{"id":98135,"date":"2026-10-01T21:31:21","date_gmt":"2026-10-01T18:31:21","guid":{"rendered":"https:\/\/twodots.gr\/?p=98135"},"modified":"2026-10-01T21:31:22","modified_gmt":"2026-10-01T18:31:22","slug":"multi-vector-embeddings-domain-finetuning-etairiki-anazitisi","status":"publish","type":"post","link":"https:\/\/twodots.gr\/bg\/multi-vector-embeddings-domain-finetuning-etairiki-anazitisi\/","title":{"rendered":"Multi-vector embeddings: \u03c0\u03ce\u03c2 \u03c4\u03bf domain finetuning \u03b1\u03bb\u03bb\u03ac\u03b6\u03b5\u03b9 \u03c4\u03b7\u03bd \u03b5\u03c4\u03b1\u03b9\u03c1\u03b9\u03ba\u03ae \u03b1\u03bd\u03b1\u03b6\u03ae\u03c4\u03b7\u03c3\u03b7"},"content":{"rendered":"<div class=\"td-article-lede\">\n<p><strong>\u0391\u03c0\u03ac\u03bd\u03c4\u03b7\u03c3\u03b7 \u03c0\u03c1\u03ce\u03c4\u03b1:<\/strong> \u03a4\u03b1 multi-vector embeddings \u03b2\u03b5\u03bb\u03c4\u03b9\u03ce\u03bd\u03bf\u03c5\u03bd \u03c4\u03b7\u03bd \u03b5\u03c4\u03b1\u03b9\u03c1\u03b9\u03ba\u03ae \u03b1\u03bd\u03b1\u03b6\u03ae\u03c4\u03b7\u03c3\u03b7 \u03cc\u03c4\u03b1\u03bd domain data, \u03bc\u03ae\u03ba\u03bf\u03c2 \u03b5\u03b3\u03b3\u03c1\u03ac\u03c6\u03c9\u03bd, evaluation \u03ba\u03b1\u03b9 index \u03bc\u03b5\u03c4\u03c1\u03ce\u03bd\u03c4\u03b1\u03b9 \u03bc\u03b1\u03b6\u03af \u03c3\u03b5 \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03ac queries. \u0394\u03b5\u03bd \u03b1\u03c1\u03ba\u03b5\u03af \u03b7 \u03b1\u03bb\u03bb\u03b1\u03b3\u03ae \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf\u03c5: \u03b7 \u03c0\u03bf\u03b9\u03cc\u03c4\u03b7\u03c4\u03b1 \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03b5\u03bb\u03b5\u03b3\u03c7\u03b8\u03b5\u03af \u03c9\u03c2 \u03b5\u03bd\u03b9\u03b1\u03af\u03bf \u03c3\u03cd\u03c3\u03c4\u03b7\u03bc\u03b1.<\/p>\n<\/div>\n<p>\u0397 \u03ad\u03ba\u03b4\u03bf\u03c3\u03b7 6.0 \u03c4\u03bf\u03c5 Sentence Transformers \u03c0\u03c1\u03cc\u03c3\u03b8\u03b5\u03c3\u03b5 \u03c4\u03bf <code>MultiVectorEncoder<\/code>, \u03bc\u03b1\u03b6\u03af \u03bc\u03b5 \u03bf\u03bb\u03bf\u03ba\u03bb\u03b7\u03c1\u03c9\u03bc\u03ad\u03bd\u03b7 \u03c5\u03c0\u03bf\u03c3\u03c4\u03ae\u03c1\u03b9\u03be\u03b7 \u03b3\u03b9\u03b1 \u03b5\u03ba\u03c0\u03b1\u03af\u03b4\u03b5\u03c5\u03c3\u03b7, \u03b1\u03be\u03b9\u03bf\u03bb\u03cc\u03b3\u03b7\u03c3\u03b7 \u03ba\u03b1\u03b9 \u03c7\u03c1\u03ae\u03c3\u03b7 ColBERT-style late-interaction \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03c9\u03bd. \u0391\u03bd\u03c4\u03af \u03bd\u03b1 \u03c3\u03c5\u03bc\u03c0\u03b9\u03ad\u03b6\u03bf\u03c5\u03bd \u03bf\u03bb\u03cc\u03ba\u03bb\u03b7\u03c1\u03bf query \u03ae document \u03c3\u03b5 \u03ad\u03bd\u03b1 vector, \u03c4\u03b1 \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03b1 \u03b1\u03c5\u03c4\u03ac \u03b4\u03b9\u03b1\u03c4\u03b7\u03c1\u03bf\u03cd\u03bd \u03bc\u03b9\u03ba\u03c1\u03cc\u03c4\u03b5\u03c1\u03b5\u03c2 \u03b1\u03bd\u03b1\u03c0\u03b1\u03c1\u03b1\u03c3\u03c4\u03ac\u03c3\u03b5\u03b9\u03c2 \u03b1\u03bd\u03ac token \u03ba\u03b1\u03b9 \u03c3\u03c5\u03b3\u03ba\u03c1\u03af\u03bd\u03bf\u03c5\u03bd \u03c4\u03b9\u03c2 \u03c0\u03b9\u03bf \u03c3\u03c7\u03b5\u03c4\u03b9\u03ba\u03ad\u03c2 \u03bb\u03b5\u03c0\u03c4\u03bf\u03bc\u03ad\u03c1\u03b5\u03b9\u03b5\u03c2 \u03bc\u03b5 \u03c4\u03bf\u03bd \u03c4\u03b5\u03bb\u03b5\u03c3\u03c4\u03ae MaxSim.<\/p>\n<p>\u039f Tom Aarsen \u03c4\u03b7\u03c2 Hugging Face \u03b4\u03bf\u03ba\u03af\u03bc\u03b1\u03c3\u03b5 \u03c4\u03b7 \u03b4\u03b9\u03b1\u03b4\u03b9\u03ba\u03b1\u03c3\u03af\u03b1 \u03c3\u03b5 \u03b9\u03b1\u03c4\u03c1\u03b9\u03ba\u03ae \u03b1\u03bd\u03b1\u03b6\u03ae\u03c4\u03b7\u03c3\u03b7 \u03bc\u03b5 \u03c4\u03bf MIRIAD. \u03a4\u03bf domain-finetuned \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf \u03c4\u03bf\u03c5 \u03be\u03b5\u03c0\u03ad\u03c1\u03b1\u03c3\u03b5 \u03c0\u03b5\u03c1\u03b9\u03c3\u03c3\u03cc\u03c4\u03b5\u03c1\u03b5\u03c2 \u03b1\u03c0\u03cc 50 \u03b4\u03b9\u03b1\u03bc\u03bf\u03c1\u03c6\u03ce\u03c3\u03b5\u03b9\u03c2 dense, sparse, lexical \u03ba\u03b1\u03b9 multi-vector retrieval \u03c3\u03c4\u03bf \u03c3\u03c5\u03b3\u03ba\u03b5\u03ba\u03c1\u03b9\u03bc\u03ad\u03bd\u03bf evaluation. \u03a4\u03bf \u03b5\u03cd\u03c1\u03b7\u03bc\u03b1 \u03b4\u03b5\u03bd \u03b5\u03af\u03bd\u03b1\u03b9 \u03ba\u03b1\u03b8\u03bf\u03bb\u03b9\u03ba\u03ae \u03ba\u03b1\u03c4\u03ac\u03c4\u03b1\u03be\u03b7 \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03c9\u03bd\u00b7 \u03b5\u03af\u03bd\u03b1\u03b9 \u03b1\u03c0\u03cc\u03b4\u03b5\u03b9\u03be\u03b7 \u03cc\u03c4\u03b9 \u03b7 \u03c3\u03c5\u03bd\u03ac\u03c6\u03b5\u03b9\u03b1 \u03bc\u03b5 \u03c4\u03bf domain, \u03c4\u03bf \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03cc \u03bc\u03ae\u03ba\u03bf\u03c2 \u03c4\u03c9\u03bd \u03b5\u03b3\u03b3\u03c1\u03ac\u03c6\u03c9\u03bd \u03ba\u03b1\u03b9 \u03b7 \u03b1\u03c1\u03c7\u03b9\u03c4\u03b5\u03ba\u03c4\u03bf\u03bd\u03b9\u03ba\u03ae \u03c4\u03bf\u03c5 index \u03bc\u03c0\u03bf\u03c1\u03bf\u03cd\u03bd \u03bd\u03b1 \u03b5\u03c0\u03b7\u03c1\u03b5\u03ac\u03c3\u03bf\u03c5\u03bd \u03c4\u03b7\u03bd \u03c0\u03bf\u03b9\u03cc\u03c4\u03b7\u03c4\u03b1 \u03c0\u03b5\u03c1\u03b9\u03c3\u03c3\u03cc\u03c4\u03b5\u03c1\u03bf \u03b1\u03c0\u03cc \u03c4\u03bf brand \u03ae \u03c4\u03bf \u03c0\u03bb\u03ae\u03b8\u03bf\u03c2 \u03c0\u03b1\u03c1\u03b1\u03bc\u03ad\u03c4\u03c1\u03c9\u03bd.<\/p>\n<p>\u0393\u03b9\u03b1 \u03bc\u03b9\u03b1 \u03b5\u03c0\u03b9\u03c7\u03b5\u03af\u03c1\u03b7\u03c3\u03b7, \u03c4\u03bf \u03b8\u03ad\u03bc\u03b1 \u03b1\u03c6\u03bf\u03c1\u03ac \u03ac\u03bc\u03b5\u03c3\u03b1 RAG, \u03b5\u03c3\u03c9\u03c4\u03b5\u03c1\u03b9\u03ba\u03ac knowledge bases, product search, support search \u03ba\u03b1\u03b9 AI assistants. \u0388\u03bd\u03b1 \u03c3\u03cd\u03c3\u03c4\u03b7\u03bc\u03b1 \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03c0\u03b1\u03c1\u03ac\u03b3\u03b5\u03b9 \u03ba\u03b1\u03bb\u03bf\u03b3\u03c1\u03b1\u03bc\u03bc\u03ad\u03bd\u03b7 \u03b1\u03c0\u03ac\u03bd\u03c4\u03b7\u03c3\u03b7 \u03ba\u03b1\u03b9 \u03c0\u03b1\u03c1\u2019 \u03cc\u03bb\u03b1 \u03b1\u03c5\u03c4\u03ac \u03bd\u03b1 \u03b1\u03c0\u03bf\u03c4\u03c5\u03b3\u03c7\u03ac\u03bd\u03b5\u03b9, \u03b5\u03c0\u03b5\u03b9\u03b4\u03ae \u03b1\u03bd\u03ad\u03ba\u03c4\u03b7\u03c3\u03b5 \u03bb\u03ac\u03b8\u03bf\u03c2 \u03c0\u03bf\u03bb\u03b9\u03c4\u03b9\u03ba\u03ae \u03b5\u03c0\u03b9\u03c3\u03c4\u03c1\u03bf\u03c6\u03ce\u03bd, \u03c0\u03b1\u03bb\u03b9\u03ac \u03ad\u03ba\u03b4\u03bf\u03c3\u03b7 \u03c3\u03c5\u03bc\u03b2\u03bf\u03bb\u03b1\u03af\u03bf\u03c5 \u03ae \u03ac\u03c3\u03c7\u03b5\u03c4\u03b7 \u03c4\u03b5\u03c7\u03bd\u03b9\u03ba\u03ae \u03bf\u03b4\u03b7\u03b3\u03af\u03b1. \u0397 \u03c0\u03bf\u03b9\u03cc\u03c4\u03b7\u03c4\u03b1 \u03b1\u03c1\u03c7\u03af\u03b6\u03b5\u03b9 \u03c0\u03c1\u03b9\u03bd \u03b1\u03c0\u03cc \u03c4\u03bf generation, \u03c3\u03c4\u03bf \u03b1\u03bd \u03bf retriever \u03b2\u03c1\u03af\u03c3\u03ba\u03b5\u03b9 \u03c4\u03bf \u03c3\u03c9\u03c3\u03c4\u03cc \u03b1\u03c0\u03cc\u03c3\u03c0\u03b1\u03c3\u03bc\u03b1.<\/p>\n<aside class=\"td-article-note\"><strong>\u03a0\u03ce\u03c2 \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03b4\u03b9\u03b1\u03b2\u03b1\u03c3\u03c4\u03bf\u03cd\u03bd \u03c4\u03b1 \u03b1\u03c0\u03bf\u03c4\u03b5\u03bb\u03ad\u03c3\u03bc\u03b1\u03c4\u03b1:<\/strong> \u03bf\u03b9 \u03b1\u03c1\u03b9\u03b8\u03bc\u03bf\u03af \u03c0\u03bf\u03c5 \u03b1\u03ba\u03bf\u03bb\u03bf\u03c5\u03b8\u03bf\u03cd\u03bd \u03c0\u03c1\u03bf\u03ad\u03c1\u03c7\u03bf\u03bd\u03c4\u03b1\u03b9 \u03b1\u03c0\u03cc \u03c4\u03bf \u03c0\u03b5\u03af\u03c1\u03b1\u03bc\u03b1 \u03c4\u03b7\u03c2 Hugging Face \u03bc\u03b5 \u03b9\u03b1\u03c4\u03c1\u03b9\u03ba\u03ad\u03c2 \u03b5\u03c1\u03c9\u03c4\u03ae\u03c3\u03b5\u03b9\u03c2 \u03ba\u03b1\u03b9 \u03bc\u03b5\u03b3\u03ac\u03bb\u03b1 passages \u03c4\u03bf\u03c5 MIRIAD. \u03a4\u03b5\u03ba\u03bc\u03b7\u03c1\u03b9\u03ce\u03bd\u03bf\u03c5\u03bd \u03c4\u03b7 \u03c3\u03c5\u03b3\u03ba\u03b5\u03ba\u03c1\u03b9\u03bc\u03ad\u03bd\u03b7 \u03b4\u03b9\u03b1\u03bc\u03cc\u03c1\u03c6\u03c9\u03c3\u03b7 \u03ba\u03b1\u03b9 \u03cc\u03c7\u03b9 \u03b5\u03b3\u03b3\u03c5\u03b7\u03bc\u03ad\u03bd\u03b7 \u03b2\u03b5\u03bb\u03c4\u03af\u03c9\u03c3\u03b7 \u03b3\u03b9\u03b1 \u03ba\u03ac\u03b8\u03b5 e-shop, CRM \u03ae \u03b5\u03c4\u03b1\u03b9\u03c1\u03b9\u03ba\u03cc corpus. \u03a4\u03bf \u03b5\u03c0\u03b9\u03c7\u03b5\u03b9\u03c1\u03b7\u03bc\u03b1\u03c4\u03b9\u03ba\u03cc \u03c3\u03c5\u03bc\u03c0\u03ad\u03c1\u03b1\u03c3\u03bc\u03b1 \u03b5\u03af\u03bd\u03b1\u03b9 \u03b7 \u03bc\u03ad\u03b8\u03bf\u03b4\u03bf\u03c2 \u03bc\u03ad\u03c4\u03c1\u03b7\u03c3\u03b7\u03c2, \u03cc\u03c7\u03b9 \u03b7 \u03bc\u03b7\u03c7\u03b1\u03bd\u03b9\u03ba\u03ae \u03b1\u03bd\u03c4\u03b9\u03b3\u03c1\u03b1\u03c6\u03ae \u03b5\u03bd\u03cc\u03c2 score.