse image search",
"Error: SERPER_API_KEY not configured",
"request timed out",
]
for pattern in fail_patterns:
if pattern in result_text:
return True
return False
def add_search_padding(bbox: List[float], frame_path: str,
padding: tuple = (0.5, 0.5),
padding_cap_px: int = 600) -> List[float]:
"""Add padding to a normalized [x1, y1, x2, y2] bbox for image search.
Padding is proportional to the BBOX size (not image size), so that:
- A tight face crop gets moderate expansion (include shoulders, some background)
- A large crop doesn't balloon to cover the entire frame
Args:
bbox: [x1, y1, x2, y2] normalized to [0.0, 1.0]
frame_path: path to the source frame (used to compute pixel cap)
padding: (pad_x_ratio, pad_y_ratio) as fraction of bbox width/height
e.g. (0.5, 0.5) means expand each side by 50% of bbox dimension
padding_cap_px: max padding in pixels on each side (prevents excessive
expansion on very large bboxes)
Returns:
Padded [x1, y1, x2, y2] clamped to [0.0, 1.0]
"""
x1, y1, x2, y2 = bbox
pad_x_ratio, pad_y_ratio = padding
# Padding proportional to bbox dimensions
bbox_w = x2 - x1
bbox_h = y2 - y1
pad_x = bbox_w * pad_x_ratio
pad_y = bbox_h * pad_y_ratio
# Cap padding at padding_cap_px pixels (convert to normalized coords)
try:
with Image.open(frame_path) as img:
img_w, img_h = img.size
cap_x = padding_cap_px / img_w
cap_y = padding_cap_px / img_h
pad_x = min(pad_x, cap_x)
pad_y = min(pad_y, cap_y)
except Exception:
pass # If we can't read dimensions, just use the bbox-proportional padding
x1 = max(0.0, x1 - pad_x)
y1 = max(0.0, y1 - pad_y)
x2 = min(1.0, x2 + pad_x)
y2 = min(1.0, y2 + pad_y)
return [x1, y1, x2, y2]
def crop_frame(frame_path: str, bbox: List[float], output_path: str) -> str:
"""Crop frame at bbox coordinates with smart format detection, clamping,
and 2x upscaling for better search quality.
Bbox format auto-detection:
- [0.0, 1.0] range โ normalized relative coordinates (standard)
- [0, 1000] range โ permille relative coordinates (some models output this)
- values > 1000 โ absolute pixel coordinates
After cropping:
- 2x LANCZOS upscale to help visual search models recognize small objects
- Save as high-quality JPEG (quality=95) to minimize compression artifacts
Args:
frame_path: path to the source frame image
bbox: [x1, y1, x2, y2] in any of the three supported formats
output_path: where to save the cropped image
Returns:
output_path
"""
with Image.open(frame_path) as img:
w, h = img.size
raw_x1, raw_y1, raw_x2, raw_y2 = bbox
# โโ Step 1: Auto-detect coordinate format and convert to pixels โโ
max_coord = max(abs(raw_x1), abs(raw_y1), abs(raw_x2), abs(raw_y2))
if max_coord <= 1.0:
# Format A: normalized [0.0, 1.0]
px_x1 = raw_x1 * w
px_y1 = raw_y1 * h
px_x2 = raw_x2 * w
px_y2 = raw_y2 * h
elif max_coord <= 1000:
# Format B: permille [0, 1000]
px_x1 = raw_x1 / 1000.0 * w
px_y1 = raw_y1 / 1000.0 * h
px_x2 = raw_x2 / 1000.0 * w
px_y2 = raw_y2 / 1000.0 * h
else:
# Format C: absolute pixel coordinates
px_x1 = raw_x1
px_y1 = raw_y1
px_x2 = raw_x2
px_y2 = raw_y2
# โโ Step 2: Clamp to image bounds โโ
x1 = max(0, min(int(round(px_x1)), w - 1))
y1 = max(0, min(int(round(px_y1)), h - 1))
x2 = max(0, min(int(round(px_x2)), w))
y2 = max(0, min(int(round(px_y2)), h))
# Ensure minimum 1px crop (prevent zero-area)
if x2 <= x1:
x2 = min(x1 + 1, w)
if y2 <= y1:
y2 = min(y1 + 1, h)
# โโ Step 3: Crop โโ
cropped_img = img.crop((x1, y1, x2, y2))
# โโ Step 4: 2x upscale with LANCZOS for better search recognition โโ
cropped_img = cropped_img.resize(
(cropped_img.width * 2, cropped_img.height * 2),
Image.Resampling.LANCZOS,
)
# โโ Step 5: Save as high-quality JPEG โโ
cropped_img.save(output_path, 'JPEG', quality=95)
return output_path
def encode_image_b64(path: str) -> str:
with open(path, "rb") as f:
return base64.b64encode(f.read()).decode("ascii")
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
# LLM Client with Multi-API Load Balancing (v3 โ Token Tracking)
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
class LLMClient:
"""Async LLM client with round-robin load balancing across API endpoints.
v3: call() now returns (content, reasoning_content, TokenUsage).
v4: Google Vertex AI authentication support (OAuth2 auto-refresh).
v5: Vertex AI Native generateContent API support (for global region).
"""
def __init__(self, api_urls: List[str], model: str,
concurrency_per_url: int = 2,
temperature: float = DEFAULT_TEMPERATURE,
top_p: float = DEFAULT_TOP_P,
max_tokens: int = DEFAULT_MAX_TOKENS,
timeout: int = DEFAULT_TIMEOUT,
max_retries: int = DEFAULT_MAX_RETRIES,
api_key: str = API_KEY):
self.api_urls = api_urls
self.model = model
self.temperature = temperature
self.top_p = top_p
self.max_tokens = max_tokens
self.timeout = timeout
self.max_retries = max_retries
self.api_key = api_key
self._url_cycle = itertools.cycle(api_urls)
self._semaphores = {url: asyncio.Semaphore(concurrency_per_url)
for url in api_urls}
self._cycle_lock = asyncio.Lock()
# โโ Project ๆฑ ๏ผ429 ๆถ่ชๅจๅๆข โโ
self._active_url_index = 0 # ๅฝๅๆดป่ท็ URL ็ดขๅผ
self._url_switch_lock = asyncio.Lock() # ๅๆข้
self._project_cooldown_until = {url: 0.0 for url in api_urls}
self._disabled_urls: Set[str] = set()
# โโ Vertex ่ฏทๆฑๆดๅฝข๏ผ้ไฝ็ง็บง burst ๅฏผ่ด็ 429 โโ
self._vertex_min_interval = max(0.0, float(VERTEX_MIN_REQUEST_INTERVAL_SECONDS or 0.0))
self._vertex_rate_limit_cooldown = max(0.0, float(VERTEX_RATE_LIMIT_COOLDOWN_SECONDS or 0.0))
self._vertex_request_jitter = max(0.0, float(VERTEX_REQUEST_JITTER_SECONDS or 0.0))
self._vertex_next_request_at = 0.0
self._vertex_pacing_lock = asyncio.Lock()
# โโ ๆฃๆต API ๆจกๅผ โโ
self._use_vertex_auth = self._detect_vertex_endpoint()
self._use_vertex_native = self._detect_vertex_native()
self._gcp_credentials = None
self._gcp_credentials_by_url: Dict[str, Any] = {}
self._gcp_auth_request = None
if self._use_vertex_auth:
self._init_gcp_auth()
def _detect_vertex_endpoint(self) -> bool:
"""ๆฃๆตๆฏๅฆไฝฟ็จ Vertex AI endpoint๏ผๆ นๆฎ URL ๅคๆญ๏ผใ"""
for url in self.api_urls:
