File size: 78,941 Bytes
54c79d6 3fd8fc1 54c79d6 04b8201 54c79d6 04b8201 0c32859 04b8201 0c32859 76004d7 04b8201 76004d7 04b8201 46481c4 54c79d6 04b8201 0c32859 04b8201 54c79d6 04b8201 54c79d6 04b8201 54c79d6 04b8201 54c79d6 dd903d9 5312454 53ce6da 5312454 07a12d0 5312454 07a12d0 8a78e83 5312454 8a78e83 5312454 53ce6da 8a78e83 5312454 8a78e83 5312454 8a78e83 5312454 53ce6da 5312454 8a78e83 f656413 8a78e83 5312454 53ce6da 5312454 53ce6da 5312454 b4bf7c4 f63a0cf 859f876 f63a0cf c6143d7 f63a0cf 859f876 8a78e83 b4bf7c4 c6143d7 859f876 b4bf7c4 5312454 f63a0cf 5312454 b4bf7c4 f63a0cf 5312454 c6143d7 5312454 f63a0cf c6143d7 5312454 c6143d7 5312454 c6143d7 5312454 f63a0cf 080fac5 f63a0cf 080fac5 f63a0cf 080fac5 f63a0cf 080fac5 5312454 080fac5 f63a0cf e8728d6 0abb551 5312454 0abb551 b361d04 0abb551 5312454 0abb551 5312454 0abb551 5312454 0abb551 b361d04 0abb551 5312454 0abb551 b361d04 0abb551 bc3f415 5312454 bc3f415 b361d04 5312454 b716750 5312454 53ce6da 989e2c4 53ce6da 989e2c4 53ce6da b716750 5312454 8a78e83 5312454 8a78e83 5312454 7079bfe 5312454 9707317 67a025a 9707317 a10ede5 7d59f1b 67a025a 5312454 7079bfe 5312454 7079bfe 5312454 4b3dbbd 7079bfe 3662f50 5312454 7079bfe 5312454 7079bfe 5312454 bd702a2 5312454 7079bfe 5312454 8a78e83 9c6c431 46481c4 5312454 53ce6da 8a78e83 5312454 53ce6da 5312454 9c6c431 53ce6da 5312454 8a78e83 53ce6da 5312454 53ce6da 8a78e83 5312454 989e2c4 53ce6da 8a78e83 5312454 989e2c4 5312454 8a78e83 53ce6da 989e2c4 704ce6a 625bf6c bd702a2 704ce6a 625bf6c 859f876 704ce6a 859f876 bd702a2 989e2c4 859f876 bd702a2 989e2c4 704ce6a bd702a2 5312454 704ce6a bd702a2 859f876 5312454 8a78e83 53ce6da 5312454 53ce6da 5312454 989e2c4 49586ec 53ce6da 5312454 8a78e83 5312454 53ce6da 9c6c431 5312454 3dd5ddd 5312454 53ce6da 20cc82e 3dd5ddd 20cc82e 3dd5ddd 07a12d0 c2362ea 3dd5ddd c2362ea 07a12d0 c2362ea 07a12d0 c2362ea 20cc82e 3dd5ddd 20cc82e 3dd5ddd 20cc82e 3dd5ddd c2362ea 20cc82e c2362ea 20cc82e 3dd5ddd 20cc82e 3dd5ddd c2362ea 3dd5ddd 8a78e83 53ce6da 5312454 53ce6da 5312454 9c6c431 5312454 53ce6da 5312454 8a78e83 53ce6da 8a78e83 5312454 49586ec 07173d5 0c72122 07173d5 d551235 6240588 d551235 48d0150 d551235 48d0150 53ce6da 5312454 53ce6da 4b3dbbd 53ce6da 5312454 53ce6da 5312454 53ce6da 5312454 53ce6da 5312454 53ce6da 5312454 53ce6da 5312454 53ce6da 5312454 53ce6da 5312454 c2362ea 2a09b09 5312454 f92c7a2 5312454 53ce6da 5312454 53ce6da 5312454 53ce6da 5312454 53ce6da 5312454 f92c7a2 5312454 53ce6da 5312454 f92c7a2 5312454 704ce6a 5312454 704ce6a 5312454 a10ede5 53ce6da 8a78e83 b361d04 5312454 b361d04 5312454 b361d04 5312454 f92c7a2 5312454 53ce6da 5312454 53ce6da 5312454 b716750 54c79d6 dd903d9 54c79d6 1a201e4 0c32859 1a201e4 34518d3 1a201e4 54c79d6 1a201e4 34518d3 1a201e4 34518d3 1a201e4 34518d3 1a201e4 54c79d6 1a201e4 54c79d6 3fd8fc1 8e8119b 3fd8fc1 8e8119b 3fd8fc1 8e8119b dd903d9 8e8119b 3fd8fc1 8e8119b 3fd8fc1 8e8119b 3fd8fc1 54c79d6 04b8201 54c79d6 04b8201 54c79d6 dd903d9 04b8201 54c79d6 04b8201 dd903d9 54c79d6 04b8201 d6a02bb 04b8201 1dc5d45 04b8201 1dc5d45 d6a02bb 04b8201 d6a02bb 76004d7 54c79d6 dd903d9 54c79d6 dd903d9 54c79d6 8e8119b 0c32859 8e8119b 0c72122 76004d7 67a025a 76004d7 67a025a 76004d7 67a025a 76004d7 67a025a 76004d7 54c79d6 76004d7 8e8119b 04b8201 54c79d6 04b8201 8e8119b 54c79d6 04b8201 54c79d6 8e8119b 76004d7 54c79d6 76004d7 dd903d9 8e8119b 04b8201 54c79d6 04b8201 54c79d6 1a201e4 54c79d6 8e8119b 04b8201 54c79d6 8e8119b 54c79d6 8e8119b 54c79d6 8e8119b 54c79d6 1a201e4 0c32859 1a201e4 54c79d6 1a201e4 54c79d6 8e8119b 54c79d6 04b8201 76004d7 04b8201 8e8119b 04b8201 8e8119b 04b8201 8e8119b 04b8201 8e8119b 04b8201 8e8119b 04b8201 dd903d9 04b8201 dd903d9 8e8119b 04b8201 8e8119b 54c79d6 04b8201 54c79d6 04b8201 8e8119b 04b8201 8e8119b 04b8201 8e8119b 04b8201 8e8119b 04b8201 dd903d9 04b8201 dd903d9 04b8201 8e8119b 54c79d6 8e8119b 54c79d6 1a201e4 54c79d6 dd903d9 54c79d6 8e8119b 54c79d6 76004d7 54c79d6 8e8119b 69abb97 8e8119b 76004d7 04b8201 76004d7 04b8201 8e8119b 54c79d6 8e8119b 54c79d6 76004d7 54c79d6 76004d7 8e8119b 54c79d6 8e8119b 54c79d6 b361d04 9707317 a10ede5 2abe70f b361d04 9707317 a10ede5 9707317 b361d04 9707317 a10ede5 9707317 b361d04 90721a6 a10ede5 9707317 b361d04 a10ede5 b361d04 b716750 154aaf2 8a78e83 b716750 c6143d7 b716750 0abb551 5312454 0abb551 1a0445f a10ede5 5312454 0abb551 5312454 b361d04 7079bfe 080fac5 b716750 5312454 1a0445f 5312454 b716750 b361d04 b716750 a841871 53ce6da b716750 f656413 b716750 7079bfe 9c6c431 b361d04 0abb551 46481c4 b361d04 4b3dbbd b361d04 9c6c431 7079bfe 0abb551 7079bfe b361d04 5312454 7079bfe 5312454 46481c4 7079bfe 8a78e83 76004d7 f656413 76004d7 0c72122 f656413 d551235 0dbb238 cf22067 bca1f37 cf22067 0dbb238 07173d5 0dbb238 a10ede5 bca1f37 a10ede5 53ce6da bca1f37 8e8119b 53ce6da 07173d5 53ce6da f656413 53ce6da 49586ec 53ce6da 07173d5 b716750 a841871 5312454 154aaf2 5312454 f656413 54c79d6 76004d7 54c79d6 b361d04 f656413 a841871 54c79d6 | 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 | import json
import os
import re
import time
from dataclasses import dataclass, field
from datetime import date
from typing import Any, Dict, List, Optional, Set, Tuple, Union
import gradio as gr
import requests
from bs4 import BeautifulSoup
from duckduckgo_search import DDGS
from huggingface_hub import InferenceClient
# --- Model configuration ---------------------------------------------------
# Our own DeepResearch model. When QUEST_BASE_URL is configured in Space
# Secrets, the app will route requests to that dedicated HF Inference Endpoint
# instead of the shared HF Inference API.
QUEST_MODEL_ID = "osunlp/QUEST-35B"
QUEST_BASE_URL = os.getenv("QUEST_BASE_URL", "").strip()
# Endpoints built from the TGI image expose a single-model OpenAI route; the
# model name passed to chat_completion is usually "tgi". vLLM endpoints usually
# want the original repo id. QUEST_ENDPOINT_MODEL overrides this if needed.
QUEST_ENDPOINT_MODEL = os.getenv("QUEST_ENDPOINT_MODEL", "tgi").strip() or "tgi"
# This Space runs exclusively on QUEST-35B served via the private HF Inference
# Endpoint pointed to by QUEST_BASE_URL. No public fallback list — the model
# field in the UI is display-only.
DEFAULT_MODEL = QUEST_MODEL_ID
# Internal defaults. Search budget is no longer user-tunable.
DEFAULT_MAX_SEARCH_RESULTS = 10
PAPER_URL = os.getenv("PAPER_URL", "https://osu-nlp-group.github.io/quest-gh-test/")
CODE_URL = os.getenv("CODE_URL", "https://github.com/OSU-NLP-Group/QUEST")
DATASET_URL = os.getenv("DATASET_URL", "https://huggingface.co/collections/osunlp/quest")
MODEL_URL = os.getenv("MODEL_URL", "https://huggingface.co/osunlp/QUEST-35B-RL")
# --- System prompt ---------------------------------------------------------
# Full QUEST SYSTEM_PROMPT (mirrors inference/prompt.py in the research repo)
# so that QUEST-35B sees the exact tool schema it was trained with. Other
# models still follow this schema just fine in practice.
QUEST_SYSTEM_PROMPT = """You are a deep research assistant. Your core function is to conduct thorough, multi-source investigations into any topic. You must handle both broad, open-domain inquiries and queries within specialized academic fields. For every request, synthesize information from credible, diverse sources to deliver a comprehensive, accurate, and objective response. When you have gathered sufficient information and are ready to provide the definitive response, you must enclose the entire final answer within <answer></answer> tags.