<\/aside>\n<div class=\"td-article-toc\">\n<div class=\"td-toc-title\">\u0421\u044a\u0434\u044a\u0440\u0436\u0430\u043d\u0438\u0435<\/div>\n<ul>\n<li><a href=\"#ena-i-polla-vectors\">\u0388\u03bd\u03b1 vector \u03ae \u03c0\u03bf\u03bb\u03bb\u03ac vectors \u03b1\u03bd\u03ac \u03ad\u03b3\u03b3\u03c1\u03b1\u03c6\u03bf;<\/a><\/li>\n<li><a href=\"#domain-finetuning\">\u0393\u03b9\u03b1\u03c4\u03af \u03c4\u03bf domain finetuning \u03b1\u03bb\u03bb\u03ac\u03b6\u03b5\u03b9 \u03c4\u03b7 \u03c3\u03c5\u03bd\u03ac\u03c6\u03b5\u03b9\u03b1<\/a><\/li>\n<li><a href=\"#checkpoint-choice\">\u03a4\u03bf \u03bc\u03b7 \u03c0\u03c1\u03bf\u03c6\u03b1\u03bd\u03ad\u03c2 \u03b1\u03c0\u03bf\u03c4\u03ad\u03bb\u03b5\u03c3\u03bc\u03b1 \u03c3\u03c4\u03b7\u03bd \u03b5\u03c0\u03b9\u03bb\u03bf\u03b3\u03ae checkpoint<\/a><\/li>\n<li><a href=\"#data-loss\">\u0394\u03b5\u03b4\u03bf\u03bc\u03ad\u03bd\u03b1 \u03ba\u03b1\u03b9 loss function \u03c7\u03c9\u03c1\u03af\u03c2 \u03bb\u03ac\u03b8\u03bf\u03c2 \u03b1\u03bd\u03c4\u03b9\u03b3\u03c1\u03b1\u03c6\u03ae<\/a><\/li>\n<li><a href=\"#document-length\">\u03a4\u03bf \u03bc\u03ae\u03ba\u03bf\u03c2 \u03b5\u03b3\u03b3\u03c1\u03ac\u03c6\u03c9\u03bd \u03b5\u03af\u03bd\u03b1\u03b9 \u03bc\u03ad\u03c1\u03bf\u03c2 \u03c4\u03bf\u03c5 \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf\u03c5<\/a><\/li>\n<li><a href=\"#evaluation\">\u0388\u03bd\u03b1 \u03b5\u03cd\u03ba\u03bf\u03bb\u03bf evaluation \u03ba\u03c1\u03cd\u03b2\u03b5\u03b9 \u03c4\u03b9\u03c2 \u03b4\u03b9\u03b1\u03c6\u03bf\u03c1\u03ad\u03c2<\/a><\/li>\n<li><a href=\"#results\">\u03a4\u03b9 \u03ad\u03b4\u03b5\u03b9\u03be\u03b1\u03bd \u03bf\u03b9 50+ retrieval \u03b4\u03b9\u03b1\u03bc\u03bf\u03c1\u03c6\u03ce\u03c3\u03b5\u03b9\u03c2<\/a><\/li>\n<li><a href=\"#index-cost\">\u039f index \u03ba\u03b1\u03b8\u03bf\u03c1\u03af\u03b6\u03b5\u03b9 \u03c4\u03bf \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03cc \u03ba\u03cc\u03c3\u03c4\u03bf\u03c2<\/a><\/li>\n<li><a href=\"#decision-framework\">\u03a0\u03cc\u03c4\u03b5 \u03ad\u03c7\u03b5\u03b9 \u03bd\u03cc\u03b7\u03bc\u03b1 \u03b3\u03b9\u03b1 \u03b5\u03c4\u03b1\u03b9\u03c1\u03b9\u03ba\u03ae \u03b1\u03bd\u03b1\u03b6\u03ae\u03c4\u03b7\u03c3\u03b7<\/a><\/li>\n<li><a href=\"#pilot-steps\">\u0395\u03c0\u03c4\u03ac \u03b2\u03ae\u03bc\u03b1\u03c4\u03b1 \u03b3\u03b9\u03b1 \u03b1\u03c3\u03c6\u03b1\u03bb\u03ad\u03c2 pilot<\/a><\/li>\n<li><a href=\"#symperasma\">\u0397 \u03c3\u03c5\u03bd\u03ac\u03c6\u03b5\u03b9\u03b1 \u03b5\u03af\u03bd\u03b1\u03b9 \u03c0\u03c1\u03bf\u03ca\u03cc\u03bd, \u03cc\u03c7\u03b9 \u03bc\u03cc\u03bd\u03bf \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf<\/a><\/li>\n<\/ul>\n<\/div>\n<h2 id=\"ena-i-polla-vectors\">\u0388\u03bd\u03b1 vector \u03ae \u03c0\u03bf\u03bb\u03bb\u03ac vectors \u03b1\u03bd\u03ac \u03ad\u03b3\u03b3\u03c1\u03b1\u03c6\u03bf;<\/h2>\n<p>\u0388\u03bd\u03b1 dense embedding model \u03b5\u03c0\u03b9\u03c3\u03c4\u03c1\u03ad\u03c6\u03b5\u03b9 \u03c3\u03c5\u03bd\u03ae\u03b8\u03c9\u03c2 \u03bc\u03af\u03b1 \u03c3\u03c5\u03bc\u03c0\u03c5\u03ba\u03bd\u03c9\u03bc\u03ad\u03bd\u03b7 \u03b1\u03bd\u03b1\u03c0\u03b1\u03c1\u03ac\u03c3\u03c4\u03b1\u03c3\u03b7 \u03b3\u03b9\u03b1 \u03bf\u03bb\u03cc\u03ba\u03bb\u03b7\u03c1\u03bf \u03c4\u03bf \u03ba\u03b5\u03af\u03bc\u03b5\u03bd\u03bf. Query \u03ba\u03b1\u03b9 document \u03bc\u03c0\u03bf\u03c1\u03bf\u03cd\u03bd \u03ad\u03c4\u03c3\u03b9 \u03bd\u03b1 \u03c3\u03c5\u03b3\u03ba\u03c1\u03b9\u03b8\u03bf\u03cd\u03bd \u03c0\u03bf\u03bb\u03cd \u03b3\u03c1\u03ae\u03b3\u03bf\u03c1\u03b1 \u03bc\u03b5 \u03ad\u03bd\u03b1 dot product, \u03b5\u03bd\u03ce \u03bf index \u03c0\u03b1\u03c1\u03b1\u03bc\u03ad\u03bd\u03b5\u03b9 \u03c3\u03c7\u03b5\u03c4\u03b9\u03ba\u03ac \u03bc\u03b9\u03ba\u03c1\u03cc\u03c2. \u0397 \u03c3\u03c5\u03bc\u03c0\u03af\u03b5\u03c3\u03b7 \u03cc\u03bc\u03c9\u03c2 \u03b5\u03af\u03bd\u03b1\u03b9 \u03b1\u03c0\u03c9\u03bb\u03b5\u03c3\u03c4\u03b9\u03ba\u03ae: \u03bc\u03b9\u03b1 \u03c3\u03c0\u03ac\u03bd\u03b9\u03b1 \u03bf\u03bd\u03bf\u03bc\u03b1\u03c3\u03af\u03b1, \u03ad\u03bd\u03b1\u03c2 \u03ba\u03c9\u03b4\u03b9\u03ba\u03cc\u03c2 \u03c0\u03c1\u03bf\u03ca\u03cc\u03bd\u03c4\u03bf\u03c2 \u03ae \u03bc\u03af\u03b1 \u03ba\u03c1\u03af\u03c3\u03b9\u03bc\u03b7 \u03c1\u03ae\u03c4\u03c1\u03b1 \u03bc\u03ad\u03c3\u03b1 \u03c3\u03b5 \u03bc\u03b5\u03b3\u03ac\u03bb\u03bf \u03ad\u03b3\u03b3\u03c1\u03b1\u03c6\u03bf \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03bc\u03bf\u03b9\u03c1\u03b1\u03c3\u03c4\u03bf\u03cd\u03bd \u03c4\u03bf\u03bd \u03af\u03b4\u03b9\u03bf \u03c7\u03ce\u03c1\u03bf \u03bc\u03b5 \u03cc\u03bb\u03bf \u03c4\u03bf \u03c5\u03c0\u03cc\u03bb\u03bf\u03b9\u03c0\u03bf \u03c0\u03b5\u03c1\u03b9\u03b5\u03c7\u03cc\u03bc\u03b5\u03bd\u03bf.<\/p>\n<p>\u03a4\u03bf multi-vector retrieval \u03b4\u03b9\u03b1\u03c4\u03b7\u03c1\u03b5\u03af \u03ad\u03bd\u03b1 \u03bc\u03b9\u03ba\u03c1\u03cc vector \u03b1\u03bd\u03ac token. \u039a\u03b1\u03c4\u03ac \u03c4\u03bf scoring, \u03ba\u03ac\u03b8\u03b5 query token \u03b2\u03c1\u03af\u03c3\u03ba\u03b5\u03b9 \u03c4\u03bf \u03ba\u03b1\u03bb\u03cd\u03c4\u03b5\u03c1\u03bf \u03c4\u03b1\u03af\u03c1\u03b9\u03b1\u03c3\u03bc\u03b1 \u03bc\u03ad\u03c3\u03b1 \u03c3\u03c4\u03b1 document tokens \u03ba\u03b1\u03b9 \u03bf\u03b9 \u03b5\u03c0\u03b9\u03bc\u03ad\u03c1\u03bf\u03c5\u03c2 \u03b2\u03b1\u03b8\u03bc\u03bf\u03bb\u03bf\u03b3\u03af\u03b5\u03c2 \u03b1\u03b8\u03c1\u03bf\u03af\u03b6\u03bf\u03bd\u03c4\u03b1\u03b9 \u03bc\u03ad\u03c3\u03c9 MaxSim. \u0397 \u03b1\u03bb\u03bb\u03b7\u03bb\u03b5\u03c0\u03af\u03b4\u03c1\u03b1\u03c3\u03b7 \u03b3\u03af\u03bd\u03b5\u03c4\u03b1\u03b9 \u03b1\u03c1\u03b3\u03ac \u2014 \u03b3\u03b9\u2019 \u03b1\u03c5\u03c4\u03cc \u03bf\u03bd\u03bf\u03bc\u03ac\u03b6\u03b5\u03c4\u03b1\u03b9 late interaction \u2014 \u03b1\u03bb\u03bb\u03ac \u03c4\u03b1 documents \u03b5\u03be\u03b1\u03ba\u03bf\u03bb\u03bf\u03c5\u03b8\u03bf\u03cd\u03bd \u03bd\u03b1 \u03ba\u03c9\u03b4\u03b9\u03ba\u03bf\u03c0\u03bf\u03b9\u03bf\u03cd\u03bd\u03c4\u03b1\u03b9 \u03b5\u03ba \u03c4\u03c9\u03bd \u03c0\u03c1\u03bf\u03c4\u03ad\u03c1\u03c9\u03bd \u03ba\u03b1\u03b9 \u03bd\u03b1 \u03b1\u03c0\u03bf\u03b8\u03b7\u03ba\u03b5\u03cd\u03bf\u03bd\u03c4\u03b1\u03b9 \u03c3\u03b5 index.<\/p>\n<p>\u0397 \u03b1\u03c1\u03c7\u03b9\u03c4\u03b5\u03ba\u03c4\u03bf\u03bd\u03b9\u03ba\u03ae \u03b2\u03c1\u03af\u03c3\u03ba\u03b5\u03c4\u03b1\u03b9 \u03b1\u03bd\u03ac\u03bc\u03b5\u03c3\u03b1 \u03c3\u03c4\u03bf\u03bd \u03b3\u03c1\u03ae\u03b3\u03bf\u03c1\u03bf bi-encoder \u03ba\u03b1\u03b9 \u03c3\u03c4\u03bf\u03bd \u03b1\u03ba\u03c1\u03b9\u03b2\u03cc\u03c4\u03b5\u03c1\u03bf cross-encoder. \u0394\u03b9\u03b1\u03c4\u03b7\u03c1\u03b5\u03af \u03c0\u03b5\u03c1\u03b9\u03c3\u03c3\u03cc\u03c4\u03b5\u03c1\u03b7 token-level \u03c0\u03bb\u03b7\u03c1\u03bf\u03c6\u03bf\u03c1\u03af\u03b1 \u03c7\u03c9\u03c1\u03af\u03c2 \u03bd\u03b1 \u03b1\u03c0\u03b1\u03b9\u03c4\u03b5\u03af \u03bd\u03b1 \u03c0\u03b5\u03c1\u03ac\u03c3\u03b5\u03b9 \u03ba\u03ac\u03b8\u03b5 \u03c3\u03c5\u03bd\u03b4\u03c5\u03b1\u03c3\u03bc\u03cc\u03c2 query\u2013document \u03b1\u03c0\u03cc \u03c4\u03bf \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf \u03c3\u03b5 \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03cc \u03c7\u03c1\u03cc\u03bd\u03bf. \u03a4\u03bf \u03c4\u03af\u03bc\u03b7\u03bc\u03b1 \u03b5\u03af\u03bd\u03b1\u03b9 \u03c0\u03b5\u03c1\u03b9\u03c3\u03c3\u03cc\u03c4\u03b5\u03c1\u03b1 vectors, \u03bc\u03b5\u03b3\u03b1\u03bb\u03cd\u03c4\u03b5\u03c1\u03bf index \u03ba\u03b1\u03b9 \u03c0\u03b9\u03bf \u03c3\u03cd\u03bd\u03b8\u03b5\u03c4\u03bf serving.<\/p>\n<p>\u0391\u03c5\u03c4\u03cc \u03b5\u03be\u03b7\u03b3\u03b5\u03af \u03b3\u03b9\u03b1\u03c4\u03af \u03b7 <a href=\"https:\/\/twodots.gr\/hybrid-ai-search-papers-with-code\/\">hybrid AI search<\/a> \u03c0\u03b1\u03c1\u03b1\u03bc\u03ad\u03bd\u03b5\u03b9 \u03c7\u03c1\u03ae\u03c3\u03b9\u03bc\u03b7: BM25, dense embeddings, multi-vector retrieval \u03ba\u03b1\u03b9 reranking \u03bb\u03cd\u03bd\u03bf\u03c5\u03bd \u03b4\u03b9\u03b1\u03c6\u03bf\u03c1\u03b5\u03c4\u03b9\u03ba\u03ac \u03c0\u03c1\u03bf\u03b2\u03bb\u03ae\u03bc\u03b1\u03c4\u03b1. \u0397 \u03c3\u03c9\u03c3\u03c4\u03ae \u03c3\u03c4\u03bf\u03af\u03b2\u03b1 \u03b4\u03b5\u03bd \u03c0\u03c1\u03bf\u03ba\u03cd\u03c0\u03c4\u03b5\u03b9 \u03b1\u03c0\u03cc \u03b9\u03b4\u03b5\u03bf\u03bb\u03bf\u03b3\u03b9\u03ba\u03ae \u03b5\u03c0\u03b9\u03bb\u03bf\u03b3\u03ae \u03b5\u03bd\u03cc\u03c2 \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf\u03c5, \u03b1\u03bb\u03bb\u03ac \u03b1\u03c0\u03cc \u03c4\u03b1 queries, \u03c4\u03b1 documents \u03ba\u03b1\u03b9 \u03c4\u03bf latency budget \u03c4\u03bf\u03c5 \u03c0\u03c1\u03bf\u03ca\u03cc\u03bd\u03c4\u03bf\u03c2.<\/p>\n<div class=\"td-comparison td-comparison-cards td-comparison-cards--horizontal\">\n<div class=\"td-comparison-grid td-comparison-grid-three\">\n<div class=\"td-platform-card\">\n<p class=\"td-comparison-title\">Dense retrieval<\/p>\n<p>\u0388\u03bd\u03b1 vector \u03b1\u03bd\u03ac \u03ba\u03b5\u03af\u03bc\u03b5\u03bd\u03bf, \u03bc\u03b9\u03ba\u03c1\u03cc\u03c4\u03b5\u03c1\u03bf\u03c2 index \u03ba\u03b1\u03b9 \u03b3\u03c1\u03ae\u03b3\u03bf\u03c1\u03bf \u03c0\u03c1\u03ce\u03c4\u03bf \u03c3\u03c4\u03ac\u03b4\u03b9\u03bf. \u03a4\u03b1\u03b9\u03c1\u03b9\u03ac\u03b6\u03b5\u03b9 \u03cc\u03c4\u03b1\u03bd \u03c4\u03b1 documents \u03b5\u03af\u03bd\u03b1\u03b9 \u03c3\u03cd\u03bd\u03c4\u03bf\u03bc\u03b1 \u03ae \u03b7 \u03b3\u03b5\u03bd\u03b9\u03ba\u03ae \u03c3\u03b7\u03bc\u03b1\u03c3\u03b9\u03bf\u03bb\u03bf\u03b3\u03b9\u03ba\u03ae \u03bf\u03bc\u03bf\u03b9\u03cc\u03c4\u03b7\u03c4\u03b1 \u03b1\u03c1\u03ba\u03b5\u03af.<\/p>\n<div class=\"td-badge-row\"><span class=\"td-badge\">1 vector<\/span><span class=\"td-badge\">\u03a7\u03b1\u03bc\u03b7\u03bb\u03cc\u03c4\u03b5\u03c1\u03bf storage<\/span><\/div>\n<\/div>\n<div class=\"td-platform-card td-platform-card--navy\">\n<p class=\"td-comparison-title\">Multi-vector retrieval<\/p>\n<p>\u0388\u03bd\u03b1 vector \u03b1\u03bd\u03ac token \u03ba\u03b1\u03b9 MaxSim \u03b3\u03b9\u03b1 \u03bb\u03b5\u03c0\u03c4\u03bf\u03bc\u03b5\u03c1\u03ae \u03b1\u03bd\u03c4\u03b9\u03c3\u03c4\u03bf\u03af\u03c7\u03b9\u03c3\u03b7. \u0388\u03c7\u03b5\u03b9 \u03b1\u03be\u03af\u03b1 \u03cc\u03c4\u03b1\u03bd \u03b7 \u03c3\u03c9\u03c3\u03c4\u03ae \u03b1\u03c0\u03ac\u03bd\u03c4\u03b7\u03c3\u03b7 \u03b5\u03be\u03b1\u03c1\u03c4\u03ac\u03c4\u03b1\u03b9 \u03b1\u03c0\u03cc \u03bc\u03b9\u03ba\u03c1\u03ae \u03b1\u03bb\u03bb\u03ac \u03ba\u03c1\u03af\u03c3\u03b9\u03bc\u03b7 \u03b5\u03bd\u03cc\u03c4\u03b7\u03c4\u03b1 \u03bc\u03b5\u03b3\u03ac\u03bb\u03bf\u03c5 \u03b5\u03b3\u03b3\u03c1\u03ac\u03c6\u03bf\u03c5.