if "aiplatform.googleapis.com" in url:
return True
return False
def _detect_vertex_native(self) -> bool:
"""ๆฃๆตๆฏๅฆไฝฟ็จ Vertex AI ๅ็ generateContent APIใ
ๅคๆญไพๆฎ๏ผURL ไธญๅ
ๅซ /publishers/google/models๏ผๅ็๏ผ
่้ /endpoints/openapi/chat/completions๏ผOpenAI ๅ
ผๅฎน๏ผ
"""
for url in self.api_urls:
if "/publishers/google/models" in url:
return True
# ไนๆฃๆฅ config ไธญ็ๆ ๅฟ
try:
use_native = getattr(config, 'USE_VERTEX_NATIVE_API', False)
if use_native:
return True
except Exception:
pass
return False
def _init_gcp_auth(self):
"""ๅๅงๅ Google Cloud ่ฎค่ฏๅญๆฎใ"""
try:
import google.auth
import google.auth.transport.requests
from google.oauth2 import service_account as sa_module
scopes = ["https://www.googleapis.com/auth/cloud-platform"]
self._gcp_auth_request = google.auth.transport.requests.Request()
pool_records = []
for item in VERTEX_CREDENTIALS_POOL:
if not isinstance(item, dict):
continue
api_url = item.get("api_url")
if api_url in self.api_urls:
pool_records.append(item)
if pool_records:
for item in pool_records:
credential = None
key_path = item.get("service_account_key") or ""
key_info = item.get("service_account_info")
if key_path and os.path.exists(key_path):
credential = sa_module.Credentials.from_service_account_file(
key_path, scopes=scopes,
)
print(
f"[GCP AUTH] ไฝฟ็จ่ดฆๅทๆฑ ๅฏ้ฅ: account={item.get('account_name', '')} "
f"project={item.get('project_id', '')} key={key_path}"
)
elif isinstance(key_info, dict):
credential = sa_module.Credentials.from_service_account_info(
key_info, scopes=scopes,
)
print(
f"[GCP AUTH] ไฝฟ็จ่ดฆๅทๆฑ ๅ
ๅตๅฏ้ฅ: account={item.get('account_name', '')} "
f"project={item.get('project_id', '')}"
)
else:
raise RuntimeError(
f"Vertex ่ดฆๅทๆฑ ๆก็ฎ็ผบๅฐๅฏ็จๅญ่ฏ: project={item.get('project_id', '')}"
)
credential.refresh(self._gcp_auth_request)
self._gcp_credentials_by_url[item["api_url"]] = credential
print(f"[GCP AUTH] ่ดฆๅทๆฑ ่ฎค่ฏๆๅ๏ผๅฏ็จๆก็ฎๆฐ: {len(self._gcp_credentials_by_url)}")
return
if GCP_SERVICE_ACCOUNT_KEY and os.path.exists(GCP_SERVICE_ACCOUNT_KEY):
# ไฝฟ็จๆๅก่ดฆๅทๅฏ้ฅๆไปถ
self._gcp_credentials = sa_module.Credentials.from_service_account_file(
GCP_SERVICE_ACCOUNT_KEY, scopes=scopes,
)
print(f"[GCP AUTH] ไฝฟ็จๆๅก่ดฆๅทๅฏ้ฅ: {GCP_SERVICE_ACCOUNT_KEY}")
else:
# ไฝฟ็จ Application Default Credentials (ADC)
# ้่ฆๅ
่ฟ่ก `gcloud auth application-default login`
self._gcp_credentials, project = google.auth.default(scopes=scopes)
print(f"[GCP AUTH] ไฝฟ็จ ADC (Application Default Credentials), project={project}")
# ้ขๅ
ๅทๆฐไธๆฌก๏ผ็กฎ่ฎคๅญๆฎๆๆ
self._gcp_credentials.refresh(self._gcp_auth_request)
print(f"[GCP AUTH] ่ฎค่ฏๆๅ๏ผtoken ๆๆๆ่ณ {self._gcp_credentials.expiry}")
except ImportError:
raise ImportError(
"ไฝฟ็จ Vertex AI ้่ฆๅฎ่ฃ
google-auth ๅบ๏ผ\n"
" pip install google-auth google-auth-httplib2"
)
except Exception as e:
raise RuntimeError(
f"Google Cloud ่ฎค่ฏๅคฑ่ดฅ: {e}\n"
f"่ฏท็กฎไฟๅทฒ้
็ฝฎ GCP_SERVICE_ACCOUNT_KEY ๆ่ฟ่ก "
f"`gcloud auth application-default login`"
)
def _get_gcp_token(self, api_url: Optional[str] = None) -> str:
"""่ทๅๆๆ็ GCP OAuth2 access token๏ผ่ชๅจๅทๆฐ่ฟๆ token๏ผใ"""
if api_url and api_url in self._gcp_credentials_by_url:
credentials = self._gcp_credentials_by_url[api_url]
if credentials.expired or not credentials.token:
credentials.refresh(self._gcp_auth_request)
return credentials.token
if self._gcp_credentials.expired or not self._gcp_credentials.token:
self._gcp_credentials.refresh(self._gcp_auth_request)
return self._gcp_credentials.token
async def _next_url(self) -> str:
async with self._cycle_lock:
return next(self._url_cycle)
def _get_active_url(self) -> str:
"""่ทๅๅฝๅๆดป่ท็ project URL๏ผ429 ๅๆขๅไผๅๅ๏ผใ"""
return self.api_urls[self._active_url_index]
async def _acquire_vertex_request_slot(self):
"""ๅฏน Vertex ่ฏทๆฑๅๅ
จๅฑ่ๆต๏ผ้ฟๅ
ๅ
ฑไบซๆฑ ็ง็บง burstใ"""
if not self._use_vertex_auth or self._vertex_min_interval <= 0:
return
while True:
async with self._vertex_pacing_lock:
now = time.monotonic()
wait = self._vertex_next_request_at - now
if wait <= 0:
reserve = self._vertex_min_interval + random.uniform(0, self._vertex_request_jitter)
self._vertex_next_request_at = now + reserve
return
if wait > 0.5:
print(f" [VERTEX PACE] waiting {wait:.1f}s before next request")
await asyncio.sleep(max(wait, 0.05))
async def _mark_project_cooldown(self, failed_url: str):
"""ๆไธช project ่ฟๅ 429 ๅ๏ผ็ญๆถ้ดๅ
ไธ่ฆ็ซๅณๆๅๅปใ"""
if self._vertex_rate_limit_cooldown <= 0:
return
cooldown = self._vertex_rate_limit_cooldown + random.uniform(0, self._vertex_request_jitter)
async with self._url_switch_lock:
until_ts = time.monotonic() + cooldown
prev = self._project_cooldown_until.get(failed_url, 0.0)
self._project_cooldown_until[failed_url] = max(prev, until_ts)
print(f" [PROJECT POOL] cooldown {cooldown:.1f}s for {failed_url[:80]}...")
async def _disable_project_url(self, failed_url: str, reason: str):
"""็ฆ็จ suspended / API-disabled / permission-denied ็ projectใ"""
async with self._url_switch_lock:
self._disabled_urls.add(failed_url)
print(f" [PROJECT POOL] disabled project: {failed_url[:80]}... reason={reason[:160]}")
if len(self._disabled_urls) >= len(self.api_urls):
raise ProjectDisabledError("All Vertex projects are disabled or suspended.")
if self.api_urls[self._active_url_index] == failed_url:
for idx, url in enumerate(self.api_urls):
if url not in self._disabled_urls:
self._active_url_index = idx
break
async def _acquire_project_url(self) -> Tuple[str, int]:
"""้ๆฉไธไธชๅฝๅๆชๅคไบๅทๅดๆ็ project๏ผๅฆๆ้ฝๅจๅทๅด๏ผๅ็ญๅพ
ๆๆฉๆขๅค็้ฃไธชใ"""
while True:
async with self._url_switch_lock:
now = time.monotonic()
count = len(self.api_urls)
for offset in range(count):
idx = (self._active_url_index + offset) % count
url = self.api_urls[idx]
if url in self._disabled_urls:
continue
if self._project_cooldown_until.get(url, 0.0) <= now:
self._active_url_index = idx
return url, idx
available_urls = [u for u in self.api_urls if u not in self._disabled_urls]
if not available_urls:
raise ProjectDisabledError("All Vertex projects are disabled or suspended.")