# Tools
You may call one or more functions to assist with the user query.
You are provided with function signatures within <tools></tools> XML tags:
<tools>
{"type": "function", "function": {"name": "search", "description": "Perform Google web searches then returns a string of the top search results. Accepts multiple queries.", "parameters": {"type": "object", "properties": {"query": {"type": "array", "items": {"type": "string", "description": "The search query."}, "minItems": 1, "description": "The list of search queries."}}, "required": ["query"]}}}
{"type": "function", "function": {"name": "visit", "description": "Visit webpage(s) and return the summary of the content.", "parameters": {"type": "object", "properties": {"url": {"type": "array", "items": {"type": "string"}, "description": "The URL(s) of the webpage(s) to visit. Can be a single URL or an array of URLs."}, "goal": {"type": "string", "description": "The specific information goal for visiting webpage(s)."}}, "required": ["url", "goal"]}}}
</tools>
# Using prev_state (Research State Summary)
If you see a "RESEARCH STATE SUMMARY (prev_state)" section in the user message, it contains a compressed summary of previous research progress. Use it to avoid repeating searches/visits that have already been executed, use verified information directly in your answer, and follow up on uncertain claims only when needed.
For each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:
<tool_call>
{"name": <function-name>, "arguments": <args-json-object>}
</tool_call>
Current date: """
def build_system_prompt() -> str:
return QUEST_SYSTEM_PROMPT + date.today().isoformat()
TOOL_RESPONSE_TEMPLATE = """<tool_response>
{payload}
</tool_response>"""
SEARCH_CACHE: Dict[str, Dict[str, Any]] = {}
VISIT_CACHE: Dict[str, Dict[str, Any]] = {}
# Quest paper palette. The Gradio shell is themed to match the OSU-NLP Quest
# microsite: soft off-white page, paper-white cards, terracotta accent, mint
# secondary, Manrope for UI type and Source Serif 4 for display headings.
APP_THEME = gr.themes.Base(
primary_hue=gr.themes.colors.orange,
secondary_hue=gr.themes.colors.teal,
neutral_hue=gr.themes.colors.slate,
font=[
gr.themes.GoogleFont("Manrope"),
"ui-sans-serif",
"system-ui",
"sans-serif",
],
font_mono=[
gr.themes.GoogleFont("JetBrains Mono"),
"ui-monospace",
"monospace",
],
).set(
body_background_fill="#F2F4F8",
body_text_color="#0D1117",
body_text_color_subdued="#64748B",
color_accent="#BE5B2B",
color_accent_soft="rgba(190,91,43,0.09)",
background_fill_primary="#FFFFFF",
background_fill_secondary="#EEF1F7",
border_color_primary="rgba(10,15,40,0.08)",
border_color_accent="#BE5B2B",
block_background_fill="#FFFFFF",
block_border_width="1px",
block_border_color="rgba(10,15,40,0.08)",
block_shadow="0 1px 2px rgba(10,15,40,0.05), 0 2px 10px rgba(10,15,40,0.06)",
block_radius="16px",
block_label_background_fill="transparent",
block_label_border_width="0px",
block_label_text_color="#64748B",
block_label_text_weight="700",
block_title_text_color="#0D1117",
block_title_text_weight="700",
block_title_border_width="0px",
panel_background_fill="transparent",
panel_border_width="0px",
panel_border_color="transparent",
input_background_fill="#FFFFFF",
input_background_fill_focus="#FFFFFF",
input_border_color="rgba(10,15,40,0.12)",
input_border_color_focus="#BE5B2B",
input_border_width="1px",
input_radius="12px",
input_shadow="none",
input_shadow_focus="0 0 0 3px rgba(190,91,43,0.15)",
code_background_fill="#EEF1F7",
slider_color="#BE5B2B",
button_primary_background_fill="#0D1117",
button_primary_background_fill_hover="#1F2A37",
button_primary_text_color="#FFFFFF",
button_primary_border_color="transparent",
button_primary_shadow="0 1px 2px rgba(10,15,40,0.08), 0 6px 18px rgba(10,15,40,0.12)",
button_secondary_background_fill="#FFFFFF",
button_secondary_background_fill_hover="rgba(190,91,43,0.09)",
button_secondary_text_color="#BE5B2B",
button_secondary_border_color="rgba(10,15,40,0.16)",
button_cancel_background_fill="#FFFFFF",
button_cancel_background_fill_hover="#FEE2E2",
button_cancel_text_color="#DC2626",
button_cancel_border_color="#FCA5A5",
table_border_color="rgba(10,15,40,0.08)",
table_even_background_fill="#FAFBFD",
table_odd_background_fill="#FFFFFF",
)
CUSTOM_CSS = """
/* === Quest paper palette applied to the Gradio shell ==================== */
/* Brings the OSU-NLP Quest microsite aesthetic into the live Space: soft
off-white background, paper-white cards with subtle 1px borders and
low-opacity shadows, terracotta accent, Source Serif 4 for display
headings, Manrope for everything else. */
:root {
--q-bg: #F2F4F8;
--q-paper: #FFFFFF;
--q-surface-alt: #EEF1F7;
--q-line: rgba(10, 15, 40, 0.08);
--q-line-strong: rgba(10, 15, 40, 0.16);
--q-text: #0D1117;
--q-muted: #64748B;
--q-accent: #BE5B2B;
--q-accent-soft: rgba(190, 91, 43, 0.09);
--q-accent-line: rgba(190, 91, 43, 0.55);
--q-mint: #0B9E8A;
--q-mint-deep: #0A8070;
--q-cover-bg: #0D1117;
--q-shadow: 0 1px 3px rgba(10,15,40,0.04), 0 8px 32px rgba(10,15,40,0.08);
--q-shadow-card: 0 1px 2px rgba(10,15,40,0.05), 0 2px 10px rgba(10,15,40,0.06);
--q-radius-xl: 20px;
--q-radius-lg: 16px;
--q-radius-md: 12px;
}
html, body, gradio-app, [class*="gradio-container"] {
background: var(--q-bg) !important;
}
/* Full-height shell ------------------------------------------------------- */
html, body { width: 100% !important; min-height: 100vh !important; margin: 0 !important; font-size: 17px !important; }
gradio-app {
display: block !important;
width: 100% !important;
min-height: 100vh !important;
margin-left: auto !important;
margin-right: auto !important;
}
gradio-app > .gradio-container,
gradio-app > div {
display: block !important;
width: 100% !important;
margin-left: auto !important;
margin-right: auto !important;
}
[class*="gradio-container"] {
max-width: 1700px !important;
width: 100% !important;
min-width: 320px !important;
margin-left: auto !important;
margin-right: auto !important;
padding: 28px 36px 72px !important;
color: var(--q-text);
box-sizing: border-box !important;
font-family: "Manrope", ui-sans-serif, system-ui, sans-serif;
font-size: 1rem !important;
}
[class*="gradio-container"] *::selection { background: rgba(190,91,43,0.18); }
/* Prevent inner wrappers from collapsing when streaming content first arrives. */
[class*="gradio-container"] .layout-gap { width: 100% !important; }
[class*="gradio-container"] .layout-gap > .gr-column,
[class*="gradio-container"] .layout-gap > div { min-width: 0 !important; }
[class*="gradio-container"] .gradio-markdown,
[class*="gradio-container"] [data-testid="markdown"] { min-height: 220px !important; }
[class*="gradio-container"] .codemirror-wrapper,
[class*="gradio-container"] .cm-editor { min-height: 220px !important; }
/* Long code / markdown cannot push the layout sideways. */
[class*="gradio-container"] .gradio-code,
[class*="gradio-container"] .gradio-markdown,
[class*="gradio-container"] .prose,
[class*="gradio-container"] .markdown,
[class*="gradio-container"] [data-testid="markdown"],
[class*="gradio-container"] .tabs,
[class*="gradio-container"] .tabitem,
[class*="gradio-container"] .tab-content {
max-width: 100% !important;
width: 100% !important;
min-width: 0 !important;
word-wrap: break-word !important;
overflow-wrap: anywhere !important;
}
[class*="gradio-container"] .codemirror-wrapper {
max-width: 100% !important;
border-radius: 14px !important;
overflow: hidden !important;
}
[class*="gradio-container"] .cm-editor { max-width: 100% !important; overflow: hidden !important; }
[class*="gradio-container"] .cm-scroller { max-width: 100% !important; overflow-x: auto !important; }
[class*="gradio-container"] .cm-content,
[class*="gradio-container"] .cm-line {
max-width: 100% !important;
white-space: pre-wrap !important;
word-break: break-word !important;
}
[class*="gradio-container"] .prose pre,
[class*="gradio-container"] .markdown pre {
max-width: 100% !important;
overflow-x: auto !important;
white-space: pre-wrap !important;
}
/* === Quest-style header ================================================= */
.quest-header {
display: flex;
align-items: center;
gap: 18px;
padding: 18px 22px;
margin: 8px 0 24px;
border: 1px solid var(--q-line);
border-radius: var(--q-radius-lg);
background: var(--q-paper);
box-shadow: var(--q-shadow-card);
}
.quest-header-mark {
display: grid;
place-items: center;
width: 48px;
height: 48px;
flex-shrink: 0;
border-radius: 12px;
background: var(--q-text);
color: #FFFFFF;
font-family: "Source Serif 4", "Source Serif Pro", ui-serif, Georgia, serif;
font-weight: 700;
font-size: 1.55rem;
}
.quest-header-text {
display: flex;
flex-direction: column;
gap: 4px;
min-width: 0;
}
.quest-header-title {
margin: 0;
font-family: "Source Serif 4", "Source Serif Pro", ui-serif, Georgia, serif;
font-weight: 600;
font-size: clamp(1.25rem, 2vw, 1.75rem);
line-height: 1.2;
letter-spacing: -0.01em;
color: var(--q-text);
}
.quest-header-byline {
color: var(--q-muted);
font-size: 0.9rem;
font-weight: 500;
text-decoration: underline;
text-decoration-color: rgba(100,116,139,0.45);
text-underline-offset: 3px;
text-decoration-thickness: 1px;
width: fit-content;
transition: color 140ms ease, text-decoration-color 140ms ease;
}
.quest-header-byline:hover {
color: var(--q-accent);
text-decoration-color: var(--q-accent);
}
/* === Cards (section-card) =============================================== */
.section-card {
background: var(--q-paper) !important;
border: 1px solid var(--q-line) !important;
border-radius: var(--q-radius-xl) !important;
box-shadow: var(--q-shadow-card) !important;
padding: 22px !important;
}
.no-frame {
background: transparent !important;
border: none !important;
box-shadow: none !important;
padding: 0 !important;
}
/* Section kicker + hero heading follow the paper treatment. */
.section-heading {
font-size: 0.7rem;
font-weight: 800;
letter-spacing: 0.14em;