<\/p>\n<div class=\"td-badge-row\"><span class=\"td-badge\">Token-level<\/span><span class=\"td-badge\">\u039c\u03b5\u03b3\u03b1\u03bb\u03cd\u03c4\u03b5\u03c1\u03bf\u03c2 index<\/span><\/div>\n<\/div>\n<div class=\"td-platform-card\">\n<p class=\"td-comparison-title\">Cross-encoder reranking<\/p>\n<p>Query \u03ba\u03b1\u03b9 candidate document \u03b1\u03be\u03b9\u03bf\u03bb\u03bf\u03b3\u03bf\u03cd\u03bd\u03c4\u03b1\u03b9 \u03bc\u03b1\u03b6\u03af, \u03bc\u03b5 \u03c5\u03c8\u03b7\u03bb\u03cc \u03c5\u03c0\u03bf\u03bb\u03bf\u03b3\u03b9\u03c3\u03c4\u03b9\u03ba\u03cc \u03ba\u03cc\u03c3\u03c4\u03bf\u03c2. \u0395\u03af\u03bd\u03b1\u03b9 \u03c0\u03c1\u03b1\u03ba\u03c4\u03b9\u03ba\u03cc \u03c9\u03c2 \u03b4\u03b5\u03cd\u03c4\u03b5\u03c1\u03bf \u03c3\u03c4\u03ac\u03b4\u03b9\u03bf \u03c0\u03ac\u03bd\u03c9 \u03c3\u03b5 \u03bc\u03b9\u03ba\u03c1\u03cc shortlist.<\/p>\n<div class=\"td-badge-row\"><span class=\"td-badge\">Joint scoring<\/span><span class=\"td-badge\">2\u03bf \u03c3\u03c4\u03ac\u03b4\u03b9\u03bf<\/span><\/div>\n<\/div>\n<\/div>\n<\/div>\n<h2 id=\"domain-finetuning\">\u0393\u03b9\u03b1\u03c4\u03af \u03c4\u03bf domain finetuning \u03b1\u03bb\u03bb\u03ac\u03b6\u03b5\u03b9 \u03c4\u03b7 \u03c3\u03c5\u03bd\u03ac\u03c6\u03b5\u03b9\u03b1<\/h2>\n<p>\u0397 \u03bb\u03ad\u03be\u03b7 \u00ab\u03c3\u03c7\u03b5\u03c4\u03b9\u03ba\u03cc\u00bb \u03b4\u03b5\u03bd \u03c3\u03b7\u03bc\u03b1\u03af\u03bd\u03b5\u03b9 \u03c4\u03bf \u03af\u03b4\u03b9\u03bf \u03c3\u03b5 \u03bd\u03bf\u03bc\u03b9\u03ba\u03ae \u03ad\u03c1\u03b5\u03c5\u03bd\u03b1, \u03c4\u03b5\u03c7\u03bd\u03b9\u03ba\u03ae \u03c5\u03c0\u03bf\u03c3\u03c4\u03ae\u03c1\u03b9\u03be\u03b7, product discovery \u03ae \u03b9\u03b1\u03c4\u03c1\u03b9\u03ba\u03ae \u03b2\u03b9\u03b2\u03bb\u03b9\u03bf\u03b3\u03c1\u03b1\u03c6\u03af\u03b1. \u0391\u03bb\u03bb\u03ac\u03b6\u03b5\u03b9 \u03c4\u03bf \u03bb\u03b5\u03be\u03b9\u03bb\u03cc\u03b3\u03b9\u03bf, \u03c4\u03bf \u03cd\u03c6\u03bf\u03c2 \u03c4\u03c9\u03bd queries \u03ba\u03b1\u03b9, \u03ba\u03c5\u03c1\u03af\u03c9\u03c2, \u03c4\u03bf \u03c0\u03bf\u03b9\u03bf \u03b1\u03c0\u03bf\u03c4\u03ad\u03bb\u03b5\u03c3\u03bc\u03b1 \u03b8\u03b5\u03c9\u03c1\u03b5\u03af \u03bf \u03c7\u03c1\u03ae\u03c3\u03c4\u03b7\u03c2 \u03c3\u03c9\u03c3\u03c4\u03cc. \u0388\u03bd\u03b1\u03c2 \u03b3\u03b5\u03bd\u03b9\u03ba\u03cc\u03c2 retriever \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03ba\u03b1\u03c4\u03b1\u03bb\u03b1\u03b2\u03b1\u03af\u03bd\u03b5\u03b9 \u03c4\u03bf \u03b8\u03ad\u03bc\u03b1, \u03b1\u03bb\u03bb\u03ac \u03bd\u03b1 \u03bc\u03b7\u03bd \u03be\u03b5\u03c7\u03c9\u03c1\u03af\u03b6\u03b5\u03b9 \u03c4\u03b7\u03bd \u03b9\u03c3\u03c7\u03cd\u03bf\u03c5\u03c3\u03b1 \u03b4\u03b9\u03b1\u03b4\u03b9\u03ba\u03b1\u03c3\u03af\u03b1 \u03b1\u03c0\u03cc \u03bc\u03b9\u03b1 \u03c0\u03b1\u03bb\u03b9\u03ac \u03ad\u03ba\u03b4\u03bf\u03c3\u03b7 \u03ae \u03ad\u03bd\u03b1\u03bd \u03b1\u03ba\u03c1\u03b9\u03b2\u03ae \u03ba\u03c9\u03b4\u03b9\u03ba\u03cc \u03b1\u03c0\u03cc \u03ad\u03bd\u03b1 \u03c0\u03b1\u03c1\u03cc\u03bc\u03bf\u03b9\u03bf \u03c0\u03c1\u03bf\u03ca\u03cc\u03bd.<\/p>\n<p>\u03a4\u03bf finetuning \u03c7\u03c1\u03b7\u03c3\u03b9\u03bc\u03bf\u03c0\u03bf\u03b9\u03b5\u03af in-domain \u03c0\u03b1\u03c1\u03b1\u03b4\u03b5\u03af\u03b3\u03bc\u03b1\u03c4\u03b1 query\u2013relevant document \u03b3\u03b9\u03b1 \u03bd\u03b1 \u03bc\u03ac\u03b8\u03b5\u03b9 \u03b1\u03c5\u03c4\u03ae \u03c4\u03b7 \u03bb\u03b5\u03b9\u03c4\u03bf\u03c5\u03c1\u03b3\u03b9\u03ba\u03ae \u03ad\u03bd\u03bd\u03bf\u03b9\u03b1 \u03c4\u03b7\u03c2 \u03c3\u03c5\u03bd\u03ac\u03c6\u03b5\u03b9\u03b1\u03c2. \u03a3\u03c4\u03bf multi-vector \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf, \u03b7 \u03c0\u03c1\u03bf\u03c3\u03b1\u03c1\u03bc\u03bf\u03b3\u03ae \u03b5\u03c0\u03b7\u03c1\u03b5\u03ac\u03b6\u03b5\u03b9 \u03c4\u03b7 \u03bb\u03b5\u03c0\u03c4\u03bf\u03bc\u03b5\u03c1\u03ae token-level \u03b1\u03bd\u03c4\u03b9\u03c3\u03c4\u03bf\u03af\u03c7\u03b9\u03c3\u03b7. \u0394\u03b5\u03bd \u03b1\u03bd\u03c4\u03b9\u03ba\u03b1\u03b8\u03b9\u03c3\u03c4\u03ac \u03cc\u03bc\u03c9\u03c2 \u03c4\u03b7\u03bd \u03c0\u03bf\u03b9\u03cc\u03c4\u03b7\u03c4\u03b1 \u03c4\u03c9\u03bd \u03b4\u03b5\u03b4\u03bf\u03bc\u03ad\u03bd\u03c9\u03bd: \u03b1\u03bd \u03c4\u03b1 \u03b8\u03b5\u03c4\u03b9\u03ba\u03ac pairs \u03b5\u03af\u03bd\u03b1\u03b9 \u03b5\u03cd\u03ba\u03bf\u03bb\u03b1, \u03bb\u03b1\u03bd\u03b8\u03b1\u03c3\u03bc\u03ad\u03bd\u03b1 \u03ae \u03b4\u03b9\u03b1\u03c1\u03c1\u03ad\u03bf\u03c5\u03bd \u03b1\u03c0\u03cc \u03c4\u03bf train \u03c3\u03c4\u03bf test, \u03c4\u03bf \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf \u03bc\u03b1\u03b8\u03b1\u03af\u03bd\u03b5\u03b9 \u03ad\u03bd\u03b1 \u03c4\u03b5\u03c7\u03bd\u03b7\u03c4\u03cc shortcut.<\/p>\n<p>\u0397 \u03c0\u03c1\u03b1\u03ba\u03c4\u03b9\u03ba\u03ae \u03b1\u03be\u03af\u03b1 \u03c6\u03b1\u03af\u03bd\u03b5\u03c4\u03b1\u03b9 \u03c3\u03b5 \u03ad\u03bd\u03b1 <a href=\"https:\/\/twodots.gr\/private-chatbot-rag-knowledge-base-business\/\">private chatbot \u03bc\u03b5 RAG<\/a>. \u0395\u03c1\u03c9\u03c4\u03ae\u03c3\u03b5\u03b9\u03c2 \u03c7\u03c1\u03b7\u03c3\u03c4\u03ce\u03bd, tickets \u03c0\u03bf\u03c5 \u03bb\u03cd\u03b8\u03b7\u03ba\u03b1\u03bd \u03c3\u03c9\u03c3\u03c4\u03ac \u03ba\u03b1\u03b9 \u03b5\u03b3\u03ba\u03b5\u03ba\u03c1\u03b9\u03bc\u03ad\u03bd\u03b1 documents \u03bc\u03c0\u03bf\u03c1\u03bf\u03cd\u03bd \u03bd\u03b1 \u03b3\u03af\u03bd\u03bf\u03c5\u03bd training \u03ae evaluation pairs. \u03a0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03cc\u03bc\u03c9\u03c2 \u03bd\u03b1 \u03b1\u03c6\u03b1\u03b9\u03c1\u03b5\u03b8\u03bf\u03cd\u03bd \u03c0\u03c1\u03bf\u03c3\u03c9\u03c0\u03b9\u03ba\u03ac \u03b4\u03b5\u03b4\u03bf\u03bc\u03ad\u03bd\u03b1, \u03bd\u03b1 \u03b4\u03b9\u03b1\u03c7\u03c9\u03c1\u03b9\u03c3\u03c4\u03bf\u03cd\u03bd \u03bf\u03b9 \u03c7\u03c1\u03bf\u03bd\u03b9\u03ba\u03ad\u03c2 \u03b5\u03ba\u03b4\u03cc\u03c3\u03b5\u03b9\u03c2 \u03ba\u03b1\u03b9 \u03bd\u03b1 \u03bf\u03c1\u03b9\u03c3\u03c4\u03b5\u03af \u03c0\u03bf\u03b9\u03b1 \u03b1\u03c0\u03ac\u03bd\u03c4\u03b7\u03c3\u03b7 \u03b5\u03af\u03bd\u03b1\u03b9 \u03ad\u03b3\u03ba\u03c5\u03c1\u03b7 \u03c3\u03ae\u03bc\u03b5\u03c1\u03b1.<\/p>\n<p>\u03a4\u03bf domain finetuning \u03b4\u03b5\u03bd \u03b4\u03b9\u03bf\u03c1\u03b8\u03ce\u03bd\u03b5\u03b9 \u03b1\u03c0\u03cc \u03bc\u03cc\u03bd\u03bf \u03c4\u03bf\u03c5 \u03ad\u03bd\u03b1 \u03c7\u03b1\u03bb\u03b1\u03c3\u03bc\u03ad\u03bd\u03bf corpus. Duplicate documents, \u03b1\u03bd\u03c4\u03b9\u03ba\u03c1\u03bf\u03c5\u03cc\u03bc\u03b5\u03bd\u03b5\u03c2 \u03c0\u03bf\u03bb\u03b9\u03c4\u03b9\u03ba\u03ad\u03c2 \u03ba\u03b1\u03b9 \u03b1\u03c3\u03b1\u03c6\u03ad\u03c2 ownership \u03b4\u03b7\u03bc\u03b9\u03bf\u03c5\u03c1\u03b3\u03bf\u03cd\u03bd \u03bb\u03ac\u03b8\u03bf\u03c2 labels \u03c0\u03c1\u03b9\u03bd \u03be\u03b5\u03ba\u03b9\u03bd\u03ae\u03c3\u03b5\u03b9 \u03b7 \u03b5\u03ba\u03c0\u03b1\u03af\u03b4\u03b5\u03c5\u03c3\u03b7. \u0397 \u03b4\u03bf\u03c5\u03bb\u03b5\u03b9\u03ac \u03b1\u03c1\u03c7\u03af\u03b6\u03b5\u03b9 \u03b1\u03c0\u03cc data governance: \u03c0\u03bf\u03b9\u03b5\u03c2 \u03c0\u03b7\u03b3\u03ad\u03c2 \u03b5\u03af\u03bd\u03b1\u03b9 authoritative, \u03c0\u03ce\u03c2 \u03b5\u03bd\u03b7\u03bc\u03b5\u03c1\u03ce\u03bd\u03bf\u03bd\u03c4\u03b1\u03b9 \u03ba\u03b1\u03b9 \u03c0\u03bf\u03b9\u03bf\u03c2 \u03b5\u03b3\u03ba\u03c1\u03af\u03bd\u03b5\u03b9 \u03c4\u03bf relevance.<\/p>\n<h2 id=\"checkpoint-choice\">\u03a4\u03bf \u03bc\u03b7 \u03c0\u03c1\u03bf\u03c6\u03b1\u03bd\u03ad\u03c2 \u03b1\u03c0\u03bf\u03c4\u03ad\u03bb\u03b5\u03c3\u03bc\u03b1 \u03c3\u03c4\u03b7\u03bd \u03b5\u03c0\u03b9\u03bb\u03bf\u03b3\u03ae checkpoint<\/h2>\n<p>\u03a4\u03bf Sentence Transformers 6.0 \u03b5\u03c0\u03b9\u03c4\u03c1\u03ad\u03c0\u03b5\u03b9 \u03b4\u03cd\u03bf \u03b2\u03b1\u03c3\u03b9\u03ba\u03ad\u03c2 \u03b1\u03c6\u03b5\u03c4\u03b7\u03c1\u03af\u03b5\u03c2. \u039c\u03b9\u03b1 \u03bf\u03bc\u03ac\u03b4\u03b1 \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03c3\u03c5\u03bd\u03b5\u03c7\u03af\u03c3\u03b5\u03b9 \u03c4\u03bf finetuning \u03c5\u03c0\u03ac\u03c1\u03c7\u03bf\u03bd\u03c4\u03bf\u03c2 multi-vector checkpoint \u03ae \u03bd\u03b1 \u03c0\u03c1\u03bf\u03c3\u03b8\u03ad\u03c3\u03b5\u03b9 \u03bd\u03ad\u03b1 multi-vector \u03ba\u03b5\u03c6\u03b1\u03bb\u03ae \u03c0\u03ac\u03bd\u03c9 \u03c3\u03b5 base transformer. \u03a3\u03c4\u03b7 \u03b4\u03b5\u03cd\u03c4\u03b5\u03c1\u03b7 \u03c0\u03b5\u03c1\u03af\u03c0\u03c4\u03c9\u03c3\u03b7 \u03b4\u03b7\u03bc\u03b9\u03bf\u03c5\u03c1\u03b3\u03b5\u03af\u03c4\u03b1\u03b9, \u03bc\u03b5\u03c4\u03b1\u03be\u03cd \u03ac\u03bb\u03bb\u03c9\u03bd, \u03c4\u03c5\u03c7\u03b1\u03af\u03b1 \u03b1\u03c1\u03c7\u03b9\u03ba\u03bf\u03c0\u03bf\u03b9\u03b7\u03bc\u03ad\u03bd\u03bf projection \u03c0\u03c1\u03bf\u03c2 \u03c4\u03b7\u03bd \u03ba\u03bb\u03b1\u03c3\u03b9\u03ba\u03ae \u03b4\u03b9\u03ac\u03c3\u03c4\u03b1\u03c3\u03b7 \u03c4\u03c9\u03bd 128, \u03bc\u03b1\u03b6\u03af \u03bc\u03b5 masking \u03ba\u03b1\u03b9 normalization. \u03a4\u03bf \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf \u03c7\u03c1\u03b5\u03b9\u03ac\u03b6\u03b5\u03c4\u03b1\u03b9 training \u03c0\u03c1\u03b9\u03bd \u03b3\u03af\u03bd\u03b5\u03b9 \u03c7\u03c1\u03ae\u03c3\u03b9\u03bc\u03bf \u03b3\u03b9\u03b1 retrieval.