soonest_url = min(available_urls, key=lambda u: self._project_cooldown_until.get(u, 0.0))
wait = max(0.0, self._project_cooldown_until.get(soonest_url, 0.0) - now)
print(f" [PROJECT POOL] all projects cooling down, waiting {wait:.1f}s")
await asyncio.sleep(max(wait, 0.1))
async def _switch_to_next_project(self, failed_index: int) -> bool:
"""429 ๆถๅๆขๅฐไธไธไธช projectใ่ฟๅ True ๅฆๆๆๅๅๆขๅฐไธๅ็ projectใ"""
async with self._url_switch_lock:
# ๅฏ่ฝๅทฒ็ป่ขซๅ
ถไปๅ็จๅๆข่ฟไบ
if self._active_url_index != failed_index:
return True # ๅทฒ็ปๅๆขไบ๏ผ็ดๆฅ้่ฏ
now = time.monotonic()
count = len(self.api_urls)
for offset in range(1, count + 1):
next_index = (failed_index + offset) % count
next_url = self.api_urls[next_index]
if next_url in self._disabled_urls:
continue
if self._project_cooldown_until.get(next_url, 0.0) <= now:
self._active_url_index = next_index
print(f" [PROJECT POOL] 429 โ ๅๆขๅฐ project #{next_index + 1}/{len(self.api_urls)}: {next_url[:80]}...")
return True
return False
def _all_projects_cooling_down(self) -> bool:
now = time.monotonic()
active_urls = [url for url in self.api_urls if url not in self._disabled_urls]
return bool(active_urls) and all(self._project_cooldown_until.get(url, 0.0) > now for url in active_urls)
@staticmethod
def _is_project_disabled_error(err_str: str) -> bool:
err_lower = err_str.lower()
return any(marker in err_lower for marker in (
"consumer_suspended",
"has been suspended",
"service_disabled",
"api has not been used",
"api is disabled",
))
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
# OpenAI โ Vertex Native ๆถๆฏๆ ผๅผ่ฝฌๆข
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
@staticmethod
def _convert_messages_to_vertex_native(
messages: List[Dict],
) -> Tuple[Optional[Dict], List[Dict]]:
"""ๅฐ OpenAI ๆ ผๅผ็ messages ่ฝฌๆขไธบ Vertex AI ๅ็ๆ ผๅผใ
OpenAI ๆ ผๅผ:
[{"role": "system", "content": "..."},
{"role": "user", "content": "..." | [{"type":"text",...}, {"type":"image_url",...}]},
{"role": "assistant", "content": "..."}]
Vertex ๅ็ๆ ผๅผ:
systemInstruction: {"parts": [{"text": "..."}]}
contents: [
{"role": "user", "parts": [{"text": "..."}, {"inlineData": {...}}]},
{"role": "model", "parts": [{"text": "..."}]}
]
Returns: (system_instruction_dict_or_None, contents_list)
"""
system_instruction = None
contents = []
for msg in messages:
role = msg.get("role", "")
content = msg.get("content", "")
if role == "system":
# ๆๅ system prompt โ systemInstruction
if isinstance(content, str):
system_instruction = {"parts": [{"text": content}]}
elif isinstance(content, list):
parts = []
for item in content:
if isinstance(item, str):
parts.append({"text": item})
elif isinstance(item, dict) and item.get("type") == "text":
parts.append({"text": item.get("text", "")})
system_instruction = {"parts": parts}
continue
# ่ง่ฒๆ ๅฐ: assistant โ model
vertex_role = "model" if role == "assistant" else "user"
# ่ฝฌๆข content โ parts
parts = []
if isinstance(content, str):
if content.strip():
parts.append({"text": content})
elif isinstance(content, list):
for item in content:
if isinstance(item, str):
if item.strip():
parts.append({"text": item})
elif isinstance(item, dict):
item_type = item.get("type", "")
if item_type == "text":
text_val = item.get("text", "")
if text_val.strip():
parts.append({"text": text_val})
elif item_type == "image_url":
# OpenAI: {"type":"image_url","image_url":{"url":"data:image/jpeg;base64,..."}}
image_url = item.get("image_url", {})
url = image_url.get("url", "") if isinstance(image_url, dict) else ""
if url.startswith("data:"):
# ่งฃๆ data URI: data:image/jpeg;base64,xxxx
# ๆๅ mimeType ๅ base64 ๆฐๆฎ
try:
header, b64_data = url.split(",", 1)
# header = "data:image/jpeg;base64"
mime_type = header.split(":")[1].split(";")[0]
parts.append({
"inlineData": {
"mimeType": mime_type,
"data": b64_data,
}
})
except (ValueError, IndexError):
# ่งฃๆๅคฑ่ดฅ๏ผ่ทณ่ฟ
pass
elif url.startswith("gs://"):
# GCS URI
parts.append({
"fileData": {
"fileUri": url,
"mimeType": "image/jpeg",
}
})
if parts:
contents.append({"role": vertex_role, "parts": parts})
return system_instruction, contents
@staticmethod
def _parse_vertex_native_response(
result: Dict[str, Any],
) -> Tuple[str, str, TokenUsage]:
"""่งฃๆ Vertex AI ๅ็ generateContent ๅๅบใ
Vertex ๅๅบๆ ผๅผ:
{
"candidates": [{
"content": {
"role": "model",
"parts": [{"text": "..."}, ...]
},
"finishReason": "STOP"
}],
"usageMetadata": {
"promptTokenCount": 100,
"candidatesTokenCount": 50,
"totalTokenCount": 150,
"thoughtsTokenCount": 20 // ๅฏ้๏ผthinking tokens
}
}
Returns: (content_text, reasoning_text, TokenUsage)
"""
# ๆฃๆตๅ
ๅฎนๅฎๅ
จๆฆๆช โ ไธๅบ้่ฏ
prompt_feedback = result.get("promptFeedback", {})
block_reason = prompt_feedback.get("blockReason", "")
if block_reason == "PROHIBITED_CONTENT":
raise ProhibitedContentError(
f"Content blocked by Vertex safety filter: {block_reason}")
candidates = result.get("candidates", [])
if not candidates:
raise ValueError(f"Empty candidates in Vertex response: {json.dumps(result)[:300]}")
candidate = candidates[0]
parts = candidate.get("content", {}).get("parts", [])
content_text = ""
reasoning_text = ""
for part in parts:
if "text" in part:
# ๆฃๆฅๆฏๅฆๆฏ thought/reasoning part
# Vertex ๅ็ API ไธญ thinking ๅ
ๅฎนๅฏ่ฝๅจ thought ๅญๆฎต
if part.get("thought", False):
reasoning_text += part["text"]
else:
content_text += part["text"]
# ่งฃๆ token ไฝฟ็จ้
usage_meta = result.get("usageMetadata", {})
prompt_tokens = usage_meta.get("promptTokenCount", 0) or 0
completion_tokens = usage_meta.get("candidatesTokenCount", 0) or 0
total_tokens = usage_meta.get("totalTokenCount", 0) or (prompt_tokens + completion_tokens)
thinking_tokens = usage_meta.get("thoughtsTokenCount", 0) or 0
cached_tokens = usage_meta.get("cachedContentTokenCount", 0) or 0
usage = TokenUsage(
prompt_tokens=prompt_tokens,
completion_tokens=completion_tokens,
total_tokens=total_tokens,
thinking_tokens=thinking_tokens,
cached_tokens=cached_tokens,
)
return content_text, reasoning_text, usage
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
# ไธป่ฐ็จๆนๆณ
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
async def call(
self,
messages: List[Dict],
session: aiohttp.ClientSession,
temperature: float = None,
max_tokens: int = None,
) -> Tuple[str, str, TokenUsage]:
"""Call LLM with load balancing and project pool failover.
่ชๅจๆ นๆฎ API ๆจกๅผ้ๆฉ OpenAI ๅ
ผๅฎน ๆ Vertex Native ๆ ผๅผใ
้ๅฐ 429 ้ๆตๆถ่ชๅจๅๆขๅฐไธไธไธช project URL ้่ฏใ
Returns (content, reasoning_content, TokenUsage).