text-transform: uppercase;
color: var(--q-accent);
margin: 0 0 14px 0;
}
.hero-heading {
font-family: "Source Serif 4", "Source Serif Pro", ui-serif, Georgia, serif !important;
font-weight: 600 !important;
font-size: 1.6rem !important;
letter-spacing: -0.01em !important;
text-transform: none !important;
color: var(--q-text) !important;
}
/* Match the .brand mark from the Quest microsite (github-page branch). */
.quest-name {
font-family: "Source Serif 4", "Source Serif Pro", ui-serif, Georgia, serif !important;
font-style: italic !important;
font-weight: 700 !important;
color: inherit !important;
letter-spacing: -0.005em;
margin: 4px 0 14px 0 !important;
}
.hero-subtitle {
color: var(--q-muted);
font-size: 0.95rem;
line-height: 1.6;
margin: -6px 0 16px 0;
}
/* Layout gap: mirror the paper's column rhythm. */
.layout-gap { gap: 24px !important; align-items: flex-start; }
.right-stack > * { margin-bottom: 14px; }
.action-row { gap: 10px !important; margin-top: 14px; }
.action-row button { min-width: 0; flex: 1; }
/* === Icon grid (Paper / Code / Dataset / Model) ========================= */
.icon-grid {
display: grid;
grid-template-columns: repeat(2, minmax(0, 1fr));
gap: 10px;
width: 100%;
}
.icon-link {
display: flex;
align-items: center;
justify-content: center;
gap: 8px;
padding: 11px 14px;
border-radius: 999px;
text-decoration: none !important;
color: var(--q-text) !important;
background: var(--q-paper);
font-weight: 600;
font-size: 0.88rem;
white-space: nowrap;
border: 1px solid var(--q-line-strong);
transition: background 140ms ease, border-color 140ms ease, color 140ms ease, transform 140ms ease;
}
.icon-link:hover {
background: var(--q-accent-soft);
border-color: var(--q-accent-line);
color: var(--q-accent) !important;
transform: translateY(-1px);
}
/* Resource cards (paper / code / data / model) — icon + label, eye-catching */
.resource-grid {
display: grid;
grid-template-columns: repeat(2, minmax(0, 1fr));
gap: 10px;
width: 100%;
}
.resource-card {
display: flex;
align-items: center;
gap: 10px;
padding: 12px 14px;
border-radius: 14px;
text-decoration: none !important;
color: var(--q-text) !important;
background: var(--q-paper);
border: 1px solid var(--q-line-strong);
transition: background 140ms ease, border-color 140ms ease, color 140ms ease, transform 140ms ease;
}
.resource-card:hover {
background: var(--q-accent-soft);
border-color: var(--q-accent-line);
color: var(--q-accent) !important;
transform: translateY(-1px);
}
.resource-card-icon {
display: inline-flex;
align-items: center;
justify-content: center;
width: 30px;
height: 30px;
flex-shrink: 0;
border-radius: 8px;
background: var(--q-surface-alt);
color: var(--q-text);
}
.resource-card-icon svg {
width: 18px;
height: 18px;
fill: currentColor;
}
.resource-card-icon.resource-card-emoji {
background: transparent;
font-size: 22px;
line-height: 1;
}
.resource-card-text {
display: flex;
flex-direction: column;
line-height: 1.15;
min-width: 0;
}
.resource-card-text strong {
font-weight: 700;
font-size: 0.92rem;
}
.resource-card-text small {
font-size: 0.72rem;
color: var(--q-muted);
margin-top: 2px;
}
/* === Buttons ============================================================ */
[class*="gradio-container"] button.primary,
[class*="gradio-container"] .gr-button-primary {
background: var(--q-text) !important;
color: #ffffff !important;
border: 1px solid var(--q-text) !important;
box-shadow: 0 1px 2px rgba(10,15,40,0.08), 0 6px 18px rgba(10,15,40,0.12) !important;
font-weight: 700 !important;
letter-spacing: 0.01em !important;
}
[class*="gradio-container"] button.primary:hover,
[class*="gradio-container"] .gr-button-primary:hover {
background: #1F2A37 !important;
border-color: #1F2A37 !important;
}
[class*="gradio-container"] button.secondary,
[class*="gradio-container"] .gr-button-secondary {
background: var(--q-paper) !important;
color: var(--q-text) !important;
border: 1px solid var(--q-line-strong) !important;
box-shadow: none !important;
font-weight: 600 !important;
}
[class*="gradio-container"] button.secondary:hover,
[class*="gradio-container"] .gr-button-secondary:hover {
background: var(--q-accent-soft) !important;
border-color: var(--q-accent-line) !important;
color: var(--q-accent) !important;
}
[class*="gradio-container"] button.stop,
[class*="gradio-container"] .gr-button-stop {
background: var(--q-paper) !important;
color: #DC2626 !important;
border: 1px solid #FCA5A5 !important;
box-shadow: none !important;
font-weight: 600 !important;
}
[class*="gradio-container"] button.stop:hover,
[class*="gradio-container"] .gr-button-stop:hover {
background: #FEE2E2 !important;
color: #B91C1C !important;
}
/* Flatten every grey block Gradio drops inside our cards. */
[class*="gradio-container"] .gr-group,
[class*="gradio-container"] fieldset,
[class*="gradio-container"] .gr-box,
[class*="gradio-container"] .gr-panel,
[class*="gradio-container"] .form,
[class*="gradio-container"] .gr-form,
[class*="gradio-container"] .container {
background: transparent !important;
}
.section-card {
--block-shadow: none;
--block-shadow-dark: none;
--block-background-fill: transparent;
--block-border-color: transparent;
--block-border-width: 0px;
--panel-background-fill: transparent;
--panel-border-width: 0px;
--background-fill-secondary: transparent;
--border-color-primary: transparent;
overflow: visible !important;
}
.section-card > div,
.section-card > div > div,
.section-card > div > div > div {
background: transparent !important;
border: none !important;
box-shadow: none !important;
overflow: visible !important;
}
.section-card .block,
.section-card .form,
.section-card .gr-form,
.section-card .gr-block,
.section-card .gr-panel,
.section-card .gr-group,
.section-card .gradio-dropdown,
.section-card .gradio-slider,
.section-card .gradio-textbox,
.section-card .gradio-markdown,
.section-card .gradio-code {
background: transparent !important;
border: none !important;
box-shadow: none !important;
overflow: visible !important;
}
.section-card .form,
.section-card .gr-form {
display: flex !important;
flex-direction: column !important;
gap: 14px !important;
}
[class*="gradio-container"] .section-card .row,
[class*="gradio-container"] .section-card [class*="row"] {
display: flex !important;
flex-direction: row !important;
flex-wrap: wrap !important;
gap: 10px !important;
}
.action-row {
display: flex !important;
flex-direction: row !important;
gap: 10px !important;
margin-top: 14px;
}
.action-row > * { flex: 1 1 0; min-width: 0; }
.section-card > * + * { margin-top: 14px; }
/* === Inputs ============================================================= */
[class*="gradio-container"] textarea,
[class*="gradio-container"] input:not([type="checkbox"]):not([type="radio"]):not([type="range"]) {
background: var(--q-paper) !important;
border: 1px solid var(--q-line-strong) !important;
box-shadow: none !important;
border-radius: var(--q-radius-md) !important;
color: var(--q-text) !important;
font-family: "Manrope", ui-sans-serif, system-ui, sans-serif !important;
}
/* Make the Model Textbox match the Memory Strategy Dropdown's height (46px outer = 44px content + 2*1px border). */
.section-card [data-testid="textbox"] textarea,
.section-card [data-testid="textbox"] input {
min-height: 44px !important;
padding: 11px 14px !important;
line-height: 1.4 !important;
box-sizing: border-box !important;
}
[class*="gradio-container"] textarea::placeholder,
[class*="gradio-container"] input::placeholder { color: #94A3B8 !important; }
[class*="gradio-container"] textarea:focus,
[class*="gradio-container"] input:focus {
border-color: var(--q-accent) !important;
box-shadow: 0 0 0 3px rgba(190,91,43,0.15) !important;
outline: none !important;
}
/* === Dropdown =========================================================== */
[class*="gradio-container"] [data-testid="dropdown"],
[class*="gradio-container"] .gradio-dropdown {
background: var(--q-paper) !important;
border: 1px solid var(--q-line-strong) !important;
border-radius: var(--q-radius-md) !important;
box-shadow: none !important;
padding: 0 !important;
min-height: 46px !important;
width: 100% !important;
box-sizing: border-box !important;
}
[class*="gradio-container"] [data-testid="dropdown"] > .wrap,
[class*="gradio-container"] [data-testid="dropdown"] .secondary-wrap,
[class*="gradio-container"] [data-testid="dropdown"] .wrap-inner,
[class*="gradio-container"] [data-testid="dropdown"] .input-container,
[class*="gradio-container"] [data-testid="dropdown"] .single-select,
[class*="gradio-container"] .gradio-dropdown .wrap,
[class*="gradio-container"] .gradio-dropdown .wrap-inner,
[class*="gradio-container"] .gradio-dropdown .secondary-wrap,
[class*="gradio-container"] .gradio-dropdown .input-container,
[class*="gradio-container"] .gradio-dropdown .single-select,
[class*="gradio-container"] [class*="dropdown"] .wrap {
background: transparent !important;
border: 0 !important;
outline: 0 !important;
box-shadow: none !important;
border-radius: 0 !important;
width: 100% !important;
min-height: 44px !important;
padding: 0 14px !important;
display: flex !important;
align-items: center !important;
box-sizing: border-box !important;
}
[class*="gradio-container"] [data-testid="dropdown"] input,
[class*="gradio-container"] .gradio-dropdown input,
[class*="gradio-container"] [data-testid="dropdown"] select,
[class*="gradio-container"] .gradio-dropdown select {
background: transparent !important;
border: 0 !important;
outline: 0 !important;
box-shadow: none !important;
padding: 0 !important;
height: 44px !important;
line-height: 44px !important;
font-size: 0.95rem !important;
width: 100% !important;
border-radius: 0 !important;
}
/* Force-remove any nested pill/rounded background that makes the dropdown
look like it has two concentric frames. */
[class*="gradio-container"] [data-testid="dropdown"] .container,
[class*="gradio-container"] [data-testid="dropdown"] .wrap > .wrap,