<\/p>\n<p>\u03a3\u03c4\u03bf \u03c0\u03b5\u03af\u03c1\u03b1\u03bc\u03b1 \u03c4\u03b7\u03c2 Hugging Face, \u03ad\u03be\u03b9 \u03b1\u03c6\u03b5\u03c4\u03b7\u03c1\u03af\u03b5\u03c2 \u03b5\u03ba\u03c0\u03b1\u03b9\u03b4\u03b5\u03cd\u03c4\u03b7\u03ba\u03b1\u03bd \u03bc\u03b5 \u03c4\u03b7\u03bd \u03af\u03b4\u03b9\u03b1 \u03c3\u03c5\u03bd\u03c4\u03b1\u03b3\u03ae \u03c0\u03ac\u03bd\u03c9 \u03c3\u03b5 25.000 \u03b9\u03b1\u03c4\u03c1\u03b9\u03ba\u03ac question\u2013passage pairs \u03ba\u03b1\u03b9 \u03b1\u03be\u03b9\u03bf\u03bb\u03bf\u03b3\u03ae\u03b8\u03b7\u03ba\u03b1\u03bd \u03c3\u03b5 1.000 held-out queries \u03bc\u03b5 corpus 50.000 passages. \u03a4\u03bf <code>mLateOn-unsupervised<\/code> \u03b1\u03bd\u03ad\u03b2\u03b7\u03ba\u03b5 \u03b1\u03c0\u03cc 0,9087 \u03c3\u03b5 0,9398 NDCG@10. \u03a4\u03bf finished <code>mLateOn<\/code> \u03c0\u03ae\u03b3\u03b5 \u03b1\u03c0\u03cc 0,9277 \u03c3\u03b5 0,9319, \u03b5\u03bd\u03ce \u03ac\u03bb\u03bb\u03b1 finished checkpoints \u03c5\u03c0\u03bf\u03c7\u03ce\u03c1\u03b7\u03c3\u03b1\u03bd.<\/p>\n<p>\u0397 \u03b5\u03c1\u03bc\u03b7\u03bd\u03b5\u03af\u03b1 \u03c4\u03bf\u03c5 \u03c3\u03c5\u03b3\u03b3\u03c1\u03b1\u03c6\u03ad\u03b1 \u03b5\u03af\u03bd\u03b1\u03b9 \u03cc\u03c4\u03b9 \u03c4\u03b1 pre-supervised checkpoints \u03b5\u03af\u03c7\u03b1\u03bd \u03ae\u03b4\u03b7 late-interaction structure, \u03c7\u03c9\u03c1\u03af\u03c2 \u03bd\u03b1 \u03ad\u03c7\u03bf\u03c5\u03bd \u03c0\u03c1\u03bf\u03c3\u03b1\u03c1\u03bc\u03bf\u03c3\u03c4\u03b5\u03af \u03c4\u03cc\u03c3\u03bf \u03ad\u03bd\u03c4\u03bf\u03bd\u03b1 \u03c3\u03b5 general-purpose supervised relevance. \u038c\u03c4\u03b1\u03bd \u03c4\u03ad\u03c4\u03bf\u03b9\u03bf checkpoint \u03b4\u03b5\u03bd \u03c5\u03c0\u03ac\u03c1\u03c7\u03b5\u03b9, fresh projection \u03c0\u03ac\u03bd\u03c9 \u03c3\u03b5 retrieval-pretrained backbone \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03b5\u03af\u03bd\u03b1\u03b9 \u03bb\u03bf\u03b3\u03b9\u03ba\u03ae \u03b5\u03bd\u03b1\u03bb\u03bb\u03b1\u03ba\u03c4\u03b9\u03ba\u03ae: \u03c4\u03bf <code>gte-modernbert-base<\/code> \u03ad\u03c6\u03c4\u03b1\u03c3\u03b5 0,9177 \u03bc\u03b5\u03c4\u03ac \u03c4\u03b1 \u03af\u03b4\u03b9\u03b1 25.000 pairs.<\/p>\n<p>\u0391\u03c5\u03c4\u03cc \u03b5\u03af\u03bd\u03b1\u03b9 \u03c5\u03c0\u03cc\u03b8\u03b5\u03c3\u03b7 \u03c0\u03c1\u03bf\u03c2 \u03ad\u03bb\u03b5\u03b3\u03c7\u03bf, \u03cc\u03c7\u03b9 \u03bc\u03cc\u03bd\u03b9\u03bc\u03bf\u03c2 \u03ba\u03b1\u03bd\u03cc\u03bd\u03b1\u03c2. \u0397 \u03bf\u03bc\u03ac\u03b4\u03b1 \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03c3\u03c5\u03b3\u03ba\u03c1\u03af\u03bd\u03b5\u03b9 \u03c4\u03bf\u03c5\u03bb\u03ac\u03c7\u03b9\u03c3\u03c4\u03bf\u03bd \u03ad\u03bd\u03b1 pre-supervised checkpoint, \u03ad\u03bd\u03b1 finished checkpoint \u03ba\u03b1\u03b9 \u03ad\u03bd\u03b1 \u03b9\u03c3\u03c7\u03c5\u03c1\u03cc dense baseline \u03bc\u03b5 \u03b1\u03ba\u03c1\u03b9\u03b2\u03ce\u03c2 \u03c4\u03b1 \u03af\u03b4\u03b9\u03b1 splits \u03ba\u03b1\u03b9 metrics. \u0397 \u03b5\u03c0\u03b9\u03bb\u03bf\u03b3\u03ae \u03c0\u03bf\u03c5 \u03c6\u03b1\u03af\u03bd\u03b5\u03c4\u03b1\u03b9 \u03c0\u03b9\u03bf \u00ab\u03ce\u03c1\u03b9\u03bc\u03b7\u00bb \u03b1\u03c0\u03cc \u03c4\u03bf \u03cc\u03bd\u03bf\u03bc\u03b1 \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03c0\u03c1\u03bf\u03c3\u03b1\u03c1\u03bc\u03cc\u03b6\u03b5\u03c4\u03b1\u03b9 \u03c7\u03b5\u03b9\u03c1\u03cc\u03c4\u03b5\u03c1\u03b1 \u03c3\u03c4\u03bf \u03b4\u03b9\u03ba\u03cc \u03c3\u03b1\u03c2 domain.<\/p>\n<h2 id=\"data-loss\">\u0394\u03b5\u03b4\u03bf\u03bc\u03ad\u03bd\u03b1 \u03ba\u03b1\u03b9 loss function \u03c7\u03c9\u03c1\u03af\u03c2 \u03bb\u03ac\u03b8\u03bf\u03c2 \u03b1\u03bd\u03c4\u03b9\u03b3\u03c1\u03b1\u03c6\u03ae<\/h2>\n<p>\u039f <code>MultiVectorEncoderTrainer<\/code> \u03b4\u03ad\u03c7\u03b5\u03c4\u03b1\u03b9 datasets \u03b1\u03c0\u03cc \u03c4\u03bf Hugging Face Hub \u03ae \u03c4\u03bf\u03c0\u03b9\u03ba\u03ac CSV, JSON, Parquet, Arrow \u03ba\u03b1\u03b9 SQL. \u0397 \u03b4\u03bf\u03ba\u03b9\u03bc\u03ae \u03c7\u03c1\u03b7\u03c3\u03b9\u03bc\u03bf\u03c0\u03bf\u03af\u03b7\u03c3\u03b5 \u03c4\u03bf MIRIAD, \u03bc\u03b5 \u03c0\u03b5\u03c1\u03af\u03c0\u03bf\u03c5 4,4 \u03b5\u03ba\u03b1\u03c4\u03bf\u03bc\u03bc\u03cd\u03c1\u03b9\u03b1 \u03b9\u03b1\u03c4\u03c1\u03b9\u03ba\u03ad\u03c2 \u03b5\u03c1\u03c9\u03c4\u03ae\u03c3\u03b5\u03b9\u03c2 \u03c3\u03c5\u03bd\u03b4\u03b5\u03b4\u03b5\u03bc\u03ad\u03bd\u03b5\u03c2 \u03bc\u03b5 \u03c4\u03bf passage \u03c0\u03bf\u03c5 \u03c0\u03b5\u03c1\u03b9\u03ad\u03c7\u03b5\u03b9 \u03c4\u03b7\u03bd \u03b1\u03c0\u03ac\u03bd\u03c4\u03b7\u03c3\u03b7. \u0393\u03b9\u03b1 \u03c4\u03bf \u03c0\u03bb\u03ae\u03c1\u03b5\u03c2 training \u03b5\u03c0\u03b9\u03bb\u03ad\u03c7\u03b8\u03b7\u03ba\u03b1\u03bd \u03ad\u03bd\u03b1 \u03b5\u03ba\u03b1\u03c4\u03bf\u03bc\u03bc\u03cd\u03c1\u03b9\u03bf pairs.<\/p>\n<p>\u0397 \u03b1\u03c0\u03bb\u03ae \u03bc\u03bf\u03c1\u03c6\u03ae query\u2013relevant document \u03b5\u03af\u03bd\u03b1\u03b9 \u03c3\u03c5\u03c7\u03bd\u03ac \u03b1\u03c1\u03ba\u03b5\u03c4\u03ae \u03b3\u03b9\u03b1 \u03ad\u03bd\u03b1 \u03c0\u03c1\u03ce\u03c4\u03bf \u03b5\u03c4\u03b1\u03b9\u03c1\u03b9\u03ba\u03cc pilot. \u0394\u03b5\u03bd \u03b1\u03c0\u03b1\u03b9\u03c4\u03b5\u03af \u03c5\u03c0\u03bf\u03c7\u03c1\u03b5\u03c9\u03c4\u03b9\u03ba\u03ac teacher scores \u03ae mined negatives. \u0397 \u03c3\u03b5\u03b9\u03c1\u03ac \u03c4\u03c9\u03bd columns \u03cc\u03bc\u03c9\u03c2 \u03ad\u03c7\u03b5\u03b9 \u03c3\u03b7\u03bc\u03b1\u03c3\u03af\u03b1: \u03b7 \u03c0\u03c1\u03ce\u03c4\u03b7 \u03b1\u03bd\u03c4\u03b9\u03bc\u03b5\u03c4\u03c9\u03c0\u03af\u03b6\u03b5\u03c4\u03b1\u03b9 \u03c9\u03c2 query \u03ba\u03b1\u03b9 \u03bf\u03b9 \u03b5\u03c0\u03cc\u03bc\u03b5\u03bd\u03b5\u03c2 \u03c9\u03c2 documents, \u03b5\u03ba\u03c4\u03cc\u03c2 \u03b1\u03bd \u03b4\u03bf\u03b8\u03b5\u03af \u03b4\u03b9\u03b1\u03c6\u03bf\u03c1\u03b5\u03c4\u03b9\u03ba\u03cc router mapping. \u038c\u03c4\u03b1\u03bd \u03b7 loss \u03b1\u03c0\u03b1\u03b9\u03c4\u03b5\u03af label, \u03b7 \u03c3\u03c4\u03ae\u03bb\u03b7 \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03bf\u03bd\u03bf\u03bc\u03ac\u03b6\u03b5\u03c4\u03b1\u03b9 <code>label<\/code> \u03ae <code>score<\/code>.<\/p>\n<p>\u0393\u03b9\u03b1 question\u2013passage pairs, \u03b7 <code>MultiVectorMultipleNegativesRankingLoss<\/code> \u03c7\u03c1\u03b7\u03c3\u03b9\u03bc\u03bf\u03c0\u03bf\u03b9\u03b5\u03af \u03c4\u03b1 \u03c5\u03c0\u03cc\u03bb\u03bf\u03b9\u03c0\u03b1 documents \u03c4\u03bf\u03c5 batch \u03c9\u03c2 negatives \u03b3\u03b9\u03b1 \u03ba\u03ac\u03b8\u03b5 query. \u039c\u03b5\u03b3\u03b1\u03bb\u03cd\u03c4\u03b5\u03c1\u03bf effective batch \u03b4\u03af\u03bd\u03b5\u03b9 \u03c0\u03b5\u03c1\u03b9\u03c3\u03c3\u03cc\u03c4\u03b5\u03c1\u03b1 negatives, \u03b1\u03bb\u03bb\u03ac \u03b7 \u03ba\u03c9\u03b4\u03b9\u03ba\u03bf\u03c0\u03bf\u03af\u03b7\u03c3\u03b7 \u03bc\u03b5\u03b3\u03ac\u03bb\u03c9\u03bd documents \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03b5\u03be\u03b1\u03bd\u03c4\u03bb\u03ae\u03c3\u03b5\u03b9 \u03c4\u03b7 GPU. \u0397 cached \u03b5\u03ba\u03b4\u03bf\u03c7\u03ae \u03bc\u03b5 GradCache \u03c7\u03c9\u03c1\u03af\u03b6\u03b5\u03b9 \u03c4\u03b7\u03bd \u03ba\u03c9\u03b4\u03b9\u03ba\u03bf\u03c0\u03bf\u03af\u03b7\u03c3\u03b7 \u03c3\u03b5 mini-batches, \u03b5\u03bd\u03ce \u03c4\u03bf <code>mini_batch_num_tokens<\/code> \u03b5\u03c0\u03b9\u03c4\u03c1\u03ad\u03c0\u03b5\u03b9 token budget \u03cc\u03c4\u03b1\u03bd \u03c4\u03b1 \u03bc\u03ae\u03ba\u03b7 \u03b4\u03b9\u03b1\u03c6\u03ad\u03c1\u03bf\u03c5\u03bd \u03ad\u03bd\u03c4\u03bf\u03bd\u03b1.<\/p>\n<p>\u03a5\u03c0\u03ac\u03c1\u03c7\u03b5\u03b9 \u03ba\u03b1\u03b9 \u03bc\u03b9\u03b1 \u03c3\u03c5\u03b3\u03ba\u03b5\u03ba\u03c1\u03b9\u03bc\u03ad\u03bd\u03b7 \u03c0\u03b1\u03b3\u03af\u03b4\u03b1 \u03bc\u03b5\u03c4\u03b1\u03c6\u03bf\u03c1\u03ac\u03c2 \u03c1\u03c5\u03b8\u03bc\u03af\u03c3\u03b5\u03c9\u03bd. \u03a4\u03bf multi-vector contrastive loss \u03ad\u03c7\u03b5\u03b9 default <code>scale=1.0<\/code>, \u03cc\u03c7\u03b9 20.0. \u0397 MaxSim \u03b2\u03b1\u03b8\u03bc\u03bf\u03bb\u03bf\u03b3\u03af\u03b1 \u03b1\u03b8\u03c1\u03bf\u03af\u03b6\u03b5\u03b9 \u03c0\u03bf\u03bb\u03bb\u03ac token matches \u03ba\u03b1\u03b9 \u03ae\u03b4\u03b7 \u03ad\u03c7\u03b5\u03b9 \u03bc\u03b5\u03b3\u03b1\u03bb\u03cd\u03c4\u03b5\u03c1\u03bf \u03b5\u03cd\u03c1\u03bf\u03c2\u00b7 \u03b7 \u03c4\u03c5\u03c6\u03bb\u03ae \u03b1\u03bd\u03c4\u03b9\u03b3\u03c1\u03b1\u03c6\u03ae \u03c4\u03bf\u03c5 20.0 \u03b1\u03c0\u03cc dense cosine training \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03ba\u03bf\u03c1\u03ad\u03c3\u03b5\u03b9 \u03c4\u03bf softmax \u03ba\u03b1\u03b9 \u03bd\u03b1 \u03ba\u03b1\u03c4\u03b1\u03c3\u03c4\u03c1\u03ad\u03c8\u03b5\u03b9 \u03c4\u03b1 gradients.<\/p>\n<h2 id=\"document-length\">\u03a4\u03bf \u03bc\u03ae\u03ba\u03bf\u03c2 \u03b5\u03b3\u03b3\u03c1\u03ac\u03c6\u03c9\u03bd \u03b5\u03af\u03bd\u03b1\u03b9 \u03bc\u03ad\u03c1\u03bf\u03c2 \u03c4\u03bf\u03c5 \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf\u03c5<\/h2>\n<p>\u03a0\u03bf\u03bb\u03bb\u03ac retrieval checkpoints \u03ad\u03c7\u03bf\u03c5\u03bd \u03b5\u03ba\u03c0\u03b1\u03b9\u03b4\u03b5\u03c5\u03c4\u03b5\u03af \u03bc\u03b5 document caps 180, 256, 300 \u03ae 512 tokens. \u03a3\u03c4\u03bf MIRIAD, \u03c4\u03b1 passages \u03b5\u03af\u03c7\u03b1\u03bd \u03bc\u03ad\u03c3\u03bf \u03bc\u03ae\u03ba\u03bf\u03c2 \u03c0\u03b5\u03c1\u03af\u03c0\u03bf\u03c5 941 tokens. \u0397 Hugging Face \u03bc\u03ad\u03c4\u03c1\u03b7\u03c3\u03b5 \u03b1\u03c0\u03ce\u03bb\u03b5\u03b9\u03b1 \u03b1\u03c0\u03cc 0,08 \u03ad\u03c9\u03c2 0,24 NDCG@10 \u03c3\u03b5 multi-vector \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03b1 \u03cc\u03c4\u03b1\u03bd \u03b5\u03c6\u03b1\u03c1\u03bc\u03cc\u03b6\u03bf\u03bd\u03c4\u03b1\u03bd length caps. \u03a3\u03c4\u03b7 \u03c3\u03c5\u03b3\u03ba\u03b5\u03ba\u03c1\u03b9\u03bc\u03ad\u03bd\u03b7 \u03b4\u03bf\u03ba\u03b9\u03bc\u03ae, \u03c4\u03bf truncation \u03bc\u03c0\u03bf\u03c1\u03bf\u03cd\u03c3\u03b5 \u03bd\u03b1 \u03b5\u03c0\u03b7\u03c1\u03b5\u03ac\u03c3\u03b5\u03b9 \u03c0\u03b5\u03c1\u03b9\u03c3\u03c3\u03cc\u03c4\u03b5\u03c1\u03bf \u03c4\u03bf \u03b1\u03c0\u03bf\u03c4\u03ad\u03bb\u03b5\u03c3\u03bc\u03b1 \u03b1\u03c0\u03cc \u03c4\u03b7 \u03b4\u03b9\u03b1\u03c6\u03bf\u03c1\u03ac \u03b1\u03bd\u03ac\u03bc\u03b5\u03c3\u03b1 \u03c3\u03b5 \u03b1\u03c1\u03c7\u03b9\u03c4\u03b5\u03ba\u03c4\u03bf\u03bd\u03b9\u03ba\u03ad\u03c2.