"""
last_err = None
start_index = self._active_url_index
# ๆๅคๅฐ่ฏๆๆ project
for _pool_attempt in range(len(self.api_urls)):
base_url, url_index = await self._acquire_project_url()
sem = self._semaphores[base_url]
try:
if self._use_vertex_native:
return await self._call_vertex_native(
messages, session, base_url, sem, temperature, max_tokens
)
else:
return await self._call_openai_compat(
messages, session, base_url, sem, temperature, max_tokens
)
except Exception as e:
err_str = str(e)
is_rate_limit = (
"429" in err_str
or "RESOURCE_EXHAUSTED" in err_str
or "rate limit" in err_str.lower()
)
if self._is_project_disabled_error(err_str):
last_err = e
await self._disable_project_url(base_url, err_str)
continue
if is_rate_limit and len(self.api_urls) > 1:
last_err = e
await self._mark_project_cooldown(base_url)
await self._switch_to_next_project(url_index)
# ๅฆๆๅๆขๅๅๅฐ่ตท็น๏ผ่ฏดๆๆๆ project ้ฝ 429 ไบ
if self._all_projects_cooling_down():
print(f" [PROJECT POOL] ๆๆ {len(self.api_urls)} ไธช project ๆญฃๅจ cooldown")
raise
if self._active_url_index == start_index:
print(f" [PROJECT POOL] ๆๆ {len(self.api_urls)} ไธช project ๅๅทฒ 429๏ผไธๆๅผๅธธ")
raise
continue # ็จๆฐ project ้่ฏ
# ้ 429 ้่ฏฏ โ ็ดๆฅๆๅบ
raise
# ๆๆ project ้ฝ 429
raise last_err or RuntimeError("All projects in pool exhausted (429)")
async def _call_openai_compat(
self,
messages: List[Dict],
session: aiohttp.ClientSession,
api_url: str,
sem: asyncio.Semaphore,
temperature: float = None,
max_tokens: int = None,
) -> Tuple[str, str, TokenUsage]:
"""OpenAI ๅ
ผๅฎน็ซฏ็น่ฐ็จ๏ผๅๆ้ป่พ๏ผไฟๆไธๅ๏ผใ"""
payload = {
"model": self.model,
"messages": messages,
"temperature": temperature or self.temperature,
"top_p": self.top_p,
"max_tokens": max_tokens or self.max_tokens,
}
for attempt in range(self.max_retries):
try:
headers = {"Content-Type": "application/json"}
if self._use_vertex_auth:
token = self._get_gcp_token(api_url)
headers["Authorization"] = f"Bearer {token}"
elif self.api_key:
headers["Authorization"] = f"Bearer {self.api_key}"
async with sem:
await self._acquire_vertex_request_slot()
async with session.post(
api_url, json=payload, headers=headers,
timeout=aiohttp.ClientTimeout(total=self.timeout),
) as resp:
if resp.status != 200:
error_body = await resp.text()
if _is_quota_exhausted(resp.status, error_body):
raise QuotaExhaustedError(
f"API ้ขๅบฆ็จๅฐฝ! HTTP {resp.status}: {error_body[:500]}")
raise RuntimeError(
f"HTTP {resp.status} from {api_url}: {error_body[:500]}")
result = await resp.json()
choices = result.get("choices", [])
if not choices:
raise ValueError(f"Empty choices: {json.dumps(result)[:300]}")
message = choices[0].get("message", {})
content = message.get("content", "") or ""
reasoning = message.get("reasoning_content", "") or ""
# Parse token usage
usage_data = result.get("usage", {})
usage = TokenUsage.from_api_response(usage_data)
return content, reasoning, usage
except QuotaExhaustedError:
raise # ้ขๅบฆ็จๅฐฝ๏ผไธ้่ฏ๏ผ็ดๆฅไธๆ
except ProhibitedContentError:
raise # ๅ
ๅฎน่ขซๅฎๅ
จ็ญ็ฅๆฆๆช๏ผไธ้่ฏ๏ผ็ดๆฅไธๆ
except ProjectDisabledError:
raise
except Exception as e:
if attempt == self.max_retries - 1:
raise
err_str = str(e)
if self._is_project_disabled_error(err_str):
raise ProjectDisabledError(err_str)
if "429" in err_str or "RESOURCE_EXHAUSTED" in err_str or "rate limit" in err_str.lower():
if len(self.api_urls) > 1:
# ๆๅคไธช project๏ผ็ซๅณๆๅบ่ฎฉ call() ๅๆข project
raise
# ๅ project ๆจกๅผ๏ผ้้ฟ้่ฏ
wait = (15 * (2 ** attempt)) + random.uniform(0, 5)
print(f" [RATE LIMIT] 429 detected, backing off {wait:.1f}s (attempt {attempt+1}/{self.max_retries})")
else:
wait = 2 ** attempt
await asyncio.sleep(wait)
raise RuntimeError("Unreachable")
async def _call_vertex_native(
self,
messages: List[Dict],
session: aiohttp.ClientSession,
base_url: str,
sem: asyncio.Semaphore,
temperature: float = None,
max_tokens: int = None,
) -> Tuple[str, str, TokenUsage]:
"""Vertex AI ๅ็ generateContent ่ฐ็จใ
base_url ๆ ผๅผ: https://aiplatform.googleapis.com/v1/projects/{P}/locations/{L}/publishers/google/models
ๅฎ้
่ฏทๆฑ URL = base_url/{model}:generateContent
"""
# ๆผๆฅๅฎๆด URL
# model ๅไธญๅฏ่ฝๆ "google/" ๅ็ผ๏ผไป OpenAI ๅ
ผๅฎน่ฟ็งป่ฟๆฅ๏ผ๏ผ้่ฆๅปๆ
model_name = self.model
if model_name.startswith("google/"):
model_name = model_name[len("google/"):]
full_url = f"{base_url}/{model_name}:generateContent"
# ่ฝฌๆขๆถๆฏๆ ผๅผ
system_instruction, contents = self._convert_messages_to_vertex_native(messages)
# ๆๅปบ่ฏทๆฑไฝ
payload: Dict[str, Any] = {
"contents": contents,
"generationConfig": {
"temperature": temperature or self.temperature,
"topP": self.top_p,
"maxOutputTokens": max_tokens or self.max_tokens,
},
}
if system_instruction:
payload["systemInstruction"] = system_instruction
for attempt in range(self.max_retries):
try:
headers = {"Content-Type": "application/json"}
if self._use_vertex_auth:
token = self._get_gcp_token(base_url)
headers["Authorization"] = f"Bearer {token}"
async with sem:
await self._acquire_vertex_request_slot()
async with session.post(
full_url, json=payload, headers=headers,
timeout=aiohttp.ClientTimeout(total=self.timeout),
) as resp:
if resp.status != 200:
error_body = await resp.text()
if _is_quota_exhausted(resp.status, error_body):
raise QuotaExhaustedError(
f"API ้ขๅบฆ็จๅฐฝ! HTTP {resp.status}: {error_body[:500]}")
raise RuntimeError(
f"HTTP {resp.status} from {full_url}: {error_body[:500]}")
result = await resp.json()
content, reasoning, usage = self._parse_vertex_native_response(result)
return content, reasoning, usage
except QuotaExhaustedError:
raise # ้ขๅบฆ็จๅฐฝ๏ผไธ้่ฏ๏ผ็ดๆฅไธๆ
except ProhibitedContentError:
raise # ๅ
ๅฎน่ขซๅฎๅ
จ็ญ็ฅๆฆๆช๏ผไธ้่ฏ๏ผ็ดๆฅไธๆ
except ProjectDisabledError:
raise
except Exception as e:
if attempt == self.max_retries - 1:
raise
err_str = str(e)
if self._is_project_disabled_error(err_str):
raise ProjectDisabledError(err_str)
if "429" in err_str or "RESOURCE_EXHAUSTED" in err_str or "rate limit" in err_str.lower():
if len(self.api_urls) > 1:
# ๆๅคไธช project๏ผ็ซๅณๆๅบ่ฎฉ call() ๅๆข project
raise
# ๅ project ๆจกๅผ๏ผ้้ฟ้่ฏ
wait = (15 * (2 ** attempt)) + random.uniform(0, 5)
print(f" [RATE LIMIT] 429 detected, backing off {wait:.1f}s (attempt {attempt+1}/{self.max_retries})")
else:
wait = 2 ** attempt
print(f" [WARN] Vertex native call failed (attempt {attempt+1}): {e}")
await asyncio.sleep(wait)
raise RuntimeError("Unreachable")
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
# Message Builders (OpenAI format)
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
def build_user_message(text: str, image_paths: List[str] = None) -> Dict:
"""Build user message with optional base64 images."""
if not image_paths:
return {"role": "user", "content": text}
content = []
for path in image_paths:
b64 = encode_image_b64(path)
content.append({
"type": "image_url",
"image_url": {"url": f"data:image/jpeg;base64,{b64}"}
})
content.append({"type": "text", "text": text})
return {"role": "user", "content": content}
def build_assistant_message(text: str) -> Dict:
return {"role": "assistant", "content": text}
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
# Response Parser (FIXED: position-based priority)
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
def parse_llm_response(text: str) -> Tuple[str, Any]:
"""Parse LLM response to extract the FIRST action by position.