[class*="gradio-container"] .gradio-dropdown .container,
[class*="gradio-container"] .gradio-dropdown .wrap > .wrap {
border: 0 !important;
outline: 0 !important;
box-shadow: none !important;
background: transparent !important;
border-radius: 0 !important;
padding: 0 !important;
}
/* The little caret/arrow icon container — vertically center it */
[class*="gradio-container"] [data-testid="dropdown"] .icon-wrap,
[class*="gradio-container"] .gradio-dropdown .icon-wrap {
top: 50% !important;
transform: translateY(-50%) !important;
right: 14px !important;
}
[class*="gradio-container"] .options ul,
[class*="gradio-container"] .options {
background: var(--q-paper) !important;
border: 1px solid var(--q-line) !important;
border-radius: var(--q-radius-md) !important;
box-shadow: 0 10px 30px rgba(10,15,40,0.12) !important;
}
[class*="gradio-container"] .options li[aria-selected="true"],
[class*="gradio-container"] .options li:hover {
background: var(--q-accent-soft) !important;
color: var(--q-accent) !important;
}
/* Info hint text under inputs */
[class*="gradio-container"] .info,
[class*="gradio-container"] [data-testid*="info"],
[class*="gradio-container"] .gr-info {
color: var(--q-muted) !important;
background: transparent !important;
font-size: 12px !important;
}
/* === Sliders ============================================================ */
/* Flatten the Slider's outer wrapper — Gradio paints a rectangular block
around the label + track + value-input by default; remove it. */
.section-card .gradio-slider,
.section-card .gradio-slider > div,
.section-card .gradio-slider .form,
.section-card .gradio-slider .gr-form,
.section-card .gradio-slider .wrap,
.section-card .gradio-slider .container,
.section-card .gradio-slider .head {
background: transparent !important;
border: 0 !important;
box-shadow: none !important;
padding: 0 !important;
}
/* === Per-component flatteners (id-based; max specificity vs Gradio defaults) === */
/* The Memory Strategy dropdown and the two sliders ship with an outer block
wrapper that paints a small rectangle. Flatten the wrapper AND any nested
div Gradio inserts (form/container/wrap/etc), keeping label + interactive
element visible. */
#quest-memory-strategy,
#quest-memory-strategy > div,
#quest-memory-strategy .form,
#quest-memory-strategy .gr-form,
#quest-memory-strategy .container,
#quest-memory-strategy .wrap-inner,
#quest-memory-strategy .head,
#quest-max-turns,
#quest-max-turns > div,
#quest-max-turns .form,
#quest-max-turns .gr-form,
#quest-max-turns .container,
#quest-max-turns .wrap-inner,
#quest-max-turns .head,
#quest-temperature,
#quest-temperature > div,
#quest-temperature .form,
#quest-temperature .gr-form,
#quest-temperature .container,
#quest-temperature .wrap-inner,
#quest-temperature .head,
#quest-model,
#quest-model > div,
#quest-model .form,
#quest-model .gr-form,
#quest-model .container,
#quest-model .wrap-inner,
#quest-model .head {
background: transparent !important;
border: 0 !important;
outline: 0 !important;
box-shadow: none !important;
padding: 0 !important;
border-radius: 0 !important;
}
/* Memory Strategy radio: stack vertically, terracotta-tinted check state. */
#quest-memory-strategy .wrap,
#quest-memory-strategy fieldset,
#quest-memory-strategy [data-testid="radio"] {
display: flex !important;
flex-direction: column !important;
gap: 6px !important;
background: transparent !important;
border: 0 !important;
padding: 0 !important;
}
#quest-memory-strategy label {
background: transparent !important;
border: 1px solid var(--q-line) !important;
border-radius: 8px !important;
padding: 8px 12px !important;
cursor: pointer !important;
font-weight: 500 !important;
font-size: 0.95rem !important;
color: var(--q-text) !important;
text-transform: none !important;
letter-spacing: 0 !important;
display: flex !important;
align-items: center !important;
gap: 10px !important;
transition: border-color 120ms ease, background 120ms ease;
}
#quest-memory-strategy label:hover {
border-color: var(--q-line-strong) !important;
}
#quest-memory-strategy input[type="radio"] {
accent-color: var(--q-accent) !important;
width: 16px !important;
height: 16px !important;
}
/* Slider head input (the "[6 ↺]" / "[1 ↺]" pill next to the slider track):
the global input rule paints a 1px border on it, which looks like a stray
rectangle. Flatten it AND hide the reset button (it's redundant — the
slider's range already shows the default value). */
#quest-max-turns input[type="number"],
#quest-temperature input[type="number"] {
border: 0 !important;
background: transparent !important;
box-shadow: none !important;
border-radius: 0 !important;
padding: 0 !important;
min-height: 0 !important;
height: auto !important;
text-align: center !important;
width: 3.5em !important;
font-weight: 600 !important;
color: var(--q-text) !important;
}
#quest-max-turns button,
#quest-temperature button {
display: none !important;
}
[class*="gradio-container"] input[type="range"] {
-webkit-appearance: none;
appearance: none;
width: 100%;
height: 6px;
background: var(--q-surface-alt);
border-radius: 999px;
outline: none;
box-shadow: none !important;
border: none !important;
}
[class*="gradio-container"] input[type="range"]::-webkit-slider-runnable-track {
height: 6px;
background: linear-gradient(90deg,var(--q-accent) var(--val,50%), var(--q-surface-alt) var(--val,50%));
border-radius: 999px;
}
[class*="gradio-container"] input[type="range"]::-webkit-slider-thumb {
-webkit-appearance: none;
appearance: none;
width: 18px;
height: 18px;
border-radius: 50%;
background: #ffffff;
border: 2px solid var(--q-accent);
box-shadow: 0 2px 6px rgba(190,91,43,0.25);
margin-top: -6px;
cursor: pointer;
}
[class*="gradio-container"] input[type="range"]::-moz-range-track {
height: 6px;
background: var(--q-surface-alt);
border-radius: 999px;
}
[class*="gradio-container"] input[type="range"]::-moz-range-progress {
height: 6px;
background: var(--q-accent);
border-radius: 999px;
}
[class*="gradio-container"] input[type="range"]::-moz-range-thumb {
width: 16px;
height: 16px;
border-radius: 50%;
background: #ffffff;
border: 2px solid var(--q-accent);
box-shadow: 0 2px 6px rgba(190,91,43,0.25);
}
/* === Tabs =============================================================== */
[class*="gradio-container"] .tabs,
[class*="gradio-container"] .tab-container,
[class*="gradio-container"] .tab-wrapper { background: transparent !important; }
[class*="gradio-container"] .tab-container::after { background: var(--q-line) !important; }
[class*="gradio-container"] .tab-wrapper button {
color: var(--q-muted) !important;
font-weight: 700 !important;
letter-spacing: 0.04em !important;
text-transform: uppercase !important;
font-size: 0.78rem !important;
}
[class*="gradio-container"] .tab-wrapper button.selected { color: var(--q-accent) !important; }
[class*="gradio-container"] .tab-wrapper button.selected::after { background: var(--q-accent) !important; }
/* Hide the orange streaming-progress bar that Gradio paints at the top of
the Markdown/Code panel while a run is in flight. */
[class*="gradio-container"] .progress,
[class*="gradio-container"] .progress-level,
[class*="gradio-container"] .progress-level-inner,
[class*="gradio-container"] .progress-bar,
[class*="gradio-container"] .progress-text,
[class*="gradio-container"] [class*="progress-level"],
[class*="gradio-container"] .generating,
[class*="gradio-container"] div[class*="progress-bar"] {
display: none !important;
background: transparent !important;
border: 0 !important;
height: 0 !important;
}
/* Kill any stray orange/thick separator that Gradio paints above the tab
panel content (border-top or ::before on the tab content wrapper). */
[class*="gradio-container"] .tabitem,
[class*="gradio-container"] .tab-content,
[class*="gradio-container"] .gradio-tabitem,
[class*="gradio-container"] .tabs > div.tabitem {
border-top: 0 !important;
box-shadow: none !important;
background: transparent !important;
}
[class*="gradio-container"] .tabitem::before,
[class*="gradio-container"] .tab-content::before,
[class*="gradio-container"] .gradio-tabitem::before { content: none !important; }
[class*="gradio-container"] .tab-nav,
[class*="gradio-container"] .tab-wrapper {
border-bottom: 1px solid var(--q-line) !important;
border-top: 0 !important;
}
[class*="gradio-container"] .tab-nav::before,
[class*="gradio-container"] .tab-wrapper::before { content: none !important; }
/* Block labels above components */
[class*="gradio-container"] .gr-block label,
[class*="gradio-container"] .gradio-slider label,
[class*="gradio-container"] .gradio-dropdown label,
[class*="gradio-container"] .gradio-textbox label {
color: var(--q-muted) !important;
font-weight: 700 !important;
font-size: 0.74rem !important;
letter-spacing: 0.08em !important;
text-transform: uppercase !important;
}
/* === Markdown / prose =================================================== */
[class*="gradio-container"] .gr-markdown,
[class*="gradio-container"] .prose,
[class*="gradio-container"] .markdown {
color: var(--q-text) !important;
font-family: "Manrope", ui-sans-serif, system-ui, sans-serif !important;
line-height: 1.75;
}
[class*="gradio-container"] .gr-markdown a,
[class*="gradio-container"] .prose a { color: var(--q-accent) !important; text-decoration: underline; text-decoration-color: rgba(190,91,43,0.35); }
[class*="gradio-container"] .gr-markdown a:hover,
[class*="gradio-container"] .prose a:hover { text-decoration-color: var(--q-accent); }
[class*="gradio-container"] .gr-markdown h1,
[class*="gradio-container"] .gr-markdown h2,
[class*="gradio-container"] .gr-markdown h3,
[class*="gradio-container"] .prose h1,
[class*="gradio-container"] .prose h2,
[class*="gradio-container"] .prose h3 {
font-family: "Source Serif 4", "Source Serif Pro", ui-serif, Georgia, serif !important;
font-weight: 600 !important;
letter-spacing: -0.01em !important;