<\/p>\n<p>\u03a3\u03c4\u03b7\u03bd \u03b5\u03c4\u03b1\u03b9\u03c1\u03b9\u03ba\u03ae \u03b1\u03bd\u03b1\u03b6\u03ae\u03c4\u03b7\u03c3\u03b7, \u03b1\u03c5\u03c4\u03cc \u03b5\u03af\u03bd\u03b1\u03b9 \u03b5\u03cd\u03ba\u03bf\u03bb\u03bf \u03bd\u03b1 \u03c0\u03b5\u03c1\u03ac\u03c3\u03b5\u03b9 \u03b1\u03c0\u03b1\u03c1\u03b1\u03c4\u03ae\u03c1\u03b7\u03c4\u03bf. \u0397 \u03ba\u03c1\u03af\u03c3\u03b9\u03bc\u03b7 \u03b5\u03be\u03b1\u03af\u03c1\u03b5\u03c3\u03b7 \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03b2\u03c1\u03af\u03c3\u03ba\u03b5\u03c4\u03b1\u03b9 \u03c3\u03c4\u03bf \u03c4\u03ad\u03bb\u03bf\u03c2 \u03bc\u03b9\u03b1\u03c2 \u03c0\u03bf\u03bb\u03b9\u03c4\u03b9\u03ba\u03ae\u03c2 \u03b5\u03c0\u03b9\u03c3\u03c4\u03c1\u03bf\u03c6\u03ce\u03bd, \u03c4\u03bf SLA \u03c3\u03c4\u03b7\u03bd \u03c4\u03b5\u03bb\u03b5\u03c5\u03c4\u03b1\u03af\u03b1 \u03c3\u03b5\u03bb\u03af\u03b4\u03b1 \u03bc\u03b9\u03b1\u03c2 \u03c3\u03cd\u03bc\u03b2\u03b1\u03c3\u03b7\u03c2 \u03ae \u03c4\u03bf \u03b2\u03ae\u03bc\u03b1 \u03b1\u03c0\u03bf\u03ba\u03b1\u03c4\u03ac\u03c3\u03c4\u03b1\u03c3\u03b7\u03c2 \u03b2\u03b1\u03b8\u03b9\u03ac \u03bc\u03ad\u03c3\u03b1 \u03c3\u03b5 \u03c4\u03b5\u03c7\u03bd\u03b9\u03ba\u03cc manual. \u0391\u03bd \u03c4\u03bf retrieval layer \u03b4\u03b5\u03bd \u03b2\u03bb\u03ad\u03c0\u03b5\u03b9 \u03c4\u03bf \u03b1\u03c0\u03cc\u03c3\u03c0\u03b1\u03c3\u03bc\u03b1, \u03ba\u03b1\u03bd\u03ad\u03bd\u03b1 \u03ba\u03b1\u03bb\u03cd\u03c4\u03b5\u03c1\u03bf prompt \u03c3\u03c4\u03bf generation layer \u03b4\u03b5\u03bd \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03c4\u03bf \u03b1\u03bd\u03b1\u03ba\u03c4\u03ae\u03c3\u03b5\u03b9.<\/p>\n<p>\u03a4\u03bf \u03c0\u03bb\u03ae\u03c1\u03b5\u03c2 run \u03c4\u03b7\u03c2 \u03bc\u03b5\u03bb\u03ad\u03c4\u03b7\u03c2 \u03ac\u03c6\u03b7\u03c3\u03b5 \u03c3\u03ba\u03cc\u03c0\u03b9\u03bc\u03b1 \u03ba\u03b5\u03bd\u03cc \u03c4\u03bf training <code>max_length<\/code>, \u03ce\u03c3\u03c4\u03b5 \u03c4\u03bf \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf \u03bd\u03b1 \u03b2\u03bb\u03ad\u03c0\u03b5\u03b9 \u03c4\u03bf \u03af\u03b4\u03b9\u03bf \u03bc\u03ae\u03ba\u03bf\u03c2 \u03c0\u03bf\u03c5 \u03b8\u03b1 \u03b5\u03be\u03c5\u03c0\u03b7\u03c1\u03b5\u03c4\u03bf\u03cd\u03c3\u03b5 \u03c3\u03c4\u03bf inference. Training \u03c3\u03c4\u03b1 512 tokens \u03ae\u03c4\u03b1\u03bd \u03c0\u03b5\u03c1\u03af\u03c0\u03bf\u03c5 \u03b4\u03cd\u03bf \u03c6\u03bf\u03c1\u03ad\u03c2 \u03b3\u03c1\u03b7\u03b3\u03bf\u03c1\u03cc\u03c4\u03b5\u03c1\u03bf, \u03b1\u03bb\u03bb\u03ac \u03ad\u03c7\u03b1\u03c3\u03b5 \u03c0\u03b5\u03c1\u03af\u03c0\u03bf\u03c5 0,015 NDCG@10 \u03ba\u03b1\u03b9 \u03b7 \u03b4\u03b9\u03b1\u03c6\u03bf\u03c1\u03ac \u03b4\u03b5\u03bd \u03ad\u03ba\u03bb\u03b5\u03b9\u03c3\u03b5 \u03bc\u03b5 \u03c0\u03b5\u03c1\u03b9\u03c3\u03c3\u03cc\u03c4\u03b5\u03c1\u03b1 \u03b4\u03b5\u03b4\u03bf\u03bc\u03ad\u03bd\u03b1. \u03a4\u03bf trade-off \u03b1\u03c6\u03bf\u03c1\u03ac \u03b1\u03c5\u03c4\u03cc \u03c4\u03bf corpus, \u03cc\u03bc\u03c9\u03c2 \u03b7 \u03b1\u03c1\u03c7\u03ae \u03b5\u03af\u03bd\u03b1\u03b9 \u03b3\u03b5\u03bd\u03b9\u03ba\u03ae: training \u03ba\u03b1\u03b9 serving length \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03c3\u03c7\u03b5\u03b4\u03b9\u03ac\u03b6\u03bf\u03bd\u03c4\u03b1\u03b9 \u03bc\u03b1\u03b6\u03af.<\/p>\n<p>\u03a0\u03c1\u03b9\u03bd \u03b5\u03c0\u03b9\u03bb\u03b5\u03b3\u03b5\u03af \u03bd\u03ad\u03bf \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf, \u03b1\u03be\u03af\u03b6\u03b5\u03b9 \u03bd\u03b1 \u03bc\u03b5\u03c4\u03c1\u03b7\u03b8\u03b5\u03af \u03b7 \u03ba\u03b1\u03c4\u03b1\u03bd\u03bf\u03bc\u03ae \u03bc\u03ae\u03ba\u03bf\u03c5\u03c2, \u03c4\u03bf \u03c0\u03bf\u03c3\u03bf\u03c3\u03c4\u03cc truncation \u03ba\u03b1\u03b9 \u03b7 \u03b8\u03ad\u03c3\u03b7 \u03c4\u03c9\u03bd \u03c3\u03c7\u03b5\u03c4\u03b9\u03ba\u03ce\u03bd \u03b1\u03c0\u03bf\u03c3\u03c0\u03b1\u03c3\u03bc\u03ac\u03c4\u03c9\u03bd. \u0397 \u03c0\u03c1\u03bf\u03c3\u03ad\u03b3\u03b3\u03b9\u03c3\u03b7 \u03c4\u03c9\u03bd <a href=\"https:\/\/twodots.gr\/semantic-compression-trees-rag-context\/\">Semantic Compression Trees<\/a> \u03b4\u03b5\u03af\u03c7\u03bd\u03b5\u03b9 \u03bc\u03b9\u03b1 \u03b4\u03b9\u03b1\u03c6\u03bf\u03c1\u03b5\u03c4\u03b9\u03ba\u03ae \u03c3\u03c4\u03c1\u03b1\u03c4\u03b7\u03b3\u03b9\u03ba\u03ae \u03bc\u03b5\u03af\u03c9\u03c3\u03b7\u03c2 context, \u03b1\u03bb\u03bb\u03ac \u03ba\u03b1\u03b9 \u03b5\u03ba\u03b5\u03af \u03b7 \u03c3\u03c5\u03bc\u03c0\u03af\u03b5\u03c3\u03b7 \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03b1\u03be\u03b9\u03bf\u03bb\u03bf\u03b3\u03b5\u03af\u03c4\u03b1\u03b9 \u03c9\u03c2 \u03c0\u03b9\u03b8\u03b1\u03bd\u03ae \u03b1\u03c0\u03ce\u03bb\u03b5\u03b9\u03b1 \u03c0\u03bb\u03b7\u03c1\u03bf\u03c6\u03bf\u03c1\u03af\u03b1\u03c2, \u03cc\u03c7\u03b9 \u03c9\u03c2 \u03b4\u03c9\u03c1\u03b5\u03ac\u03bd \u03b2\u03b5\u03bb\u03c4\u03b9\u03c3\u03c4\u03bf\u03c0\u03bf\u03af\u03b7\u03c3\u03b7.<\/p>\n<h2 id=\"evaluation\">\u0388\u03bd\u03b1 \u03b5\u03cd\u03ba\u03bf\u03bb\u03bf evaluation \u03ba\u03c1\u03cd\u03b2\u03b5\u03b9 \u03c4\u03b9\u03c2 \u03b4\u03b9\u03b1\u03c6\u03bf\u03c1\u03ad\u03c2<\/h2>\n<p>\u0388\u03bd\u03b1 \u03c7\u03b1\u03bc\u03b7\u03bb\u03cc evaluation loss \u03b4\u03b5\u03bd \u03b1\u03c0\u03b1\u03bd\u03c4\u03ac \u03b1\u03bd \u03bf \u03c7\u03c1\u03ae\u03c3\u03c4\u03b7\u03c2 \u03b8\u03b1 \u03b2\u03c1\u03b5\u03b9 \u03c4\u03bf \u03c3\u03c9\u03c3\u03c4\u03cc document. \u03a4\u03bf Sentence Transformers \u03c0\u03b1\u03c1\u03ad\u03c7\u03b5\u03b9 evaluators \u03b3\u03b9\u03b1 information retrieval, NanoBEIR, triplets, reranking \u03ba\u03b1\u03b9 distillation. \u0393\u03b9\u03b1 domain finetuning, \u03b7 \u03bf\u03c5\u03c3\u03b9\u03b1\u03c3\u03c4\u03b9\u03ba\u03ae \u03bc\u03ad\u03c4\u03c1\u03b7\u03c3\u03b7 \u03b5\u03af\u03bd\u03b1\u03b9 \u03c0\u03ac\u03bd\u03c9 \u03c3\u03b5 \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03ac held-out queries, corpus \u03ba\u03b1\u03b9 relevance mappings.<\/p>\n<p>\u039f\u03b9 \u03b5\u03c1\u03c9\u03c4\u03ae\u03c3\u03b5\u03b9\u03c2 \u03c4\u03bf\u03c5 MIRIAD \u03c0\u03b1\u03c1\u03ac\u03b3\u03bf\u03bd\u03c4\u03b1\u03b9 \u03b1\u03c0\u03cc \u03c4\u03b1 \u03af\u03b4\u03b9\u03b1 \u03c4\u03b1 source passages, \u03ac\u03c1\u03b1 \u03ad\u03c7\u03bf\u03c5\u03bd \u03b1\u03c3\u03c5\u03bd\u03ae\u03b8\u03b9\u03c3\u03c4\u03b1 \u03c5\u03c8\u03b7\u03bb\u03cc lexical overlap. \u039c\u03b5 \u03c0\u03b5\u03c1\u03af\u03c0\u03bf\u03c5 10.000 gold passages, \u03c3\u03c7\u03b5\u03b4\u03cc\u03bd \u03cc\u03bb\u03b1 \u03c4\u03b1 \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03b1 \u03be\u03b5\u03c0\u03b5\u03c1\u03bd\u03bf\u03cd\u03c3\u03b1\u03bd \u03c4\u03bf 0,97 NDCG@10 \u03ba\u03b1\u03b9 \u03c4\u03bf benchmark \u03b4\u03b5\u03bd \u03bc\u03c0\u03bf\u03c1\u03bf\u03cd\u03c3\u03b5 \u03bd\u03b1 \u03c4\u03b1 \u03b4\u03b9\u03b1\u03c7\u03c9\u03c1\u03af\u03c3\u03b5\u03b9. \u039f Aarsen \u03c0\u03c1\u03cc\u03c3\u03b8\u03b5\u03c3\u03b5 deduplicated distractors \u03ad\u03c9\u03c2 corpus 200.000 passages, \u03ce\u03c3\u03c4\u03b5 \u03b7 \u03b1\u03bd\u03ac\u03ba\u03c4\u03b7\u03c3\u03b7 \u03bd\u03b1 \u03b3\u03af\u03bd\u03b5\u03b9 \u03b1\u03c1\u03ba\u03b5\u03c4\u03ac \u03b4\u03cd\u03c3\u03ba\u03bf\u03bb\u03b7.<\/p>\n<p>\u0388\u03bd\u03b1 business evaluation \u03c7\u03c1\u03b5\u03b9\u03ac\u03b6\u03b5\u03c4\u03b1\u03b9 \u03b1\u03bc\u03c6\u03af\u03c3\u03b7\u03bc\u03b5\u03c2 \u03b4\u03b9\u03b1\u03c4\u03c5\u03c0\u03ce\u03c3\u03b5\u03b9\u03c2, \u03c0\u03b1\u03c1\u03cc\u03bc\u03bf\u03b9\u03b1 \u03c0\u03c1\u03bf\u03ca\u03cc\u03bd\u03c4\u03b1, \u03c0\u03b1\u03bb\u03b9\u03ad\u03c2 \u03b5\u03ba\u03b4\u03cc\u03c3\u03b5\u03b9\u03c2, near-duplicates, \u03c3\u03cd\u03bd\u03b8\u03b5\u03c4\u03b1 \u03c6\u03af\u03bb\u03c4\u03c1\u03b1 \u03ba\u03b1\u03b9 queries \u03cc\u03c0\u03bf\u03c5 \u03b7 \u03b1\u03c0\u03ac\u03bd\u03c4\u03b7\u03c3\u03b7 \u03b2\u03c1\u03af\u03c3\u03ba\u03b5\u03c4\u03b1\u03b9 \u03b2\u03b1\u03b8\u03b9\u03ac \u03c3\u03c4\u03bf document. \u03a4\u03bf \u03ac\u03c1\u03b8\u03c1\u03bf \u03b3\u03b9\u03b1 \u03c4\u03bf <a href=\"https:\/\/twodots.gr\/causal-rag-reranking-sosti-apantisi\/\">Causal RAG \u03ba\u03b1\u03b9 \u03c4\u03bf reranking<\/a> \u03b1\u03bd\u03b1\u03b4\u03b5\u03b9\u03ba\u03bd\u03cd\u03b5\u03b9 \u03c4\u03bf \u03af\u03b4\u03b9\u03bf \u03ba\u03b5\u03bd\u03cc \u03b1\u03c0\u03cc \u03ac\u03bb\u03bb\u03b7 \u03b3\u03c9\u03bd\u03af\u03b1: \u03b7 \u03bf\u03bc\u03bf\u03b9\u03cc\u03c4\u03b7\u03c4\u03b1 \u03b4\u03b5\u03bd \u03c4\u03b1\u03c5\u03c4\u03af\u03b6\u03b5\u03c4\u03b1\u03b9 \u03c0\u03ac\u03bd\u03c4\u03b1 \u03bc\u03b5 \u03c4\u03b7 \u03c3\u03c9\u03c3\u03c4\u03ae \u03b1\u03b9\u03c4\u03b9\u03ce\u03b4\u03b7 \u03ae \u03b5\u03c0\u03b9\u03c7\u03b5\u03b9\u03c1\u03b7\u03c3\u03b9\u03b1\u03ba\u03ae \u03b1\u03c0\u03ac\u03bd\u03c4\u03b7\u03c3\u03b7.