Returns ('tool_call', dict), ('answer', str), or ('error', str).
CRITICAL FIX: Uses position-based priority instead of always preferring
<answer>. The first action tag that appears in the text wins.
This prevents hallucinated <answer> tags from overriding valid <tool_call> tags.
"""
# Find positions of first tool_call and first answer
tc_m = re.search(r'<tool_call>\s*(.*?)\s*</tool_call>', text, re.DOTALL)
answer_m = re.search(r'<answer>\s*(.*?)\s*</answer>', text, re.DOTALL)
if tc_m and answer_m:
# Both found โ use whichever comes FIRST in the text
if tc_m.start() < answer_m.start():
try:
return ("tool_call", json.loads(tc_m.group(1)))
except json.JSONDecodeError:
return ("error", f"Malformed tool_call JSON: {tc_m.group(1)[:200]}")
else:
return ("answer", answer_m.group(1).strip())
if tc_m:
try:
return ("tool_call", json.loads(tc_m.group(1)))
except json.JSONDecodeError:
return ("error", f"Malformed tool_call JSON: {tc_m.group(1)[:200]}")
if answer_m:
return ("answer", answer_m.group(1).strip())
return ("error", "No <tool_call> or <answer> found in response")
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
# Search Utilities
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
# โโ Image Search Cache โโ
class ImageSearchCache:
"""MD5-keyed cache for image search results."""
def __init__(self, cache_file: str):
self.cache_file = cache_file
self.cache = self._load()
self._dirty_count = 0
def _load(self) -> Dict[str, str]:
if os.path.exists(self.cache_file):
try:
with open(self.cache_file, "r", encoding="utf-8") as f:
return json.load(f)
except Exception:
return {}
return {}
def save(self):
with open(self.cache_file, "w", encoding="utf-8") as f:
json.dump(self.cache, f, ensure_ascii=False, indent=2)
self._dirty_count = 0
def get(self, image_bytes: bytes) -> Optional[str]:
key = hashlib.md5(image_bytes).hexdigest()
return self.cache.get(key)
def set(self, image_bytes: bytes, result: str):
key = hashlib.md5(image_bytes).hexdigest()
self.cache[key] = result
self._dirty_count += 1
if self._dirty_count >= 10:
self.save()
# โโ Mock Search Implementations โโ
def mock_image_search(entity: str, bbox: List[float]) -> str:
"""Mock image search result for offline testing."""
return (
f"Reverse Image Search Results:\n\n"
f"Result 1:\n"
f" Title: {entity} - Character Profile\n"
f" Snippet: {entity} is a well-known character/entity.\n"
f" URL: https://example.com/{entity.lower().replace(' ', '_')}\n\n"
f"Result 2:\n"
f" Title: {entity} | Wiki\n"
f" Snippet: Detailed information about {entity}.\n"
f" URL: https://wiki.example.com/{entity.lower().replace(' ', '_')}"
)
def mock_web_search(query: str) -> str:
"""Mock web search result for offline testing."""
return (
f'Web Search Results for "{query}":\n\n'
f"Quick Answer: Information related to the query.\n\n"
f"Result 1: {query} - Overview\n"
f" Relevant information about {query}.\n\n"
f"Result 2: {query} - Details\n"
f" Additional details and facts."
)
# โโ Real Search Implementations โโ
def _format_serper_lens_results(organic_results: List[Dict[str, Any]], max_results: int = 5) -> str:
"""Format raw Serper Lens organic results into readable text."""
parts = []
for i, item in enumerate(organic_results[:max_results], 1):
title = item.get('title', '')
snippet = item.get('snippet', '')
link = item.get('link', '')
source = item.get('source', '') or item.get('domain', '')
block = [f"Result {i}:"]
if title:
block.append(f" Title: {title}")
if snippet:
block.append(f" Snippet: {snippet}")
if source:
block.append(f" Source: {source}")
if link:
block.append(f" URL: {link}")
parts.append("\n".join(block))
return "\n\n".join(parts)
def _build_serper_lens_summary_prompt(organic_results: List[Dict[str, Any]], max_results: int = 5) -> str:
"""Build English summarizer prompt using only Serper Lens result metadata."""
context_parts = []
for i, item in enumerate(organic_results[:max_results], 1):
title = item.get('title', '')
snippet = item.get('snippet', '')
link = item.get('link', '')
source = item.get('source', '') or item.get('domain', '')
block = [f"Result {i}:"]
if title:
block.append(f"Title: {title}")
if snippet:
block.append(f"Snippet: {snippet}")
if source:
block.append(f"Source: {source}")
if link:
block.append(f"Link: {link}")
context_parts.append("\n".join(block))
context_text = "\n\n".join(context_parts)
return (
"You are a helpful assistant. Your task is to summarize the main content of the given "
"Serper Lens reverse image search results in no more than five sentences.\n\n"
"Your summary should cover the overall key points across the results, not just the parts "
"most related to the user's question.\n\n"
"If any part of the results is helpful for identifying the entity or answering the user's "
"question, include it clearly in the summary. Do not ignore relevant information, but make "
"sure the general structure and main ideas of the results are preserved.\n\n"
"Your summary should be concise, factual, and informative. If the results are ambiguous, "
"conflicting, or insufficient, clearly state that uncertainty.\n\n"
"Use only the provided result titles, snippets, and source/link metadata. Do not invent facts "
"and do not assume content from the linked pages.\n\n"
f"{context_text}"
)
async def summarize_serper_image_results(
organic_results: List[Dict[str, Any]],
session: aiohttp.ClientSession,
summarizer_address: str = "",
summarizer_model: str = "",
max_results: int = 5,
max_tokens: int = 512,
) -> Optional[str]:
"""Summarize Serper Lens results without fetching linked webpages."""
summarizer_addr = summarizer_address or IMAGE_SEARCH_SUMMARIZER_ADDRESS
sum_model = summarizer_model or IMAGE_SEARCH_SUMMARIZER_MODEL
if not organic_results or not summarizer_addr or not sum_model:
return None
summarizer_prompt = _build_serper_lens_summary_prompt(organic_results, max_results=max_results)
summarizer_payload = {
"model": sum_model,
"messages": [{"role": "user", "content": summarizer_prompt}],
"max_tokens": max_tokens,
"temperature": 0.3,
"chat_template_kwargs": {"enable_thinking": False},
}
try:
async with session.post(
f"http://{summarizer_addr}/v1/chat/completions",
json=summarizer_payload,
headers={"Content-Type": "application/json"},
timeout=aiohttp.ClientTimeout(total=120),
) as resp:
if resp.status != 200:
print(f" [IMAGE_SEARCH] Summarizer returned HTTP {resp.status}, falling back to raw results")
return None
data = await resp.json()
choices = data.get("choices", [])
if choices and isinstance(choices, list):
msg = choices[0].get("message", {})
summary = msg.get("content", "")
if summary and summary.strip():
summary = _strip_thinking_tags(summary).strip()
return summary or None
return None
except asyncio.TimeoutError:
print(" [IMAGE_SEARCH] Summarizer timeout, falling back to raw results")
return None
except Exception as e:
print(f" [IMAGE_SEARCH] Summarizer error: {e}, falling back to raw results")
return None
async def real_image_search(
image_b64_or_path: str,
session: aiohttp.ClientSession,
api_key: str,
crop_path: str = None,
) -> str:
"""ๅๅๅพ็ๆ็ดข๏ผๆ นๆฎ IMAGE_SEARCH_MODE ้ๆฉ็ด่ฟ Serper ๆๅ
ฌๅธๅ
้จ็ฝๅ
ณใ"""
if IMAGE_SEARCH_MODE == "gateway":
return await _gateway_image_search(image_b64_or_path, session, crop_path)
else:
return await _serper_image_search(image_b64_or_path, session, api_key, crop_path)
async def _gateway_image_search(
image_b64_or_path: str,
session: aiohttp.ClientSession,
crop_path: str = None,
) -> str:
"""ๅๅๅพ็ๆ็ดข via ๅ
ฌๅธๅ
้จ็ฝๅ
ณ โ Serper Google Lens + optional LLM summarization."""
if not GATEWAY_TOKEN:
return "Error: GATEWAY_TOKEN not configured."