color: var(--q-text) !important;
}
[class*="gradio-container"] .gr-markdown code,
[class*="gradio-container"] .prose code {
background: var(--q-surface-alt);
border: 1px solid var(--q-line);
padding: 1px 6px;
border-radius: 6px;
font-size: 0.9em;
}
/* === Code block (Record tab) ============================================ */
[class*="gradio-container"] .codemirror-wrapper,
[class*="gradio-container"] .cm-editor,
[class*="gradio-container"] .cm-scroller,
[class*="gradio-container"] .cm-gutters,
[class*="gradio-container"] .cm-content {
background: var(--q-surface-alt) !important;
color: var(--q-text) !important;
border: none !important;
font-family: "JetBrains Mono", ui-monospace, monospace !important;
}
[class*="gradio-container"] .cm-gutters {
border-right: 1px solid var(--q-line) !important;
color: var(--q-muted) !important;
}
/* === Rounded corners on everything ====================================== */
[class*="gradio-container"] .block,
[class*="gradio-container"] .form,
[class*="gradio-container"] .gr-box,
[class*="gradio-container"] .gr-panel,
[class*="gradio-container"] .gr-group,
[class*="gradio-container"] [data-testid="textbox"],
[class*="gradio-container"] [data-testid="dropdown"],
[class*="gradio-container"] .tabitem,
[class*="gradio-container"] .tab-content,
[class*="gradio-container"] .gradio-markdown,
[class*="gradio-container"] .gradio-code { border-radius: var(--q-radius-md) !important; }
[class*="gradio-container"] button { border-radius: 999px !important; }
/* === Example buttons ==================================================== */
.example-note { color: var(--q-muted); font-size: 13px; margin: 0 0 12px 0; line-height: 1.5; }
.memory-help {
color: var(--q-muted);
font-size: 12.5px;
line-height: 1.55;
margin: 6px 0 0 0;
padding: 10px 12px;
background: var(--q-surface-alt);
border: 1px solid var(--q-line);
border-radius: 8px;
}
.memory-help b { color: var(--q-text); font-weight: 600; }
.example-buttons { display: grid; gap: 10px; margin-top: 4px; }
[class*="gradio-container"] .example-btn {
text-align: left !important;
justify-content: flex-start !important;
white-space: normal !important;
line-height: 1.5 !important;
padding: 14px 16px !important;
font-size: 14px !important;
color: var(--q-text) !important;
background: var(--q-paper) !important;
border: 1px solid var(--q-line) !important;
border-radius: var(--q-radius-md) !important;
box-shadow: none !important;
font-weight: 500 !important;
letter-spacing: normal !important;
text-transform: none !important;
}
[class*="gradio-container"] .example-btn:hover {
background: var(--q-accent-soft) !important;
border-color: var(--q-accent-line) !important;
color: var(--q-accent) !important;
}
[class*="gradio-container"] .example-btn > * {
color: inherit !important;
white-space: normal !important;
display: inline !important;
}
/* Footer tagline block */
.quest-footer {
margin-top: 28px;
padding: 18px 24px;
border: 1px solid var(--q-line);
border-radius: var(--q-radius-xl);
background: var(--q-paper);
box-shadow: var(--q-shadow-card);
display: flex;
align-items: center;
justify-content: space-between;
gap: 20px;
color: var(--q-muted);
font-size: 0.86rem;
line-height: 1.65;
}
.quest-footer a { color: var(--q-muted); text-decoration: none; }
.quest-footer a:hover { color: var(--q-text); }
.quest-footer-links { display: flex; gap: 16px; flex-wrap: wrap; }
/* Tiny mark that replaces the HF watermark block. */
footer { display: none !important; }
/* === Responsive ========================================================= */
@media (max-width: 1100px) {
.quest-cover-inner { grid-template-columns: 1fr; }
.quest-cover-panel.wide { grid-column: auto; min-height: 180px; }
}
@media (max-width: 760px) {
[class*="gradio-container"] { padding: 16px !important; }
.quest-footer { flex-direction: column; align-items: flex-start; }
}
"""
@dataclass
class AgentState:
searched_queries: List[str] = field(default_factory=list)
visited_urls: List[str] = field(default_factory=list)
searched_query_set: Set[str] = field(default_factory=set)
visited_url_set: Set[str] = field(default_factory=set)
trusted_notes: List[str] = field(default_factory=list)
trace: List[Dict[str, Any]] = field(default_factory=list)
# Accept a variety of placeholder-only answers: a bare ellipsis (ASCII `...`
# or unicode `…`), a single interpunct, and any whitespace-only content. These
# show up when the model echoes a literal `<answer>...</answer>` template
# from the prompt instead of producing a real answer.
_PLACEHOLDER_ANSWER_RE = re.compile(r"^[\s.\u2026\u00b7]*$")
# Pipe-table separator line, e.g. `| --- | :---: |`. The outer pipes are
# optional in some GFM dialects, so we accept both.
_TABLE_SEPARATOR_RE = re.compile(
r"^\s*\|?\s*:?-{2,}:?(?:\s*\|\s*:?-{2,}:?)+\s*\|?\s*$"
)
def strip_think_blocks(text: str) -> str:
"""Remove any <think>...</think> reasoning blocks.
QUEST-35B (Qwen3 family) emits `<think>` reasoning before the final
answer. When the endpoint is deployed without a reasoning parser, the raw
tags leak into chat completion `content`; stripping them here keeps the
extracted answer clean for Markdown rendering.
"""
return re.sub(
r"<think>.*?</think>", "", text, flags=re.DOTALL | re.IGNORECASE
)
def decode_escaped_whitespace(text: str) -> str:
"""Decode literal `\\n`/`\\t`/`\\r` sequences back to real whitespace.
Some OpenAI-compatible servers (and some vLLM builds when a tokenizer's
chat template escapes control characters) return `choices[0].message.content`
with newlines stored as the two-character backslash+n sequence rather than
as a real newline. That breaks Markdown rendering because a pipe table on
a single line is not a table — it is just a sentence with `|` in it, which
is exactly the symptom we saw with:
\\n| Color | Hex |\\n|---|---|\\n| Red | #FF0000 |...
We only decode when the escapes dominate (at least 3 of them, and at
least as many as the real newlines in the text). That keeps us from
corrupting legitimate backslash-n pairs that happen to appear in a code
sample the model produced.
"""
if not text:
return text
escaped_newlines = text.count("\\n")
if escaped_newlines == 0 and "\\t" not in text and "\\r" not in text:
return text
real_newlines = text.count("\n")
if escaped_newlines < max(3, real_newlines + 1):
return text
# Preserve real backslashes so that `\\\\n` (an actual `\n` the model
# wrote) doesn't get collapsed to a newline.
sentinel = "\x00__BS__\x00"
out = text.replace("\\\\", sentinel)
out = out.replace("\\n", "\n").replace("\\r", "\r").replace("\\t", "\t")
out = out.replace(sentinel, "\\")
return out
def _is_placeholder_answer(text: str) -> bool:
return bool(_PLACEHOLDER_ANSWER_RE.match(text or ""))
def ensure_markdown_table_blank_lines(text: str) -> str:
"""Insert a blank line before any pipe-table header row.
GitHub-Flavored Markdown requires a pipe table to be preceded by a
paragraph break; otherwise the header row is folded into the previous
paragraph and the whole table renders as raw text. Models sometimes glue
the table directly under a sentence (e.g. "Here's the comparison: | Col
..."), so we fix that up defensively.
"""
lines = text.split("\n")
out: List[str] = []
for idx, line in enumerate(lines):
is_header = (
"|" in line
and idx + 1 < len(lines)
and _TABLE_SEPARATOR_RE.match(lines[idx + 1]) is not None
)
if is_header and out and out[-1].strip() != "":
out.append("")
out.append(line)
return "\n".join(out)
def extract_answer(text: str) -> Optional[str]:
"""Return the content of the first `<answer>...</answer>` block.
Tries two strategies, in order, and discards placeholder-only content
(bare ellipses) that the model sometimes echoes from the prompt:
1. Well-formed `<answer>...</answer>` block.
2. Truncated `<answer>...` with no closing tag (tokens ran out);
in that case we take everything after the opening tag.
"""
# Decode escaped whitespace on the whole output first so the <answer>
# regex can actually match the opening and closing tags across lines.
decoded = decode_escaped_whitespace(text or "")
cleaned = strip_think_blocks(decoded)
full_match = re.search(
r"<answer>\s*(.*?)\s*</answer>",
cleaned,
flags=re.DOTALL | re.IGNORECASE,
)
if full_match is not None:
candidate = decode_escaped_whitespace(full_match.group(1).strip())
if candidate and not _is_placeholder_answer(candidate):
return candidate
# Closed block was a placeholder / empty: fail fast. Do NOT fall
# through to the open-ended strategy, or it would re-match the same
# tag and incorrectly capture `...</answer>` as the answer.
return None
open_match = re.search(
r"<answer>\s*(.*)$", cleaned, flags=re.DOTALL | re.IGNORECASE
)
if open_match is not None:
candidate = decode_escaped_whitespace(open_match.group(1).strip())
if candidate and not _is_placeholder_answer(candidate):
return candidate
return None
def parse_tool_call(text: str) -> Tuple[Optional[str], Optional[Dict[str, Any]], Optional[str]]:
cleaned = strip_think_blocks(text or "")
match = re.search(r"<tool_call>\s*(.*?)\s*</tool_call>", cleaned, flags=re.DOTALL | re.IGNORECASE)
if not match:
return None, None, None
payload = match.group(1).strip()
try:
data = json.loads(payload)
except json.JSONDecodeError:
return None, None, "Invalid JSON in <tool_call> block."
name = data.get("name")
arguments = data.get("arguments", {})
if not isinstance(name, str) or not isinstance(arguments, dict):
return None, None, "Invalid tool format. Expect name(str) and arguments(dict)."
return name, arguments, None
_SEARCH_UNAVAILABLE_HINT = (
"The web-search backend is currently rate-limited or unreachable. "
"If this question can be answered confidently from your own training "
"knowledge (e.g. common product specs, historical facts, definitions), "
"please produce your best answer now inside <answer>...</answer>, and "
"mention any value that might be out of date. Only ask the user to "
"retry later if the question truly requires a fresh web lookup."