<\/p>\n<p>\u03a4\u03b1 metrics \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03c3\u03c5\u03bd\u03b4\u03ad\u03bf\u03bd\u03c4\u03b1\u03b9 \u03bc\u03b5 \u03c4\u03bf \u03c0\u03c1\u03bf\u03ca\u03cc\u03bd. \u03a4\u03bf NDCG@10 \u03bc\u03b5\u03c4\u03c1\u03ac \u03c4\u03b7\u03bd \u03c0\u03bf\u03b9\u03cc\u03c4\u03b7\u03c4\u03b1 \u03c4\u03b7\u03c2 \u03ba\u03b1\u03c4\u03ac\u03c4\u03b1\u03be\u03b7\u03c2, \u03c4\u03bf accuracy@1 \u03b1\u03bd \u03c4\u03bf \u03c3\u03c9\u03c3\u03c4\u03cc document \u03ad\u03c1\u03c7\u03b5\u03c4\u03b1\u03b9 \u03c0\u03c1\u03ce\u03c4\u03bf \u03ba\u03b1\u03b9 \u03c4\u03bf recall \u03b1\u03bd \u03b5\u03bc\u03c6\u03b1\u03bd\u03af\u03b6\u03b5\u03c4\u03b1\u03b9 \u03bc\u03ad\u03c3\u03b1 \u03c3\u03c4\u03bf shortlist. \u0393\u03b9\u03b1 \u03c0\u03b1\u03c1\u03b1\u03b3\u03c9\u03b3\u03ae \u03c7\u03c1\u03b5\u03b9\u03ac\u03b6\u03bf\u03bd\u03c4\u03b1\u03b9 \u03b5\u03c0\u03b9\u03c0\u03bb\u03ad\u03bf\u03bd latency, index size, \u03ba\u03cc\u03c3\u03c4\u03bf\u03c2 \u03b1\u03bd\u03ac query, freshness \u03ba\u03b1\u03b9 \u03c0\u03bf\u03c3\u03bf\u03c3\u03c4\u03cc queries \u03cc\u03c0\u03bf\u03c5 \u03bf \u03c7\u03c1\u03ae\u03c3\u03c4\u03b7\u03c2 \u03c3\u03c5\u03bd\u03b5\u03c7\u03af\u03b6\u03b5\u03b9 \u03bd\u03b1 \u03c8\u03ac\u03c7\u03bd\u03b5\u03b9 \u03ae \u03b1\u03bd\u03bf\u03af\u03b3\u03b5\u03b9 \u03bb\u03ac\u03b8\u03bf\u03c2 \u03b1\u03c0\u03bf\u03c4\u03ad\u03bb\u03b5\u03c3\u03bc\u03b1.<\/p>\n<h2 id=\"results\">\u03a4\u03b9 \u03ad\u03b4\u03b5\u03b9\u03be\u03b1\u03bd \u03bf\u03b9 50+ retrieval \u03b4\u03b9\u03b1\u03bc\u03bf\u03c1\u03c6\u03ce\u03c3\u03b5\u03b9\u03c2<\/h2>\n<p>\u03a3\u03c4\u03bf \u03c4\u03b5\u03bb\u03b9\u03ba\u03cc evaluation \u03bc\u03b5 1.000 held-out \u03b9\u03b1\u03c4\u03c1\u03b9\u03ba\u03ad\u03c2 \u03b5\u03c1\u03c9\u03c4\u03ae\u03c3\u03b5\u03b9\u03c2 \u03ba\u03b1\u03b9 200.000 \u03bc\u03bf\u03bd\u03b1\u03b4\u03b9\u03ba\u03ac passages, \u03c4\u03bf finetuned <code>mLateOn-medical<\/code> \u03c0\u03ad\u03c4\u03c5\u03c7\u03b5 0,9139 NDCG@10. \u03a4\u03bf \u03b9\u03c3\u03c7\u03c5\u03c1\u03cc\u03c4\u03b5\u03c1\u03bf zero-shot multi-vector \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf \u03b5\u03af\u03c7\u03b5 0,8520, \u03c4\u03bf Qwen3-Embedding-4B 0,7817, \u03c4\u03bf BM25 0,7501 \u03ba\u03b1\u03b9 \u03c4\u03bf SPLADE-v3 0,6853.<\/p>\n<p>\u0397 \u03b4\u03b9\u03b1\u03c6\u03bf\u03c1\u03ac \u03b1\u03c0\u03cc \u03c4\u03bf \u03b9\u03c3\u03c7\u03c5\u03c1\u03cc\u03c4\u03b5\u03c1\u03bf zero-shot configuration \u03ae\u03c4\u03b1\u03bd +0,062 NDCG@10. \u03a3\u03c4\u03bf accuracy@1, \u03c4\u03bf finetuned \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf \u03ad\u03c6\u03c4\u03b1\u03c3\u03b5 0,849 \u03ad\u03bd\u03b1\u03bd\u03c4\u03b9 0,758 \u03c4\u03bf\u03c5 \u03ba\u03b1\u03bb\u03cd\u03c4\u03b5\u03c1\u03bf\u03c5 zero-shot. \u039f\u03b9 \u03b1\u03c1\u03b9\u03b8\u03bc\u03bf\u03af \u03b4\u03b5\u03af\u03c7\u03bd\u03bf\u03c5\u03bd \u03b9\u03c3\u03c7\u03c5\u03c1\u03ae \u03c0\u03c1\u03bf\u03c3\u03b1\u03c1\u03bc\u03bf\u03b3\u03ae \u03c3\u03c4\u03bf \u03c3\u03c5\u03b3\u03ba\u03b5\u03ba\u03c1\u03b9\u03bc\u03ad\u03bd\u03bf medical retrieval task, \u03cc\u03c7\u03b9 \u03cc\u03c4\u03b9 \u03c4\u03bf \u03af\u03b4\u03b9\u03bf checkpoint \u03b5\u03af\u03bd\u03b1\u03b9 \u03b3\u03b5\u03bd\u03b9\u03ba\u03ac \u03ba\u03b1\u03bb\u03cd\u03c4\u03b5\u03c1\u03bf \u03c3\u03b5 \u03ba\u03ac\u03b8\u03b5 \u03b3\u03bb\u03ce\u03c3\u03c3\u03b1, corpus \u03ae \u03c4\u03cd\u03c0\u03bf query.<\/p>\n<p>\u03a4\u03bf BM25 \u03c0\u03b1\u03c1\u03ad\u03bc\u03b5\u03b9\u03bd\u03b5 \u03b1\u03bd\u03c4\u03b1\u03b3\u03c9\u03bd\u03b9\u03c3\u03c4\u03b9\u03ba\u03cc \u03b5\u03c0\u03b5\u03b9\u03b4\u03ae \u03bf\u03b9 \u03b5\u03c1\u03c9\u03c4\u03ae\u03c3\u03b5\u03b9\u03c2 \u03b5\u03af\u03c7\u03b1\u03bd \u03c5\u03c8\u03b7\u03bb\u03cc lexical overlap \u03bc\u03b5 \u03c4\u03b1 gold passages \u03ba\u03b1\u03b9 \u03c4\u03bf lexical retrieval \u03b4\u03b9\u03ac\u03b2\u03b1\u03b6\u03b5 \u03cc\u03bb\u03bf \u03c4\u03bf \u03ba\u03b5\u03af\u03bc\u03b5\u03bd\u03bf. \u0391\u03c5\u03c4\u03cc \u03b4\u03b5\u03bd \u03b5\u03b3\u03b3\u03c5\u03ac\u03c4\u03b1\u03b9 \u03c4\u03bf \u03af\u03b4\u03b9\u03bf \u03b1\u03c0\u03bf\u03c4\u03ad\u03bb\u03b5\u03c3\u03bc\u03b1 \u03b1\u03bb\u03bb\u03bf\u03cd, \u03b1\u03bb\u03bb\u03ac \u03b5\u03c0\u03b9\u03b2\u03b5\u03b2\u03b1\u03b9\u03ce\u03bd\u03b5\u03b9 \u03cc\u03c4\u03b9 \u03ad\u03bd\u03b1 \u03c6\u03b8\u03b7\u03bd\u03cc baseline \u03b5\u03af\u03bd\u03b1\u03b9 \u03b1\u03c0\u03b1\u03c1\u03b1\u03af\u03c4\u03b7\u03c4\u03bf. \u03a7\u03c9\u03c1\u03af\u03c2 BM25 \u03ba\u03b1\u03b9 \u03c4\u03bf \u03c5\u03c0\u03ac\u03c1\u03c7\u03bf\u03bd dense retriever, \u03b7 \u03bf\u03bc\u03ac\u03b4\u03b1 \u03b4\u03b5\u03bd \u03b3\u03bd\u03c9\u03c1\u03af\u03b6\u03b5\u03b9 \u03b1\u03bd \u03b7 \u03c0\u03bf\u03bb\u03c5\u03c0\u03bb\u03bf\u03ba\u03cc\u03c4\u03b7\u03c4\u03b1 \u03c4\u03bf\u03c5 multi-vector stack \u03b1\u03b3\u03bf\u03c1\u03ac\u03b6\u03b5\u03b9 \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03ae \u03b2\u03b5\u03bb\u03c4\u03af\u03c9\u03c3\u03b7.<\/p>\n<div class=\"td-chart td-chart--metrics\">\n<div class=\"td-chart-head\">\n<p class=\"td-chart-kicker\">\u039c\u03b5\u03c4\u03c1\u03ae\u03c3\u03b5\u03b9\u03c2 \u03b1\u03c0\u03cc \u03c4\u03bf medical retrieval experiment<\/p>\n<p class=\"td-chart-title\">\u0397 \u03c0\u03bf\u03b9\u03cc\u03c4\u03b7\u03c4\u03b1 \u03ba\u03c1\u03af\u03bd\u03b5\u03c4\u03b1\u03b9 \u03bc\u03b1\u03b6\u03af \u03bc\u03b5 training \u03ba\u03b1\u03b9 index<\/p>\n<p class=\"td-chart-intro\">\u039f\u03b9 \u03c4\u03b9\u03bc\u03ad\u03c2 \u03b1\u03c6\u03bf\u03c1\u03bf\u03cd\u03bd \u03b1\u03c0\u03bf\u03ba\u03bb\u03b5\u03b9\u03c3\u03c4\u03b9\u03ba\u03ac \u03c4\u03b7 \u03c3\u03c5\u03b3\u03ba\u03b5\u03ba\u03c1\u03b9\u03bc\u03ad\u03bd\u03b7 \u03bc\u03b5\u03bb\u03ad\u03c4\u03b7 Hugging Face \u03c0\u03ac\u03bd\u03c9 \u03c3\u03c4\u03bf MIRIAD. \u0394\u03b5\u03bd \u03b5\u03af\u03bd\u03b1\u03b9 \u03b3\u03b5\u03bd\u03b9\u03ba\u03ac benchmarks \u03b3\u03b9\u03b1 \u03ba\u03ac\u03b8\u03b5 \u03b5\u03c4\u03b1\u03b9\u03c1\u03b9\u03ba\u03ae \u03b1\u03bd\u03b1\u03b6\u03ae\u03c4\u03b7\u03c3\u03b7.<\/p>\n<\/div>\n<div class=\"td-metric-grid\">\n<div class=\"td-metric-card\"><span class=\"td-metric-value\">0,9139<\/span><span class=\"td-metric-label\">NDCG@10 \u03c4\u03bf\u03c5 finetuned mLateOn-medical \u03c3\u03b5 corpus 200.000 passages<\/span><\/div>\n<div class=\"td-metric-card\"><span class=\"td-metric-value\">+0,062<\/span><span class=\"td-metric-label\">NDCG@10 \u03ad\u03bd\u03b1\u03bd\u03c4\u03b9 \u03c4\u03bf\u03c5 \u03b9\u03c3\u03c7\u03c5\u03c1\u03cc\u03c4\u03b5\u03c1\u03bf\u03c5 zero-shot configuration \u03c4\u03b7\u03c2 \u03c3\u03cd\u03b3\u03ba\u03c1\u03b9\u03c3\u03b7\u03c2<\/span><\/div>\n<div class=\"td-metric-card\"><span class=\"td-metric-value\">14,5 \u03ce\u03c1\u03b5\u03c2<\/span><span class=\"td-metric-label\">\u03a0\u03bb\u03ae\u03c1\u03b5\u03c2 training \u03b5\u03bd\u03cc\u03c2 \u03b5\u03ba\u03b1\u03c4\u03bf\u03bc\u03bc\u03c5\u03c1\u03af\u03bf\u03c5 pairs \u03c3\u03b5 \u03bc\u03af\u03b1 RTX 3090 \u03bc\u03b5 peak 17,5 GB VRAM<\/span><\/div>\n<div class=\"td-metric-card\"><span class=\"td-metric-value\">45 GB \u2192 3,37 GB<\/span><span class=\"td-metric-label\">Raw FP16 index \u03ad\u03bd\u03b1\u03bd\u03c4\u03b9 1-bit PLAID \u03bc\u03b5 \u03cc\u03bb\u03b1 \u03c4\u03b1 vectors \u03ba\u03b1\u03b9 \u03bc\u03b9\u03ba\u03c1\u03cc\u03c4\u03b5\u03c1\u03b7 \u03b5\u03c0\u03af\u03b4\u03bf\u03c3\u03b7 0,8984<\/span><\/div>\n<\/div>\n<p class=\"td-chart-source\">\u03a0\u03b7\u03b3\u03ae \u03c4\u03b9\u03bc\u03ce\u03bd: Hugging Face, \u00abTraining and Finetuning Multi-Vector Embedding Models with Sentence Transformers\u00bb, \u0391\u03cd\u03b3\u03bf\u03c5\u03c3\u03c4\u03bf\u03c2 2026.<\/p>\n<\/div>\n<h2 id=\"index-cost\">\u039f index \u03ba\u03b1\u03b8\u03bf\u03c1\u03af\u03b6\u03b5\u03b9 \u03c4\u03bf \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03cc \u03ba\u03cc\u03c3\u03c4\u03bf\u03c2<\/h2>\n<p>\u03a3\u03c4\u03bf MIRIAD, \u03ba\u03ac\u03b8\u03b5 passage \u03c0\u03b1\u03c1\u03ae\u03b3\u03b1\u03b3\u03b5 \u03ba\u03b1\u03c4\u03ac \u03bc\u03ad\u03c3\u03bf \u03cc\u03c1\u03bf \u03c0\u03b5\u03c1\u03af\u03c0\u03bf\u03c5 878 vectors. \u0393\u03b9\u03b1 200.000 \u03bc\u03b5\u03b3\u03ac\u03bb\u03b1 passages, \u03c4\u03b1 raw FP16 embeddings \u03c7\u03c1\u03b5\u03b9\u03ac\u03c3\u03c4\u03b7\u03ba\u03b1\u03bd \u03c0\u03b5\u03c1\u03af\u03c0\u03bf\u03c5 45 GB, \u03b5\u03bd\u03ce \u03ad\u03bd\u03b1 dense index \u03c7\u03c1\u03b5\u03b9\u03b1\u03b6\u03cc\u03c4\u03b1\u03bd \u03bb\u03b9\u03b3\u03cc\u03c4\u03b5\u03c1\u03bf \u03b1\u03c0\u03cc 1 GB. \u03a3\u03b5 \u03bc\u03b9\u03ba\u03c1\u03cc\u03c4\u03b5\u03c1\u03b1 Natural Questions passages, \u03bf \u03bc\u03ad\u03c3\u03bf\u03c2 \u03cc\u03c1\u03bf\u03c2 \u03ae\u03c4\u03b1\u03bd \u03c0\u03b5\u03c1\u03af\u03c0\u03bf\u03c5 125 vectors, \u03b5\u03c0\u03c4\u03ac \u03c6\u03bf\u03c1\u03ad\u03c2 \u03c7\u03b1\u03bc\u03b7\u03bb\u03cc\u03c4\u03b5\u03c1\u03bf\u03c2. \u03a4\u03bf document length \u03b1\u03bb\u03bb\u03ac\u03b6\u03b5\u03b9 \u03b5\u03c0\u03bf\u03bc\u03ad\u03bd\u03c9\u03c2 \u03ba\u03b1\u03b9 \u03c4\u03b7\u03bd \u03bf\u03b9\u03ba\u03bf\u03bd\u03bf\u03bc\u03af\u03b1 \u03c4\u03b7\u03c2 \u03b1\u03c1\u03c7\u03b9\u03c4\u03b5\u03ba\u03c4\u03bf\u03bd\u03b9\u03ba\u03ae\u03c2.