image_url, prep_error = _prepare_image_search_url(
image_b64_or_path, crop_path, "IMAGE_SEARCH/GATEWAY"
)
if prep_error:
return prep_error
headers = {
'Content-Type': 'application/json',
'User-Agent': 'ifbook-http-client',
}
serper_params = {
"url": image_url,
"type": "lens",
}
gateway_payload = {
"sec_info": {
"username": GATEWAY_USERNAME,
"userid": GATEWAY_USERID,
"token": GATEWAY_TOKEN,
},
"model_type": "openai",
"model_name": "serper",
"params": json.dumps(serper_params),
}
max_api_retries = 2
last_error = None
for api_attempt in range(max_api_retries):
try:
async with session.post(
GATEWAY_URL,
headers=headers,
json=gateway_payload,
timeout=aiohttp.ClientTimeout(total=60),
) as resp:
if resp.status != 200:
error_body = await resp.text()
last_error = (f"Gateway error: HTTP {resp.status}: "
f"{error_body[:300]}")
if api_attempt < max_api_retries - 1:
print(f" [IMAGE_SEARCH/GATEWAY] HTTP {resp.status}, "
f"retrying ({api_attempt+1}/{max_api_retries})...")
await asyncio.sleep(3 * (api_attempt + 1))
continue
return last_error
gateway_resp = await resp.json()
model_output_str = gateway_resp.get("model_output", "{}")
data = json.loads(model_output_str)
organic = data.get('organic', [])
if not organic:
return "No results found from reverse image search."
raw_results = _format_serper_lens_results(
organic,
max_results=IMAGE_SEARCH_SUMMARIZER_MAX_RESULTS,
)
if not IMAGE_SEARCH_SUMMARIZE_SERPER:
return raw_results
summary = await summarize_serper_image_results(
organic,
session,
summarizer_address=IMAGE_SEARCH_SUMMARIZER_ADDRESS,
summarizer_model=IMAGE_SEARCH_SUMMARIZER_MODEL,
max_results=IMAGE_SEARCH_SUMMARIZER_MAX_RESULTS,
max_tokens=IMAGE_SEARCH_SUMMARIZER_MAX_TOKENS,
)
if summary:
return f"Summary: {summary}\n\nTop Lens Results:\n\n{raw_results}"
return raw_results
except asyncio.TimeoutError:
last_error = "Image search error: request timed out after 60s"
if api_attempt < max_api_retries - 1:
print(f" [IMAGE_SEARCH/GATEWAY] Timeout, "
f"retrying ({api_attempt+1}/{max_api_retries})...")
await asyncio.sleep(3 * (api_attempt + 1))
continue
return last_error
except Exception as e:
last_error = f"Image search error: {e}"
if api_attempt < max_api_retries - 1:
print(f" [IMAGE_SEARCH/GATEWAY] Error: {e}, "
f"retrying ({api_attempt+1}/{max_api_retries})...")
await asyncio.sleep(3 * (api_attempt + 1))
continue
return last_error
return last_error or "Image search error: unknown failure"
async def _serper_image_search(
image_b64_or_path: str,
session: aiohttp.ClientSession,
api_key: str,
crop_path: str = None,
) -> str:
"""ๅๅๅพ็ๆ็ดข via Serper Google Lens + optional LLM summarization๏ผๅๅง็ด่ฟๆนๅผ๏ผใ"""
if not api_key:
return "Error: SERPER_API_KEY not configured."
headers = {'X-API-KEY': api_key, 'Content-Type': 'application/json'}
image_url, prep_error = _prepare_image_search_url(
image_b64_or_path, crop_path, "IMAGE_SEARCH"
)
if prep_error:
return prep_error
max_api_retries = 2
last_error = None
for api_attempt in range(max_api_retries):
try:
async with session.post(
"https://google.serper.dev/lens",
headers=headers,
json={"url": image_url},
timeout=aiohttp.ClientTimeout(total=60),
) as resp:
if resp.status != 200:
error_body = await resp.text()
last_error = (f"Image search error: HTTP {resp.status}: "
f"{error_body[:300]}")
if api_attempt < max_api_retries - 1:
print(f" [IMAGE_SEARCH] HTTP {resp.status}, "
f"retrying ({api_attempt+1}/{max_api_retries})...")
await asyncio.sleep(3 * (api_attempt + 1))
continue
return last_error
data = await resp.json()
organic = data.get('organic', [])
if not organic:
return "No results found from reverse image search."
raw_results = _format_serper_lens_results(
organic,
max_results=IMAGE_SEARCH_SUMMARIZER_MAX_RESULTS,
)
if not IMAGE_SEARCH_SUMMARIZE_SERPER:
return raw_results
summary = await summarize_serper_image_results(
organic,
session,
summarizer_address=IMAGE_SEARCH_SUMMARIZER_ADDRESS,
summarizer_model=IMAGE_SEARCH_SUMMARIZER_MODEL,
max_results=IMAGE_SEARCH_SUMMARIZER_MAX_RESULTS,
max_tokens=IMAGE_SEARCH_SUMMARIZER_MAX_TOKENS,
)
if summary:
return f"Summary: {summary}\n\nTop Lens Results:\n\n{raw_results}"
return raw_results
except asyncio.TimeoutError:
last_error = "Image search error: request timed out after 60s"
if api_attempt < max_api_retries - 1:
print(f" [IMAGE_SEARCH] Timeout, "
f"retrying ({api_attempt+1}/{max_api_retries})...")
await asyncio.sleep(3 * (api_attempt + 1))
continue
return last_error
except Exception as e:
last_error = f"Image search error: {e}"
if api_attempt < max_api_retries - 1:
print(f" [IMAGE_SEARCH] Error: {e}, "
f"retrying ({api_attempt+1}/{max_api_retries})...")
await asyncio.sleep(3 * (api_attempt + 1))
continue
return last_error
return last_error or "Image search error: unknown failure"
async def real_web_search(
query: str,
session: aiohttp.ClientSession,
address: str = WEB_SEARCH_ADDRESS,
) -> str:
"""Web search via internal search server (SenseNova pattern)."""
payload = {
"query": query.strip().replace("\n", " "),
"top_k": 3,
"retrieval_mode": "google_serper",
**WEB_SEARCH_CONFIG,
}
try:
async with session.post(
f"http://{address}/search", json=payload,
timeout=aiohttp.ClientTimeout(total=100),
) as resp:
resp.raise_for_status()
return await resp.text()
except asyncio.TimeoutError:
return f"Error: Web search timeout for query: {query[:100]}"
except Exception as e:
return f"Error: Web search failed: {e}"
async def serper_web_search(
query: str,
session: aiohttp.ClientSession,
api_key: str,
) -> str:
"""Fallback web search via Serper Google Search."""
if not api_key:
return "Error: SERPER_API_KEY not configured."
headers = {'X-API-KEY': api_key, 'Content-Type': 'application/json'}
try:
async with session.post(
"https://google.serper.dev/search",
headers=headers,
json={"q": query},
timeout=aiohttp.ClientTimeout(total=20),
) as resp:
resp.raise_for_status()
data = await resp.json()
parts = []
ab = data.get('answerBox', {})
if ab:
answer = ab.get('answer') or ab.get('snippet', '')
if answer:
parts.append(f"Quick Answer: {answer}")
kg = data.get('knowledgeGraph', {})
if kg:
parts.append(
f"Knowledge Graph: {kg.get('title', '')} - "
f"{kg.get('description', '')}")
for i, item in enumerate(data.get('organic', [])[:5], 1):
parts.append(
f"Result {i}: {item.get('title', '')}\n"
f" {item.get('snippet', '')}")
return "\n".join(parts) if parts else "No relevant results found."
except Exception as e:
return f"Web search error: {e}"
# โโ MARS retrieval concurrency semaphore (lazy init) โโ
_mars_retrieval_semaphore: Optional[asyncio.Semaphore] = None
def _get_mars_retrieval_semaphore() -> Optional[asyncio.Semaphore]:
global _mars_retrieval_semaphore
if MARS_RETRIEVAL_CONCURRENCY > 0 and _mars_retrieval_semaphore is None:
_mars_retrieval_semaphore = asyncio.Semaphore(MARS_RETRIEVAL_CONCURRENCY)
return _mars_retrieval_semaphore if MARS_RETRIEVAL_CONCURRENCY > 0 else None
async def mars_web_search(
query: str,
session: aiohttp.ClientSession,
retrieval_address: str = "",
summarizer_address: str = "",
retrieval_topk: int = 3,
summarizer_model: str = "",
) -> str:
"""SenseNova-MARS style web search: retrieve from local Wikipedia + summarize via LLM.