)
# Google Serper API key. Either SERPER_API_KEY or SERPER_KEY_ID is accepted
# so that the Space matches the env-var name used by the research repo.
SERPER_API_KEY = (
os.getenv("SERPER_API_KEY") or os.getenv("SERPER_KEY_ID") or ""
).strip()
SERPER_ENDPOINT = os.getenv("SERPER_ENDPOINT", "https://google.serper.dev/search")
def _serper_search(query: str, max_results: int) -> Dict[str, Any]:
"""Hit the Google Serper API. Returns the same shape as `_ddg_search`.
Serper responds in well under a second and is not subject to the 202
Ratelimit we get from html.duckduckgo.com, so preferring it when the
key is set cuts latency dramatically and eliminates most search
failures on shared Space IPs.
"""
try:
resp = requests.post(
SERPER_ENDPOINT,
headers={
"X-API-KEY": SERPER_API_KEY,
"Content-Type": "application/json",
},
json={"q": query, "num": max_results},
timeout=15,
)
resp.raise_for_status()
data = resp.json()
except Exception as exc:
return {
"ok": False,
"query": query,
"error": f"Serper error: {type(exc).__name__}: {exc}",
"results": [],
"backend": "serper",
}
rows: List[Dict[str, str]] = []
for item in (data.get("organic") or [])[:max_results]:
rows.append(
{
"title": item.get("title", ""),
"href": item.get("link", ""),
"body": item.get("snippet", ""),
}
)
# Fold in the answer box and knowledge graph when present; these often
# carry the exact fact the model is looking for in a compact form.
answer_box = data.get("answerBox") or {}
if answer_box:
rows.insert(
0,
{
"title": answer_box.get("title", "Answer box"),
"href": answer_box.get("link", ""),
"body": answer_box.get("snippet")
or answer_box.get("answer")
or "",
},
)
if not rows:
return {
"ok": False,
"query": query,
"error": "Serper returned no organic results",
"results": [],
"backend": "serper",
}
return {
"ok": True,
"query": query,
"results": rows,
"cached": False,
"backend": "serper",
}
def _ddg_search(query: str, max_results: int) -> Dict[str, Any]:
"""Fallback path: scrape DuckDuckGo. Rate-limits on shared IPs."""
last_exc: Optional[BaseException] = None
for attempt in range(2):
try:
rows: List[Dict[str, str]] = []
with DDGS() as ddgs:
for item in ddgs.text(query, max_results=max_results):
rows.append(
{
"title": item.get("title", ""),
"href": item.get("href", ""),
"body": item.get("body", ""),
}
)
return {
"ok": True,
"query": query,
"results": rows,
"cached": False,
"backend": "duckduckgo",
}
except Exception as exc:
last_exc = exc
if attempt == 0:
time.sleep(1.5)
continue
err = f"{type(last_exc).__name__}: {last_exc}" if last_exc else "unknown error"
return {
"ok": False,
"query": query,
"error": f"DuckDuckGo unavailable ({err}).",
"results": [],
"backend": "duckduckgo",
}
def _run_search_single(query: str, max_results: int) -> Dict[str, Any]:
"""Run one search query, preferring Serper when the key is set.
Returns a structured dict on both success and failure; never raises.
Order of preference:
1. Google Serper (fast, no scraping, requires `SERPER_API_KEY` /
`SERPER_KEY_ID`).
2. DuckDuckGo HTML backend (free, but rate-limits on shared Space IPs).
3. Graceful `ok: False` payload with a hint that tells the agent to
answer from its own knowledge if it reasonably can.
"""
if not query.strip():
return {"ok": False, "error": "Search query cannot be empty."}
cache_key = f"{query.strip().lower()}::{max_results}"
if cache_key in SEARCH_CACHE:
return {**SEARCH_CACHE[cache_key], "cached": True}
tried: List[Dict[str, Any]] = []
if SERPER_API_KEY:
serper_result = _serper_search(query, max_results)
if serper_result.get("ok"):
SEARCH_CACHE[cache_key] = serper_result
return serper_result
tried.append(serper_result)
ddg_result = _ddg_search(query, max_results)
if ddg_result.get("ok"):
SEARCH_CACHE[cache_key] = ddg_result
return ddg_result
tried.append(ddg_result)
# Both backends failed (or no Serper key and DDG rate-limited).
errors = "; ".join(
f"{r.get('backend', 'unknown')}: {r.get('error', 'no results')}"
for r in tried
)
return {
"ok": False,
"query": query,
"error": f"All search backends failed ({errors}).",
"results": [],
"hint": _SEARCH_UNAVAILABLE_HINT,
}
def run_search(query: Union[str, List[str]], max_results: int = 5) -> Dict[str, Any]:
"""Runs one or more queries through DuckDuckGo.
QUEST's schema passes `query` as an array of strings, while the simpler
starter schema used a single string. We accept both shapes.
"""
if isinstance(query, list):
sub_results: List[Dict[str, Any]] = []
for q in query:
if not isinstance(q, str) or not q.strip():
continue
sub_results.append(_run_search_single(q, max_results))
return {"ok": True, "queries": query, "results": sub_results}
return _run_search_single(str(query or "").strip(), max_results)
def _clean_html_to_text(html: str, max_chars: int) -> str:
soup = BeautifulSoup(html, "html.parser")
for tag in soup(["script", "style", "noscript"]):
tag.decompose()
text = soup.get_text(separator=" ", strip=True)
text = re.sub(r"\s+", " ", text)
return text[:max_chars]
def _run_visit_single(url: str, max_chars: int, goal: str = "") -> Dict[str, Any]:
if not url.strip():
return {"ok": False, "error": "URL cannot be empty."}
cache_key = f"{url.strip()}::{max_chars}"
if cache_key in VISIT_CACHE:
return {**VISIT_CACHE[cache_key], "cached": True, "goal": goal}
try:
resp = requests.get(
url,
timeout=20,
headers={"User-Agent": "Mozilla/5.0 (compatible; DeepResearchSpace/1.0)"},
)
resp.raise_for_status()
content_type = resp.headers.get("content-type", "")
if "text/html" in content_type or "<html" in resp.text[:200].lower():
text = _clean_html_to_text(resp.text, max_chars=max_chars)
else:
text = resp.text[:max_chars]
payload = {"ok": True, "url": url, "content": text, "cached": False, "goal": goal}
VISIT_CACHE[cache_key] = payload
return payload
except Exception as exc:
return {"ok": False, "url": url, "error": str(exc), "goal": goal}
def run_visit(
url: Union[str, List[str]],
max_chars: int = 6000,
goal: str = "",
) -> Dict[str, Any]:
"""Fetches one or more URLs. Accepts string or list (QUEST schema)."""
if isinstance(url, list):
sub_results: List[Dict[str, Any]] = []
for u in url:
if not isinstance(u, str) or not u.strip():
continue
sub_results.append(_run_visit_single(u, max_chars, goal))
return {"ok": True, "goal": goal, "results": sub_results}
return _run_visit_single(str(url or "").strip(), max_chars, goal)
def _build_client_for_model(model: str) -> Tuple[InferenceClient, str, List[str]]:
"""Returns (client, primary_model_id, fallback_model_ids).
When the user picks the Quest model and QUEST_BASE_URL is configured, the
InferenceClient is pointed at the dedicated endpoint; otherwise we hit the
shared HF Inference API and let the starter fall back across free models.
"""
token = os.getenv("HF_TOKEN")
quest_timeout = int(os.getenv("QUEST_REQUEST_TIMEOUT", "600"))
if model == QUEST_MODEL_ID and QUEST_BASE_URL:
# Prefer a dedicated key for the self-hosted endpoint so the real HF
# token never travels into vLLM / tunnel logs.
endpoint_token = os.getenv("QUEST_API_KEY") or token
client = InferenceClient(
base_url=QUEST_BASE_URL,
token=endpoint_token,
timeout=quest_timeout,
)
return client, QUEST_ENDPOINT_MODEL, []
client = InferenceClient(token=token, timeout=quest_timeout)
return client, model, []
def call_model(
client: InferenceClient,
messages: List[Dict[str, str]],
preferred_model: str,
candidate_models: List[str],
temperature: float,
max_new_tokens: int,
) -> Tuple[str, str]:
model_order: List[str] = []
for m in [preferred_model] + candidate_models:
if m and m not in model_order:
model_order.append(m)
last_error = None
for model_name in model_order:
try:
completion = client.chat_completion(
model=model_name,
messages=messages,
temperature=temperature,
max_tokens=max_new_tokens,
)
return completion.choices[0].message.content or "", model_name
except Exception as exc:
last_error = exc
continue
raise RuntimeError(f"All model candidates failed. Last error: {last_error}")
def _render_progress(
lines: List[str],
used_model: str,
question: str,
) -> str:
"""Render the in-progress status view that replaces the Markdown panel
while the agent is still running, so the user is not staring at a blank
box for the 20-60 seconds a full QUEST-35B research run can take."""
header = (
f"### ⏳ Researching…\n\n"
f"**Model:** `{used_model}` \n"
f"**Question:** {question.strip()[:200]}"
)
if not lines:
body = "_Starting agent…_"
else:
body = "\n".join(f"- {line}" for line in lines)
return f"{header}\n\n{body}"
def _trace_to_json(state: "AgentState", used_model: str) -> str:
return json.dumps(
{
"used_model": used_model,
"searched_queries": state.searched_queries,
"visited_urls": state.visited_urls,
"trusted_notes": state.trusted_notes[-10:],
"trace": state.trace,
},
ensure_ascii=False,
indent=2,
)
MEMORY_STRATEGIES = ("condenser", "vanilla", "discard_all", "hide_tool_result")
def _normalize_memory_strategy(strategy: str) -> str:
s = (strategy or "condenser").strip().lower().replace("-", "_")
if s == "hide_tool_results":
s = "hide_tool_result"
return s if s in MEMORY_STRATEGIES else "condenser"
def _apply_memory_strategy(messages: List[Dict[str, str]], strategy: str, turn: int) -> None:
"""Lightweight port of the strategies defined in the Quest inference
code (`inference/react_agent.py`). Upstream is token-threshold-driven;
this Space approximates each strategy on a turn-count basis for demo
purposes.