<\/p>\n<p>\u03a4\u03bf token pooling \u03bc\u03b5\u03b9\u03ce\u03bd\u03b5\u03b9 \u03c0\u03cc\u03c3\u03b1 vectors \u03b1\u03c0\u03bf\u03b8\u03b7\u03ba\u03b5\u03cd\u03bf\u03bd\u03c4\u03b1\u03b9. \u0397 \u03b4\u03b9\u03b1\u03c4\u03ae\u03c1\u03b7\u03c3\u03b7 \u03c4\u03c9\u03bd \u03bc\u03b9\u03c3\u03ce\u03bd \u03ba\u03cc\u03c3\u03c4\u03b9\u03c3\u03b5 0,0033 NDCG@10 \u03c7\u03c9\u03c1\u03af\u03c2 \u03b1\u03bb\u03bb\u03b1\u03b3\u03ae \u03c3\u03c4\u03bf rank-1 accuracy, \u03b5\u03bd\u03ce \u03c4\u03bf \u03ad\u03bd\u03b1 \u03c4\u03ad\u03c4\u03b1\u03c1\u03c4\u03bf \u03ad\u03b4\u03c9\u03c3\u03b5 index 11,2 GB \u03ba\u03b1\u03b9 score 0,8991. \u039c\u03b5 1-bit residual quantization \u03bc\u03ad\u03c3\u03c9 fast-plaid, \u03cc\u03bb\u03b1 \u03c4\u03b1 vectors \u03c7\u03ce\u03c1\u03b5\u03c3\u03b1\u03bd \u03c3\u03b5 3,37 GB \u03bc\u03b5 0,8984 NDCG@10.<\/p>\n<p>\u039c\u03b5 \u03c0\u03c1\u03cc\u03c3\u03b8\u03b5\u03c4\u03bf pruning \u03c3\u03c4\u03bf 65% \u03c4\u03c9\u03bd vectors, \u03bf index \u03ad\u03c0\u03b5\u03c3\u03b5 \u03c3\u03c4\u03b1 2,23 GB \u03ba\u03b1\u03b9 0,8830. \u03a3\u03c4\u03bf 42% \u03ad\u03c6\u03c4\u03b1\u03c3\u03b5 1,45 GB \u03ba\u03b1\u03b9 0,8642. \u0397 \u03af\u03b4\u03b9\u03b1 \u03b7 \u03c0\u03b7\u03b3\u03ae \u03c7\u03b1\u03c1\u03b1\u03ba\u03c4\u03b7\u03c1\u03af\u03b6\u03b5\u03b9 \u03c4\u03bf pruning \u03b1\u03c0\u03bb\u03cc \u03ba\u03b1\u03b9 \u03b4\u03b9\u03b5\u03c1\u03b5\u03c5\u03bd\u03b7\u03c4\u03b9\u03ba\u03cc, \u03ac\u03c1\u03b1 \u03c4\u03b1 \u03c4\u03b5\u03bb\u03b5\u03c5\u03c4\u03b1\u03af\u03b1 \u03c3\u03b7\u03bc\u03b5\u03af\u03b1 \u03b4\u03b5\u03bd \u03b5\u03af\u03bd\u03b1\u03b9 \u03c4\u03b5\u03bb\u03b9\u03ba\u03cc\u03c2 \u03b2\u03ad\u03bb\u03c4\u03b9\u03c3\u03c4\u03bf\u03c2 \u03c3\u03c7\u03b5\u03b4\u03b9\u03b1\u03c3\u03bc\u03cc\u03c2. \u0394\u03b5\u03af\u03c7\u03bd\u03bf\u03c5\u03bd \u03cc\u03bc\u03c9\u03c2 \u03cc\u03c4\u03b9 quantization, pooling \u03ba\u03b1\u03b9 pruning \u03b5\u03af\u03bd\u03b1\u03b9 \u03b1\u03c0\u03bf\u03c6\u03ac\u03c3\u03b5\u03b9\u03c2 \u03c0\u03c1\u03bf\u03ca\u03cc\u03bd\u03c4\u03bf\u03c2, \u03cc\u03c7\u03b9 \u03b4\u03b5\u03c5\u03c4\u03b5\u03c1\u03b5\u03cd\u03bf\u03c5\u03c3\u03b5\u03c2 \u03bb\u03b5\u03c0\u03c4\u03bf\u03bc\u03ad\u03c1\u03b5\u03b9\u03b5\u03c2 \u03c5\u03c0\u03bf\u03b4\u03bf\u03bc\u03ae\u03c2.<\/p>\n<p>\u0397 \u03b5\u03c0\u03b9\u03c7\u03b5\u03af\u03c1\u03b7\u03c3\u03b7 \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03c3\u03c5\u03b3\u03ba\u03c1\u03af\u03bd\u03b5\u03b9 quality\u2013cost curves, \u03cc\u03c7\u03b9 \u03ad\u03bd\u03b1 \u03bc\u03cc\u03bd\u03bf score. \u0388\u03bd\u03b1 \u03b1\u03c3\u03c6\u03b1\u03bb\u03ad\u03c2 enterprise RAG \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03b5\u03c0\u03af\u03c3\u03b7\u03c2 \u03bd\u03b1 \u03c7\u03c1\u03b5\u03b9\u03ac\u03b6\u03b5\u03c4\u03b1\u03b9 on-premises \u03ae \u03b1\u03c0\u03bf\u03bc\u03bf\u03bd\u03c9\u03bc\u03ad\u03bd\u03b7 \u03c5\u03c0\u03bf\u03b4\u03bf\u03bc\u03ae, \u03cc\u03c0\u03c9\u03c2 \u03b4\u03b5\u03af\u03c7\u03bd\u03b5\u03b9 \u03b7 \u03c3\u03c5\u03b6\u03ae\u03c4\u03b7\u03c3\u03b7 \u03b3\u03b9\u03b1 <a href=\"https:\/\/twodots.gr\/spyre-linuxone-secure-enterprise-rag\/\">enterprise RAG \u03bc\u03b5 \u03b5\u03c5\u03b1\u03af\u03c3\u03b8\u03b7\u03c4\u03b1 \u03b4\u03b5\u03b4\u03bf\u03bc\u03ad\u03bd\u03b1<\/a>. Storage, latency, \u03c0\u03c1\u03cc\u03c3\u03b2\u03b1\u03c3\u03b7, observability \u03ba\u03b1\u03b9 freshness \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03bc\u03c0\u03bf\u03c5\u03bd \u03c3\u03c4\u03bf \u03af\u03b4\u03b9\u03bf business case.<\/p>\n<div class=\"td-decision-band\">\n<p class=\"td-decision-kicker\">\u039a\u03b1\u03bd\u03cc\u03bd\u03b1\u03c2 \u03b1\u03c0\u03cc\u03c6\u03b1\u03c3\u03b7\u03c2<\/p>\n<p class=\"td-decision-title\">\u039c\u03b7\u03bd \u03c5\u03b9\u03bf\u03b8\u03b5\u03c4\u03b5\u03af\u03c4\u03b5 multi-vector retrieval \u03b5\u03c0\u03b5\u03b9\u03b4\u03ae \u03ba\u03ad\u03c1\u03b4\u03b9\u03c3\u03b5 \u03ad\u03bd\u03b1 benchmark<\/p>\n<p>\u03a0\u03c1\u03bf\u03c7\u03c9\u03c1\u03ae\u03c3\u03c4\u03b5 \u03cc\u03c4\u03b1\u03bd \u03c4\u03bf dense \u03ae lexical baseline \u03c7\u03ac\u03bd\u03b5\u03b9 \u03ba\u03c1\u03af\u03c3\u03b9\u03bc\u03b5\u03c2 \u03bb\u03b5\u03c0\u03c4\u03bf\u03bc\u03ad\u03c1\u03b5\u03b9\u03b5\u03c2 \u03c3\u03b5 \u03bc\u03b5\u03b3\u03ac\u03bb\u03b1 domain documents \u03ba\u03b1\u03b9 \u03b7 \u03bc\u03b5\u03c4\u03c1\u03b7\u03bc\u03ad\u03bd\u03b7 \u03ac\u03bd\u03bf\u03b4\u03bf\u03c2 \u03c3\u03b5 NDCG, accuracy \u03ae task success \u03b4\u03b9\u03ba\u03b1\u03b9\u03bf\u03bb\u03bf\u03b3\u03b5\u03af index size, latency \u03ba\u03b1\u03b9 \u03bb\u03b5\u03b9\u03c4\u03bf\u03c5\u03c1\u03b3\u03b9\u03ba\u03ae \u03c0\u03bf\u03bb\u03c5\u03c0\u03bb\u03bf\u03ba\u03cc\u03c4\u03b7\u03c4\u03b1. \u0391\u03bd \u03c4\u03b1 documents \u03b5\u03af\u03bd\u03b1\u03b9 \u03c3\u03cd\u03bd\u03c4\u03bf\u03bc\u03b1 \u03ae \u03ad\u03bd\u03b1 hybrid baseline \u03ba\u03b1\u03bb\u03cd\u03c0\u03c4\u03b5\u03b9 \u03c4\u03bf use case, \u03b7 \u03b1\u03c0\u03bb\u03bf\u03cd\u03c3\u03c4\u03b5\u03c1\u03b7 \u03bb\u03cd\u03c3\u03b7 \u03c0\u03b1\u03c1\u03b1\u03bc\u03ad\u03bd\u03b5\u03b9 \u03c3\u03c5\u03c7\u03bd\u03ac \u03ba\u03b1\u03bb\u03cd\u03c4\u03b5\u03c1\u03b7.<\/p>\n<\/div>\n<h2 id=\"decision-framework\">\u03a0\u03cc\u03c4\u03b5 \u03ad\u03c7\u03b5\u03b9 \u03bd\u03cc\u03b7\u03bc\u03b1 \u03b3\u03b9\u03b1 \u03b5\u03c4\u03b1\u03b9\u03c1\u03b9\u03ba\u03ae \u03b1\u03bd\u03b1\u03b6\u03ae\u03c4\u03b7\u03c3\u03b7<\/h2>\n<p>\u03a4\u03bf \u03c0\u03c1\u03ce\u03c4\u03bf \u03b8\u03b5\u03c4\u03b9\u03ba\u03cc \u03c3\u03ae\u03bc\u03b1 \u03b5\u03af\u03bd\u03b1\u03b9 \u03c4\u03bf \u03bc\u03ae\u03ba\u03bf\u03c2: policies, \u03c3\u03c5\u03bc\u03b2\u03ac\u03c3\u03b5\u03b9\u03c2, \u03b5\u03b3\u03c7\u03b5\u03b9\u03c1\u03af\u03b4\u03b9\u03b1 \u03ae product catalogs \u03cc\u03c0\u03bf\u03c5 \u03c4\u03bf \u03c7\u03c1\u03ae\u03c3\u03b9\u03bc\u03bf \u03b1\u03c0\u03cc\u03c3\u03c0\u03b1\u03c3\u03bc\u03b1 \u03c7\u03ac\u03bd\u03b5\u03c4\u03b1\u03b9 \u03bc\u03b5\u03c4\u03ac \u03b1\u03c0\u03cc truncation. \u03a4\u03bf \u03b4\u03b5\u03cd\u03c4\u03b5\u03c1\u03bf \u03b5\u03af\u03bd\u03b1\u03b9 \u03b7 \u03bf\u03c1\u03bf\u03bb\u03bf\u03b3\u03af\u03b1: \u03bf\u03bd\u03bf\u03bc\u03b1\u03c3\u03af\u03b5\u03c2, \u03b5\u03be\u03b1\u03c1\u03c4\u03ae\u03bc\u03b1\u03c4\u03b1, \u03ba\u03c9\u03b4\u03b9\u03ba\u03bf\u03af \u03ae \u03c6\u03c1\u03ac\u03c3\u03b5\u03b9\u03c2 \u03c0\u03bf\u03c5 \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03c4\u03b1\u03b9\u03c1\u03b9\u03ac\u03be\u03bf\u03c5\u03bd \u03bb\u03b5\u03c0\u03c4\u03bf\u03bc\u03b5\u03c1\u03ce\u03c2. \u03a4\u03bf \u03c4\u03c1\u03af\u03c4\u03bf \u03b5\u03af\u03bd\u03b1\u03b9 \u03b7 \u03b1\u03be\u03af\u03b1 \u03c4\u03bf\u03c5 \u03bb\u03ac\u03b8\u03bf\u03c5\u03c2: \u03cc\u03c4\u03b1\u03bd \u03bc\u03b9\u03b1 \u03bb\u03b1\u03bd\u03b8\u03b1\u03c3\u03bc\u03ad\u03bd\u03b7 \u03b1\u03bd\u03ac\u03ba\u03c4\u03b7\u03c3\u03b7 \u03bf\u03b4\u03b7\u03b3\u03b5\u03af \u03c3\u03b5 \u03bb\u03ac\u03b8\u03bf\u03c2 \u03b1\u03c0\u03ac\u03bd\u03c4\u03b7\u03c3\u03b7 \u03c0\u03b5\u03bb\u03ac\u03c4\u03b7, \u03b5\u03c0\u03b9\u03c3\u03c4\u03c1\u03bf\u03c6\u03ae \u03c0\u03c1\u03bf\u03ca\u03cc\u03bd\u03c4\u03bf\u03c2 \u03ae \u03c4\u03b5\u03c7\u03bd\u03b9\u03ba\u03ae \u03b4\u03b9\u03b1\u03ba\u03bf\u03c0\u03ae.<\/p>\n<p>\u03a3\u03b5 e-commerce, \u03c4\u03bf use case \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03b5\u03af\u03bd\u03b1\u03b9 product discovery \u03bc\u03b5 \u03c0\u03bf\u03bb\u03bb\u03ac \u03c4\u03b1\u03c5\u03c4\u03cc\u03c7\u03c1\u03bf\u03bd\u03b1 \u03c7\u03b1\u03c1\u03b1\u03ba\u03c4\u03b7\u03c1\u03b9\u03c3\u03c4\u03b9\u03ba\u03ac. \u03a3\u03b5 support, \u03b7 \u03c3\u03c9\u03c3\u03c4\u03ae \u03bb\u03cd\u03c3\u03b7 \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03b5\u03be\u03b1\u03c1\u03c4\u03ac\u03c4\u03b1\u03b9 \u03b1\u03c0\u03cc \u03ad\u03ba\u03b4\u03bf\u03c3\u03b7, \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf \u03ba\u03b1\u03b9 \u03c3\u03c5\u03b3\u03ba\u03b5\u03ba\u03c1\u03b9\u03bc\u03ad\u03bd\u03bf error condition. \u03a3\u03b5 compliance \u03ae \u03bd\u03bf\u03bc\u03b9\u03ba\u03ac documents, \u03bc\u03b9\u03b1 \u03b5\u03be\u03b1\u03af\u03c1\u03b5\u03c3\u03b7 \u03ad\u03c7\u03b5\u03b9 \u03bc\u03b5\u03b3\u03b1\u03bb\u03cd\u03c4\u03b5\u03c1\u03b7 \u03c3\u03b7\u03bc\u03b1\u03c3\u03af\u03b1 \u03b1\u03c0\u03cc \u03c4\u03b7 \u03b3\u03b5\u03bd\u03b9\u03ba\u03ae \u03b8\u03b5\u03bc\u03b1\u03c4\u03b9\u03ba\u03ae \u03bf\u03bc\u03bf\u03b9\u03cc\u03c4\u03b7\u03c4\u03b1. \u0397 <a href=\"https:\/\/twodots.gr\/multimodal-anazitisi-eikonon-embeddings-frontier-llms\/\">multimodal \u03b1\u03bd\u03b1\u03b6\u03ae\u03c4\u03b7\u03c3\u03b7 \u03b5\u03b9\u03ba\u03cc\u03bd\u03c9\u03bd \u03bc\u03b5 embeddings<\/a> \u03c5\u03c0\u03b5\u03bd\u03b8\u03c5\u03bc\u03af\u03b6\u03b5\u03b9 \u03cc\u03c4\u03b9 \u03b7 \u03c3\u03c9\u03c3\u03c4\u03ae \u03b1\u03bd\u03b1\u03c0\u03b1\u03c1\u03ac\u03c3\u03c4\u03b1\u03c3\u03b7 \u03b5\u03be\u03b1\u03c1\u03c4\u03ac\u03c4\u03b1\u03b9 \u03ba\u03b1\u03b9 \u03b1\u03c0\u03cc \u03c4\u03bf modality, \u03cc\u03c7\u03b9 \u03bc\u03cc\u03bd\u03bf \u03b1\u03c0\u03cc \u03c4\u03bf\u03bd \u03b1\u03bb\u03b3\u03cc\u03c1\u03b9\u03b8\u03bc\u03bf \u03ba\u03b1\u03c4\u03ac\u03c4\u03b1\u03be\u03b7\u03c2.