Two-step pipeline:
1. POST to Search-R1 retrieval server โ get top-k document passages
2. POST to summarizer LLM (OpenAI-compatible /v1/chat/completions) โ get concise summary
This mirrors the SenseNova-MARS web_search_server architecture but without
the intermediate FastAPI layer.
"""
retrieval_addr = retrieval_address or MARS_RETRIEVAL_ADDRESS
summarizer_addr = summarizer_address or MARS_SUMMARIZER_ADDRESS
topk = retrieval_topk or MARS_RETRIEVAL_TOPK
sum_model = summarizer_model or MARS_SUMMARIZER_MODEL
if not retrieval_addr:
return "Error: MARS_RETRIEVAL_ADDRESS not configured."
if not summarizer_addr:
return "Error: MARS_SUMMARIZER_ADDRESS not configured."
clean_query = query.strip().replace("\n", " ")
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
# Step 1: Retrieve passages from Search-R1 retrieval server
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
retrieval_payload = {
"queries": [clean_query], # Search-R1 ่ฆๆฑ queries ๆฏ List[str]
"return_scores": True, # ๅฟ
้กปไธบ True๏ผๅฆๅๆๅก็ซฏ่งฃๅ
ๅดฉๆบ่ฟๅ 500
"topk": topk,
}
retrieved_passages = []
sem = _get_mars_retrieval_semaphore()
try:
if sem:
await sem.acquire()
try:
async with session.post(
f"http://{retrieval_addr}/retrieve",
json=retrieval_payload,
timeout=aiohttp.ClientTimeout(total=MARS_RETRIEVAL_TIMEOUT),
proxy="", # โ ็ป่ฟ็ณป็ปไปฃ็๏ผ็ด่ฟๅ
็ฝ
) as resp:
if resp.status != 200:
error_body = await resp.text()
return (f"Error: Retrieval server returned HTTP {resp.status}: "
f"{error_body[:300]}")
data = await resp.json()
# ๅทฒ็กฎ่ฎค็ Search-R1 ่ฟๅๆ ผๅผ:
# {"result": [[{"document": {"id": "xx", "contents": "\"title\"\ntext"}, "score": 0.84}, ...]]}
raw_results = data.get("result", [])
for query_results in raw_results:
if not isinstance(query_results, list):
continue
for item in query_results:
if not isinstance(item, dict):
continue
doc = item.get("document", {})
if isinstance(doc, dict):
contents = doc.get("contents", "")
elif isinstance(doc, str):
contents = doc
else:
continue
if contents:
lines = contents.split("\n", 1)
title = lines[0].strip('"') if lines else ""
text = lines[1] if len(lines) > 1 else contents
# ๆชๆญ่ฟ้ฟ็ passage๏ผ้ฟๅ
่ถ
ๅบ summarizer context window
retrieved_passages.append({"title": title, "text": text[:2000]})
finally:
if sem:
sem.release()
except asyncio.TimeoutError:
return f"Error: Retrieval server timeout for query: {clean_query[:100]}"
except (aiohttp.ClientConnectorError,
aiohttp.ServerDisconnectedError,
ConnectionRefusedError,
ConnectionResetError,
OSError) as e:
raise RetrieverDownError(
f"Retriever ๆๅก {retrieval_addr} ่ฟๆฅๅคฑ่ดฅ: {e}"
) from e
except Exception as e:
err_str = str(e).lower()
if any(kw in err_str for kw in (
"cannot connect", "connection refused", "connect call failed",
"server disconnected",
)):
raise RetrieverDownError(
f"Retriever ๆๅก {retrieval_addr} ่ฟๆฅๅคฑ่ดฅ: {e}"
) from e
return f"Error: Retrieval failed: {e}"
if not retrieved_passages:
return f"No relevant passages found for query: {clean_query}"
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
# Step 2: Summarize via LLM (OpenAI-compatible API)
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
context_parts = []
for i, p in enumerate(retrieved_passages, 1):
title_str = f" (Title: {p['title']})" if p['title'] else ""
context_parts.append(f"Passage {i}{title_str}:\n{p['text']}")
context_text = "\n\n".join(context_parts)
summarizer_prompt = (
f"Based on the following retrieved passages, provide a concise and informative "
f"summary that answers the query: \"{clean_query}\"\n\n"
f"{context_text}\n\n"
f"Please provide a concise summary focusing on the most relevant information. "
f"If the passages do not contain relevant information, say so."
)
summarizer_payload = {
"model": sum_model,
"messages": [{"role": "user", "content": summarizer_prompt}],
"max_tokens": 1024,
"temperature": 0.3,
"chat_template_kwargs": {"enable_thinking": False},
}
try:
async with session.post(
f"http://{summarizer_addr}/v1/chat/completions",
json=summarizer_payload,
headers={"Content-Type": "application/json"},
timeout=aiohttp.ClientTimeout(total=120),
proxy="", # โ ็ป่ฟ็ณป็ปไปฃ็๏ผ็ด่ฟๅ
็ฝ
) as resp:
if resp.status != 200:
print(f" [MARS_SEARCH] Summarizer returned HTTP {resp.status}, "
f"falling back to raw passages")
return _format_raw_passages(clean_query, retrieved_passages)
data = await resp.json()
choices = data.get("choices", [])
if choices and isinstance(choices, list):
msg = choices[0].get("message", {})
summary = msg.get("content", "")
if summary and summary.strip():
summary = _strip_thinking_tags(summary)
return (
f"Web Search Results for \"{clean_query}\" "
f"(MARS retrieve+summarize, {len(retrieved_passages)} passages):\n\n"
f"{summary.strip()}"
)
return _format_raw_passages(clean_query, retrieved_passages)
except asyncio.TimeoutError:
print(f" [MARS_SEARCH] Summarizer timeout, falling back to raw passages")
return _format_raw_passages(clean_query, retrieved_passages)
except Exception as e:
print(f" [MARS_SEARCH] Summarizer error: {e}, falling back to raw passages")
return _format_raw_passages(clean_query, retrieved_passages)
def _strip_thinking_tags(text: str) -> str:
"""Remove thinking content from summarizer output."""
# 1. ็งป้ค XML ้ฃๆ ผ็ <think>...</think>
text = re.sub(r'<think>.*?</think>', '', text, flags=re.DOTALL).strip()
# 2. ็งป้ค "Thinking Process:" ๅผๅคด็ๅ
ๅฎนๅ๏ผๅฐ็ฌฌไธไธช่ฟ็ปญ็ฉบ่กไธบๆญข๏ผ
text = re.sub(r'^Thinking Process:.*?(?=\n\n)', '', text, flags=re.DOTALL).strip()
# 3. ๅฆๆๆธ
็ๅ่ฟๆฏไปฅ "Thinking" ๅผๅคด๏ผ่พน็ๆ
ๅต๏ผ๏ผๅๆชไธๆฌก
if text.startswith("Thinking"):
parts = text.split("\n\n", 1)
if len(parts) > 1:
text = parts[1].strip()
return text
def _format_raw_passages(query: str, passages: list) -> str:
"""Format raw retrieved passages as fallback when summarizer fails."""
parts = [f"Web Search Results for \"{query}\" (raw retrieval, {len(passages)} passages):"]
for i, p in enumerate(passages, 1):
title_str = f" โ {p['title']}" if p['title'] else ""
text_preview = p['text'][:500] + ("..." if len(p['text']) > 500 else "")
parts.append(f"\nResult {i}{title_str}:\n {text_preview}")
return "\n".join(parts)
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
# I/O Helpers
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
def make_uid(entry: Dict) -> str:
"""Generate a unique ID from id + video_filename.