- vanilla: no-op (matches MEMORY_ENABLED=false upstream).
- condenser: no-op here; the main loop injects a compact research-state
summary every few turns (a poor-man's stand-in for the upstream
State Summarizer LLM that emits a structured trusted/untrusted/
uncertain JSON when the token threshold is hit).
- discard_all: every 8 turns, reset history to [system, user question]
(upstream resets when token_count crosses the threshold).
- hide_tool_result: keep only the most recent tool-response user
message; older ones get their content replaced with a stub
(mirrors upstream behavior).
"""
if strategy == "discard_all":
if turn > 1 and turn % 8 == 0 and len(messages) > 2:
system_msg = messages[0]
question_msg = messages[1]
messages.clear()
messages.append(system_msg)
messages.append(question_msg)
messages.append(
{
"role": "user",
"content": "[memory discarded at turn "
f"{turn} — continue the research from the original question]",
}
)
elif strategy == "hide_tool_result":
keep_tail = 1
tool_indices = [
i for i, m in enumerate(messages)
if m.get("role") == "user" and str(m.get("content", "")).startswith("<tool_response>")
]
if len(tool_indices) > keep_tail:
for i in tool_indices[:-keep_tail]:
if messages[i]["content"] != "<tool_response>[hidden]</tool_response>":
messages[i] = {
"role": "user",
"content": "<tool_response>[hidden]</tool_response>",
}
def build_research_agent(
question: str,
model: str,
max_turns: int,
temperature: float,
memory_strategy: str = "condenser",
):
"""Run the ReAct research loop as a generator.
Each `yield` emits a `(markdown_for_answer_panel, json_for_record_panel)`
tuple. Intermediate yields show progress so that Gradio streams the
status lines into the UI as work happens. The last yield contains the
final answer and the final trace.
"""
client, primary_model, fallback_models = _build_client_for_model(model)
# Display label: the real HF repo id is nicer than the TGI shim name.
display_primary = model if (model == QUEST_MODEL_ID) else primary_model
state = AgentState()
used_model = display_primary
status_lines: List[str] = []
def _emit():
"""Yield the current progress snapshot to Gradio."""
return (
_render_progress(status_lines, used_model, question),
_trace_to_json(state, used_model),
)
messages: List[Dict[str, str]] = [
{"role": "system", "content": build_system_prompt()},
{"role": "user", "content": question},
]
final_answer: Optional[str] = None
status_lines.append("🚀 Starting research agent")
yield _emit()
strategy = _normalize_memory_strategy(memory_strategy)
os.environ["MEMORY_STRATEGY"] = strategy
for turn in range(1, max_turns + 1):
_apply_memory_strategy(messages, strategy, turn)
if strategy == "condenser" and state.trusted_notes and turn > 1 and turn % 3 == 0:
summary_lines = "\n".join(f"- {n}" for n in state.trusted_notes[-6:])
messages.append(
{
"role": "user",
"content": f"RESEARCH STATE SUMMARY\n{summary_lines}\nUse this summary to avoid repeating work.",
}
)
status_lines.append(f"🧠 turn {turn}: thinking…")
yield _emit()
t0 = time.time()
raw_output, endpoint_model = call_model(
client=client,
messages=messages,
preferred_model=primary_model,
candidate_models=fallback_models,
temperature=temperature,
max_new_tokens=int(os.getenv("QUEST_MAX_NEW_TOKENS", "4096")),
)
dt = time.time() - t0
model_output = raw_output
# Preserve the human-friendly model id for the trace even if the
# endpoint ignores the "model" param and returns the TGI shim name.
used_model = display_primary if endpoint_model == primary_model == QUEST_ENDPOINT_MODEL else endpoint_model
messages.append({"role": "assistant", "content": model_output})
state.trace.append({"turn": turn, "assistant": model_output, "elapsed_s": round(dt, 2)})
status_lines[-1] = f"🧠 turn {turn}: model reply in {dt:.1f}s"
yield _emit()
extracted_answer = extract_answer(model_output)
if extracted_answer:
final_answer = extracted_answer
status_lines.append("✍️ writing final answer")
yield _emit()
break
tool_name, tool_args, tool_err = parse_tool_call(model_output)
if tool_err:
tool_response = {"ok": False, "error": tool_err}
status_lines.append(f"⚠️ turn {turn}: malformed tool call — {tool_err}")
yield _emit()
elif not tool_name:
# No explicit tool call and no final answer: force finalization.
# IMPORTANT: do not write the literal characters `<answer>...</answer>`
# here. Some models (notably the Qwen3 family that QUEST-35B is
# built on) will echo the template verbatim, which means the
# extracted answer ends up being the three-dot placeholder `...`
# and the user sees an empty-looking result.
messages.append(
{
"role": "user",
"content": (
"You did not call a tool and did not produce a final "
"answer. Please now write your best final answer, "
"wrapped between an opening <answer> tag and a "
"closing </answer> tag. Put the real answer text "
"between those tags; do not write a literal ellipsis "
"or other placeholder. If the question asks for "
"tabular data, use GitHub-Flavored Markdown pipe "
"tables (`| col1 | col2 |` + `|---|---|`) and put a "
"blank line before the first row so the table renders."
),
}
)
status_lines.append(f"🙃 turn {turn}: model stalled; asking for an answer")
yield _emit()
continue
else:
if tool_name == "search":
raw_query = tool_args.get("query", "")
queries: List[str]
if isinstance(raw_query, list):
queries = [str(q).strip() for q in raw_query if str(q).strip()]
else:
queries = [str(raw_query).strip()] if str(raw_query).strip() else []
max_results = int(tool_args.get("max_results", DEFAULT_MAX_SEARCH_RESULTS))
max_results = max(1, min(max_results, DEFAULT_MAX_SEARCH_RESULTS))
queries_preview = ", ".join(f"`{q}`" for q in queries) or "_(empty)_"
status_lines.append(f"🔍 turn {turn}: searching {queries_preview}")
yield _emit()
per_query: List[Dict[str, Any]] = []
backend_labels: List[str] = []
hits_total = 0
for q in queries:
if q in state.searched_query_set:
per_query.append({
"ok": True,
"query": q,
"cached": True,
"note": "Already searched; reusing cached result.",
"results": [],
})
backend_labels.append("cache")
continue
state.searched_queries.append(q)
state.searched_query_set.add(q)
single = _run_search_single(q, max_results)
per_query.append(single)
backend_labels.append(single.get("backend", "unknown"))
if single.get("ok"):
hits_total += len(single.get("results", []))
first_titles = [r.get("title", "") for r in single.get("results", [])[:2]]
if first_titles:
state.trusted_notes.append(
f"Searched '{q}' and found leads: {', '.join(t for t in first_titles if t)}"
)
else:
status_lines.append(
f"⚠️ search failed on `{q}` via {single.get('backend', 'unknown')}: "
f"{single.get('error', 'no results')}"
)
tool_response = (
per_query[0]
if len(per_query) == 1
else {"ok": True, "queries": queries, "results": per_query}
)
unique_backends = sorted(set(backend_labels))
backend_str = "/".join(unique_backends) if unique_backends else "?"
status_lines.append(
f"✅ turn {turn}: got {hits_total} hit(s) via {backend_str}"
)
yield _emit()
elif tool_name == "visit":
raw_url = tool_args.get("url", "")
urls: List[str]
if isinstance(raw_url, list):
urls = [str(u).strip() for u in raw_url if str(u).strip()]
else:
urls = [str(raw_url).strip()] if str(raw_url).strip() else []
goal = str(tool_args.get("goal", "")).strip()
max_chars = int(tool_args.get("max_chars", 6000))
max_chars = max(500, min(max_chars, 20000))
urls_preview = ", ".join(f"`{u[:60]}`" for u in urls) or "_(empty)_"
status_lines.append(f"🌐 turn {turn}: visiting {urls_preview}")
yield _emit()
per_url: List[Dict[str, Any]] = []
visit_ok = 0
for u in urls:
if u in state.visited_url_set:
per_url.append({
"ok": True,
"url": u,
"cached": True,
"note": "Already visited; reusing cached result.",
})
visit_ok += 1
continue
state.visited_urls.append(u)
state.visited_url_set.add(u)
single = _run_visit_single(u, max_chars, goal)
per_url.append(single)
if single.get("ok"):
visit_ok += 1
snippet = str(single.get("content", ""))[:180]
if snippet:
state.trusted_notes.append(
f"Visited {u} and extracted key context: {snippet}"
)
tool_response = (
per_url[0]
if len(per_url) == 1
else {"ok": True, "goal": goal, "results": per_url}
)
status_lines.append(
f"✅ turn {turn}: read {visit_ok}/{len(urls)} page(s)"
)
yield _emit()
else:
tool_response = {"ok": False, "error": f"Unknown tool: {tool_name}"}
status_lines.append(f"⚠️ turn {turn}: unknown tool `{tool_name}`")
yield _emit()
state.trace.append({"turn": turn, "tool": tool_name, "tool_response": tool_response})
messages.append(
{
"role": "user",
"content": TOOL_RESPONSE_TEMPLATE.format(
payload=json.dumps(tool_response, ensure_ascii=False)
),
}
)
if final_answer is None:
final_answer = (
"I could not finish a complete research answer within the configured turns. "
"Try increasing max turns or switching to a stronger model."