A single 'id' can map to multiple videos, so we need a composite key."""
eid = entry.get("id", "unknown")
vf = entry.get("video_filename", "")
if vf:
stem = os.path.splitext(vf)[0]
return f"{eid}__{stem}"
return eid
def load_completed_ids(output_file: str) -> Set[str]:
"""Load already-completed entry UIDs from output JSONL for resume."""
completed = set()
if not os.path.exists(output_file):
return completed
with open(output_file, "r", encoding="utf-8") as f:
for line in f:
line = line.strip()
if not line:
continue
try:
record = json.loads(line)
if "uid" in record and "error" not in record:
completed.add(record["uid"])
except json.JSONDecodeError:
pass
return completed
def get_question(entry: Dict) -> str:
"""Get the appropriate question field based on verdict."""
if entry.get("verdict") == "rewrite":
return entry.get("rewritten_question") or entry.get("original_question", "")
return entry.get("original_question", "")
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
# OSS Image Upload for Search (่งฃๅณ Serper Lens 400 Bad Request)
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
# ๅฏน้ฝ็ฌฌไบไปฝไปฃ็ ็ๆ่ทฏ๏ผๅ
ๆ่ฃๅชๅพไธไผ ๅฐๅ
ฌ็ฝๅฏ่ฎฟ้ฎ็ๅฏน่ฑกๅญๅจ โ
# ๆฟๅฐ https://... URL โ ไผ ็ป Serper Lens โ ่ทๅๆ็ดข็ปๆใ
#
# Serper Lens ๅฏน base64 data URI ๆฏๆไธ็จณๅฎ๏ผ้ข็น่ฟๅ 400๏ผ๏ผ
# ไฝๅฏน HTTPS URL ๅทฅไฝๆญฃๅธธใ
_oss_bucket = None # ๅปถ่ฟๅๅงๅ๏ผ้ฟๅ
ๆชๅฎ่ฃ
ๆถ็ดๆฅๆฅ้
_oss_init_attempted = False # ๅชๅฐ่ฏๅๅงๅไธๆฌก
_oss_last_error = ""
def _get_oss_bucket():
"""ๅปถ่ฟๅๅงๅ OSS Bucket ๅฏน่ฑก๏ผๅชๅๅงๅไธๆฌกใ"""
global _oss_bucket, _oss_init_attempted, _oss_last_error
if _oss_init_attempted:
return _oss_bucket
_oss_init_attempted = True
try:
import oss2
if not OSS_ACCESS_KEY_ID or not OSS_ACCESS_KEY_SECRET:
_oss_last_error = "OSS_ACCESS_KEY_ID / OSS_ACCESS_KEY_SECRET ๆช้
็ฝฎ"
print(f"[ERROR] {_oss_last_error}")
return None
auth = oss2.Auth(OSS_ACCESS_KEY_ID, OSS_ACCESS_KEY_SECRET)
_oss_bucket = oss2.Bucket(auth, f"https://{OSS_ENDPOINT}", OSS_BUCKET_NAME)
_oss_last_error = ""
print(f"[OSS] Bucket initialized: {OSS_BUCKET_NAME} @ {OSS_ENDPOINT}")
return _oss_bucket
except ImportError:
_oss_last_error = "oss2 ๆชๅฎ่ฃ
๏ผ่ฏทๆง่ก: pip install oss2 --break-system-packages"
print(f"[ERROR] {_oss_last_error}")
return None
except Exception as e:
_oss_last_error = f"OSS Bucket ๅๅงๅๅคฑ่ดฅ: {e}"
print(f"[ERROR] {_oss_last_error}")
return None
def optimize_crop_for_search(crop_path: str, output_path: str = None,
max_size: int = SEARCH_CROP_MAX_SIZE,
quality: int = SEARCH_CROP_JPEG_QUALITY) -> str:
"""ไผๅ่ฃๅชๅพ็จไบๆ็ดข๏ผ็ผฉๅฐๅฐบๅฏธ + ้ไฝ่ดจ้๏ผๅๅฐไธไผ ไฝ็งฏใ
Args:
crop_path: ๅๅง่ฃๅชๅพ่ทฏๅพ
output_path: ไผๅๅไฟๅญ่ทฏๅพ๏ผ้ป่ฎคๅจๅ็ฎๅฝๅ _opt ๅ็ผ๏ผ
max_size: ๆๅคง่พน้ฟ๏ผๅ็ด ๏ผ
quality: JPEG ่ดจ้
Returns:
ไผๅๅ็ๅพ็่ทฏๅพ๏ผๅฆๆไผๅๅคฑ่ดฅๅ่ฟๅๅ่ทฏๅพ๏ผใ
"""
if output_path is None:
base, ext = os.path.splitext(crop_path)
output_path = f"{base}_opt{ext}"
try:
with Image.open(crop_path) as img:
w, h = img.size
if max(w, h) > max_size:
ratio = max_size / max(w, h)
new_w, new_h = int(w * ratio), int(h * ratio)
img = img.resize((new_w, new_h), Image.LANCZOS)
img.save(output_path, 'JPEG', quality=quality)
return output_path
except Exception as e:
print(f" [WARN] optimize_crop_for_search failed: {e}, using original")
return crop_path
def _upload_to_oss(local_path: str, oss_object_name: str = None) -> Optional[str]:
"""ไธไผ ๅพ็ๅฐ้ฟ้ไบ OSS๏ผ่ฟๅๅ
ฌ็ฝ URLใ
ๅป้็ญ็ฅ๏ผ็จๆไปถๅ
ๅฎน็ MD5 ไฝไธบๅฏน่ฑกๅ๏ผ็ธๅๅ
ๅฎนไธ้ๅคไธไผ ใ
Args:
local_path: ๆฌๅฐๅพ็่ทฏๅพ
oss_object_name: OSS ๅฏน่ฑกๅ๏ผ้ป่ฎค็จ md5 hash ๅป้
Returns:
ๅ
ฌ็ฝ URL๏ผๅคฑ่ดฅ่ฟๅ None
"""
global _oss_last_error
bucket = _get_oss_bucket()
if bucket is None:
return None
try:
# ๅ
ไผๅๅพ็ๅฐบๅฏธๅ่ดจ้
opt_path = optimize_crop_for_search(local_path)
# ็จ md5 ไฝไธบๆไปถๅๅป้
with open(opt_path, "rb") as f:
file_bytes = f.read()
file_hash = hashlib.md5(file_bytes).hexdigest()
if oss_object_name is None:
oss_object_name = f"{OSS_UPLOAD_PREFIX}/{file_hash}.jpg"
# ๆฃๆฅๆฏๅฆๅทฒๅญๅจ๏ผhead_object ๆๅ่ฏดๆๅทฒไธไผ ่ฟ๏ผ
try:
bucket.head_object(oss_object_name)
# ๅทฒๅญๅจ๏ผ็ดๆฅ่ฟๅ URL
public_url = f"https://{OSS_BUCKET_NAME}.{OSS_ENDPOINT}/{oss_object_name}"
print(f" [OSS] Cache hit: {public_url}")
return public_url
except Exception:
# ไธๅญๅจๆๅ
ถไปๅผๅธธ๏ผ็ปง็ปญไธไผ
pass
# ไธไผ
bucket.put_object(oss_object_name, file_bytes, headers={
'Content-Type': 'image/jpeg',
})
public_url = f"https://{OSS_BUCKET_NAME}.{OSS_ENDPOINT}/{oss_object_name}"
print(f" [OSS] Uploaded: {public_url}")
return public_url
except Exception as e:
_oss_last_error = f"OSS upload failed: {e}"
print(f" [WARN] {_oss_last_error}")
return None
def _prepare_image_search_url(
image_b64_or_path: str,
crop_path: Optional[str],
log_prefix: str,
) -> Tuple[Optional[str], Optional[str]]:
"""ไธบๅพ็ๆ็ดขๅๅค URL๏ผ้ป่ฎค่ฆๆฑๅ
ไธไผ ๅฐ OSSใ"""
image_url = None
if crop_path and os.path.exists(crop_path):
image_url = _upload_to_oss(crop_path)
if image_url:
print(f" [{log_prefix}] Using OSS URL: {image_url}")
return image_url, None
if IMAGE_SEARCH_ALLOW_BASE64_FALLBACK:
image_url = f"data:image/jpeg;base64,{image_b64_or_path}"
print(f" [{log_prefix}] WARNING: OSS upload failed, falling back to "
f"base64 data URI (len={len(image_b64_or_path)}, may get 400)")
return image_url, None
failure_reason = _oss_last_error or "OSS upload returned None"
return None, (
"Image search unavailable: OSS upload failed and base64 fallback is disabled. "
f"Root cause: {failure_reason}"
)
|