)
else:
final_answer = ensure_markdown_table_blank_lines(final_answer)
citations = "\n".join(f"- {url}" for url in sorted(set(state.visited_urls)))
final_answer = f"**Model used:** `{used_model}`\n\n{final_answer}"
if citations:
final_answer = f"{final_answer}\n\n### Visited Sources\n{citations}"
trace_text = _trace_to_json(state, used_model)
yield (final_answer, trace_text)
def run_ui(
question: str,
max_turns: int,
memory_strategy: str,
temperature: float,
):
if not question.strip():
yield "Please input a question.", "{}"
return
if not os.getenv("HF_TOKEN"):
warning = (
"HF_TOKEN is not configured in Space Secrets. "
"Go to Settings -> Secrets -> add `HF_TOKEN`, then retry."
)
yield warning, json.dumps({"error": warning}, ensure_ascii=False, indent=2)
return
if not QUEST_BASE_URL:
warning = (
f"`{QUEST_MODEL_ID}` needs a private HF Inference Endpoint. "
"Create one at https://ui.endpoints.huggingface.co/, then set "
"`QUEST_BASE_URL` in Space Secrets to the endpoint's `/v1/` URL."
)
yield warning, json.dumps({"error": warning}, ensure_ascii=False, indent=2)
return
try:
for partial_answer, partial_trace in build_research_agent(
question=question,
model=QUEST_MODEL_ID,
max_turns=max_turns,
temperature=temperature,
memory_strategy=memory_strategy,
):
yield partial_answer, partial_trace
except Exception as exc:
yield f"Error: {exc}", json.dumps({"error": str(exc)}, ensure_ascii=False, indent=2)
EXAMPLES = [
{
"category": "Multi-hop facts",
"icon": "🎯",
"text": "Who was the first person to walk on the Moon, and which U.S. President set that goal in his famous 1962 “Moon speech”?",
},
{
"category": "Time-varying + multi-hop",
"icon": "📈",
"text": "Who is the current CEO of the company that acquired GitHub in 2018, and what was that company's market capitalization at the close of the most recent quarter?",
},
{
"category": "Multi-constraint",
"icon": "🧩",
"text": "Find a 2-day itinerary in Tokyo under $250 focused on contemporary art museums and vegetarian restaurants, including transit between sites.",
},
{
"category": "Research Report",
"icon": "📚",
"text": "Compare the LLM-safety research approaches of Anthropic, OpenAI, and Google DeepMind over the past 18 months, focusing on alignment techniques and red-teaming methodologies.",
},
]
def _example_label(ex: Dict[str, str]) -> str:
return f"{ex['icon']} {ex['category']} — {ex['text']}"
with gr.Blocks(
title="QUEST · Deep Research by OSU NLP",
theme=APP_THEME,
css=CUSTOM_CSS,
fill_width=True,
) as demo:
# --- Quest-style header (Q mark + title + byline) ---
gr.HTML(
"""
<header class="quest-header">
<div class="quest-header-text">
<h1 class="quest-header-title"><span class="quest-name">QUEST</span>: A Fully Open Recipe for Training Deep Research Agents from Scratch</h1>
<a class="quest-header-byline" href="https://x.com/osunlp" target="_blank" rel="noopener noreferrer">Built by OSU NLP Group</a>
</div>
</header>
"""
)
# --- Main two-column layout ---
with gr.Row(elem_classes="layout-gap"):
with gr.Column(scale=6, min_width=420):
with gr.Group(elem_classes="section-card"):
gr.HTML(
'<div class="section-heading">Ask the agent</div>'
'<div class="hero-heading"><span class="quest-name">QUEST</span>: What I can research for you?</div>'
)
question = gr.Textbox(
show_label=False,
placeholder="Ask anything you want to research in depth...",
lines=6,
)
with gr.Row(elem_classes="action-row"):
run_btn = gr.Button("Run Research", variant="primary", size="lg")
stop_btn = gr.Button("Stop", variant="stop", size="lg")
clear_btn = gr.Button("Clear", variant="secondary", size="lg")
with gr.Group(elem_classes="section-card"):
gr.HTML(
'<div class="section-heading">Try examples</div>'
'<div class="example-note"><span class="quest-name">QUEST</span> can handle multiple types of queries as shown below.</div>'
)
with gr.Column(elem_classes="example-buttons"):
example_buttons = [
gr.Button(_example_label(ex), variant="secondary", elem_classes="example-btn")
for ex in EXAMPLES
]
with gr.Group(elem_classes="section-card"):
gr.HTML('<div class="section-heading">Output</div>')
with gr.Tabs():
with gr.TabItem("Result"):
answer = gr.Markdown(label="Final Answer")
with gr.TabItem("Record"):
trace = gr.Code(label="Execution Trace (JSON)", language="json")
with gr.Column(scale=4, min_width=340, elem_classes="right-stack"):
with gr.Group(elem_classes="section-card"):
gr.HTML(
f"""
<div class="section-heading">Open release</div>
<div class="resource-grid">
<a class="resource-card" href="{PAPER_URL}" target="_blank" rel="noopener noreferrer">
<span class="resource-card-icon" aria-hidden="true">
<svg viewBox="0 0 24 24" role="img" focusable="false"><path d="M6 2.5h8.2L19 7.3v14.2H6V2.5Zm8 1.9v3.2h3.2L14 4.4ZM8.1 9.8h8.8V8.4H8.1v1.4Zm0 3.3h8.8v-1.4H8.1v1.4Zm0 3.3h6.4V15H8.1v1.4Z"/></svg>
</span>
<span class="resource-card-text"><strong>Paper</strong><small>Blog</small></span>
</a>
<a class="resource-card" href="{CODE_URL}" target="_blank" rel="noopener noreferrer">
<span class="resource-card-icon" aria-hidden="true">
<svg viewBox="0 0 24 24" role="img" focusable="false"><path d="M12 1.8c-5.7 0-10.3 4.6-10.3 10.3 0 4.6 3 8.5 7.1 9.8.5.1.7-.2.7-.5v-1.8c-2.9.6-3.5-1.2-3.5-1.2-.5-1.2-1.1-1.5-1.1-1.5-.9-.6.1-.6.1-.6 1 .1 1.6 1.1 1.6 1.1.9 1.6 2.4 1.1 3 .8.1-.7.4-1.1.7-1.3-2.3-.3-4.7-1.2-4.7-5.1 0-1.1.4-2.1 1.1-2.8-.1-.3-.5-1.4.1-2.8 0 0 .9-.3 2.9 1.1.8-.2 1.7-.3 2.6-.3s1.8.1 2.6.3c2-1.4 2.9-1.1 2.9-1.1.6 1.4.2 2.5.1 2.8.7.8 1.1 1.7 1.1 2.8 0 4-2.4 4.8-4.7 5.1.4.3.7 1 .7 2v2.9c0 .3.2.6.7.5 4.1-1.4 7.1-5.2 7.1-9.8C22.3 6.4 17.7 1.8 12 1.8Z"/></svg>
</span>
<span class="resource-card-text"><strong>Code</strong><small>GitHub</small></span>
</a>
<a class="resource-card" href="{DATASET_URL}" target="_blank" rel="noopener noreferrer">
<span class="resource-card-icon resource-card-emoji" aria-hidden="true">🤗</span>
<span class="resource-card-text"><strong>Data</strong><small>Collection</small></span>
</a>
<a class="resource-card" href="{MODEL_URL}" target="_blank" rel="noopener noreferrer">
<span class="resource-card-icon resource-card-emoji" aria-hidden="true">🤗</span>
<span class="resource-card-text"><strong>Model</strong><small>QUEST-35B-RL</small></span>
</a>
</div>
"""
)
with gr.Group(elem_classes="section-card"):
gr.HTML('<div class="section-heading">Settings</div>')
gr.Textbox(
label="Model",
value=QUEST_MODEL_ID,
interactive=False,
elem_id="quest-model",
)
memory_strategy = gr.Radio(
label="Memory Strategy",
choices=[
("Condenser (default)", "condenser"),
("Vanilla", "vanilla"),
("Discard-all", "discard_all"),
("Hide-tool-result", "hide_tool_result"),
],
value="condenser",
elem_id="quest-memory-strategy",
)
gr.HTML(
'<div class="memory-help">'
'<b>Condenser</b> (default) — when context grows large, a State Summarizer LLM compresses earlier turns into a structured JSON of trusted/untrusted/uncertain claims, visited sources, and prior search queries; the agent continues with that compact state.<br>'
'<b>Vanilla</b> — memory management disabled; the full conversation history is kept.<br>'
'<b>Discard-all</b> — when context grows large, the entire message history is reset, restarting the agent from the original question with no accumulated context.<br>'
'<b>Hide-tool-result</b> — when context grows large, older tool responses are pruned; only the most recent tool result is kept.'
'</div>'
)
max_turns = gr.Slider(
label="Max Turns",
minimum=2,
maximum=50,
value=6,
step=1,
elem_id="quest-max-turns",
)
temperature = gr.Slider(
label="Temperature",
minimum=0.0,
maximum=1.5,
value=1.0,
step=0.1,
elem_id="quest-temperature",
)
gr.HTML(
"""
<footer class="quest-footer">
<p>QUEST is a fully open recipe for training deep research agents from scratch — covering data synthesis, memory management, infrastructure, and long-horizon training.</p>
<div class="quest-footer-links">
<a href="https://nlp.osu.edu/" target="_blank" rel="noopener noreferrer">OSU NLP</a>
<a href="https://huggingface.co/osunlp" target="_blank" rel="noopener noreferrer">Hugging Face</a>
</div>
</footer>
"""
)
run_event = run_btn.click(
fn=run_ui,
inputs=[question, max_turns, memory_strategy, temperature],
outputs=[answer, trace],
)
for btn, ex in zip(example_buttons, EXAMPLES):
btn.click(
fn=(lambda text=ex["text"]: text),
inputs=[],
outputs=[question],
)
stop_btn.click(fn=None, cancels=[run_event])
clear_btn.click(
fn=lambda: ("", "", "{}"),
inputs=[],
outputs=[question, answer, trace],
)
if __name__ == "__main__":
demo.launch()
|