<\/p>\n<p>\u03a5\u03c0\u03ac\u03c1\u03c7\u03bf\u03c5\u03bd \u03ba\u03b1\u03b9 \u03c3\u03b1\u03c6\u03ae \u03b1\u03c1\u03bd\u03b7\u03c4\u03b9\u03ba\u03ac \u03c3\u03ae\u03bc\u03b1\u03c4\u03b1. \u0391\u03bd \u03b4\u03b5\u03bd \u03c5\u03c0\u03ac\u03c1\u03c7\u03bf\u03c5\u03bd \u03b1\u03be\u03b9\u03cc\u03c0\u03b9\u03c3\u03c4\u03b1 relevance labels, \u03b1\u03bd \u03c4\u03bf corpus \u03b1\u03bb\u03bb\u03ac\u03b6\u03b5\u03b9 \u03c7\u03c9\u03c1\u03af\u03c2 versioning, \u03b1\u03bd \u03b4\u03b5\u03bd \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03bc\u03b5\u03c4\u03c1\u03b7\u03b8\u03b5\u03af latency \u03ae \u03b1\u03bd \u03b7 \u03bf\u03bc\u03ac\u03b4\u03b1 \u03b4\u03b5\u03bd \u03b4\u03b9\u03b1\u03b8\u03ad\u03c4\u03b5\u03b9 \u03b4\u03b9\u03b1\u03b4\u03b9\u03ba\u03b1\u03c3\u03af\u03b1 rebuild \u03c4\u03bf\u03c5 index, \u03c4\u03bf finetuning \u03ba\u03b9\u03bd\u03b4\u03c5\u03bd\u03b5\u03cd\u03b5\u03b9 \u03bd\u03b1 \u03b2\u03b5\u03bb\u03c4\u03b9\u03c3\u03c4\u03bf\u03c0\u03bf\u03b9\u03ae\u03c3\u03b5\u03b9 \u03bc\u03b9\u03b1 \u03b1\u03c3\u03c4\u03b1\u03b8\u03ae \u03b2\u03ac\u03c3\u03b7. \u03a0\u03c1\u03ce\u03c4\u03b1 \u03b4\u03b9\u03bf\u03c1\u03b8\u03ce\u03bd\u03b5\u03c4\u03b1\u03b9 \u03c4\u03bf information architecture \u03ba\u03b1\u03b9 \u03bc\u03b5\u03c4\u03ac \u03b1\u03c5\u03be\u03ac\u03bd\u03b5\u03c4\u03b1\u03b9 \u03b7 \u03c0\u03bf\u03bb\u03c5\u03c0\u03bb\u03bf\u03ba\u03cc\u03c4\u03b7\u03c4\u03b1.<\/p>\n<p>\u0397 \u03b4\u03c1\u03bf\u03bc\u03bf\u03bb\u03cc\u03b3\u03b7\u03c3\u03b7 \u03b4\u03b5\u03b4\u03bf\u03bc\u03ad\u03bd\u03c9\u03bd \u03ad\u03c7\u03b5\u03b9 \u03b5\u03c0\u03af\u03c3\u03b7\u03c2 \u03c3\u03b7\u03bc\u03b1\u03c3\u03af\u03b1. \u0388\u03bd\u03b1 \u03c3\u03cd\u03c3\u03c4\u03b7\u03bc\u03b1 \u03cc\u03c0\u03c9\u03c2 \u03c4\u03bf <a href=\"https:\/\/twodots.gr\/schemarouter-ai-agents-field-aware-rag\/\">SchemaRouter \u03b3\u03b9\u03b1 field-aware RAG<\/a> \u03b1\u03bd\u03c4\u03b9\u03bc\u03b5\u03c4\u03c9\u03c0\u03af\u03b6\u03b5\u03b9 \u03c4\u03bf \u03b5\u03c1\u03ce\u03c4\u03b7\u03bc\u03b1 \u03c0\u03bf\u03b9\u03b1 \u03c0\u03b5\u03b4\u03af\u03b1 \u03ba\u03b1\u03b9 \u03c0\u03b7\u03b3\u03ad\u03c2 \u03c7\u03c1\u03b5\u03b9\u03ac\u03b6\u03bf\u03bd\u03c4\u03b1\u03b9 \u03c0\u03c1\u03b9\u03bd \u03b3\u03b5\u03bc\u03af\u03c3\u03b5\u03b9 \u03c4\u03bf context. \u03a4\u03bf multi-vector retrieval \u03b1\u03c0\u03b1\u03bd\u03c4\u03ac \u03b4\u03b9\u03b1\u03c6\u03bf\u03c1\u03b5\u03c4\u03b9\u03ba\u03ac: \u03c0\u03bf\u03b9\u03b1 \u03bb\u03b5\u03c0\u03c4\u03bf\u03bc\u03b5\u03c1\u03ae tokens \u03bc\u03ad\u03c3\u03b1 \u03c3\u03c4\u03b1 \u03c5\u03c0\u03bf\u03c8\u03ae\u03c6\u03b9\u03b1 documents \u03c4\u03b1\u03b9\u03c1\u03b9\u03ac\u03b6\u03bf\u03c5\u03bd \u03ba\u03b1\u03bb\u03cd\u03c4\u03b5\u03c1\u03b1. \u039f\u03b9 \u03b4\u03cd\u03bf \u03b5\u03c0\u03b9\u03bb\u03bf\u03b3\u03ad\u03c2 \u03bc\u03c0\u03bf\u03c1\u03bf\u03cd\u03bd \u03bd\u03b1 \u03c3\u03c5\u03bd\u03b5\u03c1\u03b3\u03b1\u03c3\u03c4\u03bf\u03cd\u03bd, \u03b1\u03bb\u03bb\u03ac \u03b4\u03b5\u03bd \u03b5\u03af\u03bd\u03b1\u03b9 \u03c5\u03c0\u03bf\u03ba\u03b1\u03c4\u03ac\u03c3\u03c4\u03b1\u03c4\u03b1.<\/p>\n<h2 id=\"pilot-steps\">\u0395\u03c0\u03c4\u03ac \u03b2\u03ae\u03bc\u03b1\u03c4\u03b1 \u03b3\u03b9\u03b1 \u03b1\u03c3\u03c6\u03b1\u03bb\u03ad\u03c2 pilot<\/h2>\n<p>\u0388\u03bd\u03b1 pilot \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03b1\u03c0\u03bf\u03bc\u03bf\u03bd\u03ce\u03bd\u03b5\u03b9 \u03c4\u03b9\u03c2 \u03bc\u03b5\u03c4\u03b1\u03b2\u03bb\u03b7\u03c4\u03ad\u03c2 \u03ba\u03b1\u03b9 \u03bd\u03b1 \u03b1\u03c6\u03ae\u03bd\u03b5\u03b9 audit trail. \u0397 \u03bf\u03bc\u03ac\u03b4\u03b1 \u03b4\u03b5\u03bd \u03c7\u03c1\u03b5\u03b9\u03ac\u03b6\u03b5\u03c4\u03b1\u03b9 \u03bd\u03b1 \u03be\u03b5\u03ba\u03b9\u03bd\u03ae\u03c3\u03b5\u03b9 \u03bc\u03b5 \u03ad\u03bd\u03b1 \u03b5\u03ba\u03b1\u03c4\u03bf\u03bc\u03bc\u03cd\u03c1\u03b9\u03bf pairs \u03ae production traffic. \u03a3\u03c4\u03b1 scaling experiments \u03c4\u03b7\u03c2 Hugging Face, 100.000 pairs \u03ba\u03b1\u03b9 75 \u03bb\u03b5\u03c0\u03c4\u03ac training \u03b2\u03c1\u03ad\u03b8\u03b7\u03ba\u03b1\u03bd \u03b5\u03bd\u03c4\u03cc\u03c2 0,012 NDCG@10 \u03b1\u03c0\u03cc \u03c4\u03bf \u03c0\u03bb\u03ae\u03c1\u03b5\u03c2 run\u00b7 \u03c4\u03bf \u03b1\u03c0\u03bf\u03c4\u03ad\u03bb\u03b5\u03c3\u03bc\u03b1 \u03b1\u03c6\u03bf\u03c1\u03ac \u03c4\u03b7 \u03c3\u03c5\u03b3\u03ba\u03b5\u03ba\u03c1\u03b9\u03bc\u03ad\u03bd\u03b7 GPU \u03ba\u03b1\u03b9 dataset, \u03b1\u03bb\u03bb\u03ac \u03b4\u03b5\u03af\u03c7\u03bd\u03b5\u03b9 \u03b3\u03b9\u03b1\u03c4\u03af \u03ad\u03bd\u03b1 \u03bc\u03b9\u03ba\u03c1\u03cc\u03c4\u03b5\u03c1\u03bf \u03c0\u03b5\u03af\u03c1\u03b1\u03bc\u03b1 \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03b5\u03af\u03bd\u03b1\u03b9 \u03c0\u03bb\u03b7\u03c1\u03bf\u03c6\u03bf\u03c1\u03b9\u03b1\u03ba\u03ac \u03c7\u03c1\u03ae\u03c3\u03b9\u03bc\u03bf.<\/p>\n<div class=\"td-step-list\">\n<p class=\"td-step-list-title\">\u0391\u03c0\u03cc \u03c4\u03bf corpus \u03c3\u03b5 \u03bc\u03b5\u03c4\u03c1\u03ae\u03c3\u03b9\u03bc\u03b7 \u03b5\u03c4\u03b1\u03b9\u03c1\u03b9\u03ba\u03ae \u03b1\u03bd\u03b1\u03b6\u03ae\u03c4\u03b7\u03c3\u03b7<\/p>\n<ol>\n<li><span class=\"td-step-kicker\">\u0421\u0442\u044a\u043f\u043a\u0430 1<\/span><strong>\u039f\u03c1\u03af\u03c3\u03c4\u03b5 \u03c4\u03bf \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03cc search task<\/strong>\n<p>\u039a\u03b1\u03c4\u03b1\u03b3\u03c1\u03ac\u03c8\u03c4\u03b5 \u03c0\u03bf\u03b9\u03bf\u03c2 \u03c1\u03c9\u03c4\u03ac, \u03c0\u03bf\u03b9\u03bf document \u03b8\u03b5\u03c9\u03c1\u03b5\u03af\u03c4\u03b1\u03b9 \u03c3\u03c9\u03c3\u03c4\u03cc \u03ba\u03b1\u03b9 \u03c0\u03bf\u03b9\u03b1 \u03b5\u03c0\u03b9\u03c7\u03b5\u03b9\u03c1\u03b7\u03c3\u03b9\u03b1\u03ba\u03ae \u03c3\u03c5\u03bd\u03ad\u03c0\u03b5\u03b9\u03b1 \u03ad\u03c7\u03b5\u03b9 \u03bc\u03b9\u03b1 \u03bb\u03ac\u03b8\u03bf\u03c2 \u03b1\u03bd\u03ac\u03ba\u03c4\u03b7\u03c3\u03b7 \u03c3\u03b5 support, e-commerce \u03ae \u03b5\u03c3\u03c9\u03c4\u03b5\u03c1\u03b9\u03ba\u03ae \u03b3\u03bd\u03ce\u03c3\u03b7.<\/p>\n<\/li>\n<li><span class=\"td-step-kicker\">\u0421\u0442\u044a\u043f\u043a\u0430 2<\/span><strong>\u039a\u03b1\u03b8\u03b1\u03c1\u03af\u03c3\u03c4\u03b5 \u03ba\u03b1\u03b9 \u03b5\u03ba\u03b4\u03ce\u03c3\u03c4\u03b5 \u03c4\u03bf corpus<\/strong>\n<p>\u0391\u03c6\u03b1\u03b9\u03c1\u03ad\u03c3\u03c4\u03b5 duplicates, \u03b5\u03c0\u03b9\u03c3\u03b7\u03bc\u03ac\u03bd\u03b5\u03c4\u03b5 \u03c0\u03b1\u03bb\u03b9\u03ad\u03c2 \u03b5\u03ba\u03b4\u03cc\u03c3\u03b5\u03b9\u03c2, \u03bf\u03c1\u03af\u03c3\u03c4\u03b5 authoritative sources \u03ba\u03b1\u03b9 \u03ba\u03c1\u03b1\u03c4\u03ae\u03c3\u03c4\u03b5 \u03c3\u03c4\u03b1\u03b8\u03b5\u03c1\u03cc snapshot \u03b3\u03b9\u03b1 train, evaluation \u03ba\u03b1\u03b9 index rebuild.<\/p>\n<\/li>\n<li><span class=\"td-step-kicker\">\u0421\u0442\u044a\u043f\u043a\u0430 3<\/span><strong>\u039c\u03b5\u03c4\u03c1\u03ae\u03c3\u03c4\u03b5 \u03bc\u03ae\u03ba\u03bf\u03c2 \u03ba\u03b1\u03b9 truncation<\/strong>\n<p>\u0392\u03c1\u03b5\u03af\u03c4\u03b5 \u03c0\u03cc\u03c3\u03b1 documents \u03be\u03b5\u03c0\u03b5\u03c1\u03bd\u03bf\u03cd\u03bd \u03c4\u03b1 length caps \u03ba\u03b1\u03b9 \u03c0\u03cc\u03c3\u03bf \u03c3\u03c5\u03c7\u03bd\u03ac \u03c4\u03bf relevant passage \u03b2\u03c1\u03af\u03c3\u03ba\u03b5\u03c4\u03b1\u03b9 \u03b5\u03ba\u03c4\u03cc\u03c2 \u03c4\u03bf\u03c5 \u03c4\u03bc\u03ae\u03bc\u03b1\u03c4\u03bf\u03c2 \u03c0\u03bf\u03c5 \u03b2\u03bb\u03ad\u03c0\u03b5\u03b9 \u03c4\u03bf \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf.<\/p>\n<\/li>\n<li><span class=\"td-step-kicker\">\u0421\u0442\u044a\u043f\u043a\u0430 4<\/span><strong>\u03a7\u03c4\u03af\u03c3\u03c4\u03b5 \u03b4\u03cd\u03c3\u03ba\u03bf\u03bb\u03bf held-out set<\/strong>\n<p>\u03a7\u03c1\u03b7\u03c3\u03b9\u03bc\u03bf\u03c0\u03bf\u03b9\u03ae\u03c3\u03c4\u03b5 \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03ac queries, near-duplicates, \u03c0\u03b1\u03bb\u03b9\u03ad\u03c2 \u03b5\u03ba\u03b4\u03cc\u03c3\u03b5\u03b9\u03c2 \u03ba\u03b1\u03b9 hard distractors, \u03c7\u03c9\u03c1\u03af\u03c2 \u03b4\u03b9\u03b1\u03c1\u03c1\u03bf\u03ae \u03b1\u03c0\u03cc \u03c4\u03bf training set \u03ba\u03b1\u03b9 \u03c7\u03c9\u03c1\u03af\u03c2 \u03c3\u03c5\u03bd\u03b8\u03b5\u03c4\u03b9\u03ba\u03ae \u03b5\u03c5\u03ba\u03bf\u03bb\u03af\u03b1.<\/p>\n<\/li>\n<li><span class=\"td-step-kicker\">\u0421\u0442\u044a\u043f\u043a\u0430 5<\/span><strong>\u03a3\u03c5\u03b3\u03ba\u03c1\u03af\u03bd\u03b5\u03c4\u03b5 \u03c4\u03c1\u03af\u03b1 baselines<\/strong>\n<p>\u039c\u03b5\u03c4\u03c1\u03ae\u03c3\u03c4\u03b5 BM25, \u03c4\u03bf \u03c5\u03c0\u03ac\u03c1\u03c7\u03bf\u03bd dense retriever \u03ba\u03b1\u03b9 \u03ad\u03bd\u03b1 multi-vector checkpoint \u03bc\u03b5 \u03c4\u03b1 \u03af\u03b4\u03b9\u03b1 labels, metrics, document lengths \u03ba\u03b1\u03b9 hardware conditions.<\/p>\n<\/li>\n<li><span class=\"td-step-kicker\">\u0421\u0442\u044a\u043f\u043a\u0430 6<\/span><strong>\u0395\u03ba\u03c0\u03b1\u03b9\u03b4\u03b5\u03cd\u03c3\u03c4\u03b5 \u03ba\u03b1\u03b9 \u03c3\u03c5\u03bc\u03c0\u03b9\u03ad\u03c3\u03c4\u03b5 \u03be\u03b5\u03c7\u03c9\u03c1\u03b9\u03c3\u03c4\u03ac<\/strong>\n<p>\u0394\u03bf\u03ba\u03b9\u03bc\u03ac\u03c3\u03c4\u03b5 pre-supervised checkpoint, finished checkpoint \u03ba\u03b1\u03b9 fresh projection \u03ba\u03b1\u03b9 \u03bc\u03b5\u03c4\u03ac \u03b1\u03be\u03b9\u03bf\u03bb\u03bf\u03b3\u03ae\u03c3\u03c4\u03b5 quantization, pooling \u03ae pruning \u03c7\u03c9\u03c1\u03af\u03c2 \u03bd\u03b1 \u03bc\u03c0\u03b5\u03c1\u03b4\u03b5\u03cd\u03b5\u03c4\u03b5 \u03c4\u03b9\u03c2 \u03b1\u03b9\u03c4\u03af\u03b5\u03c2 \u03c4\u03b7\u03c2 \u03b2\u03b5\u03bb\u03c4\u03af\u03c9\u03c3\u03b7\u03c2.<\/p>\n<\/li>\n<li><span class=\"td-step-kicker\">\u0421\u0442\u044a\u043f\u043a\u0430 7<\/span><strong>\u039a\u03ac\u03bd\u03c4\u03b5 shadow test \u03c0\u03c1\u03b9\u03bd \u03c4\u03bf rollout<\/strong>\n<p>\u039c\u03b5\u03c4\u03c1\u03ae\u03c3\u03c4\u03b5 relevance, latency, index size, freshness \u03ba\u03b1\u03b9 failure cases \u03b4\u03af\u03c0\u03bb\u03b1 \u03c3\u03c4\u03bf \u03c5\u03c0\u03ac\u03c1\u03c7\u03bf\u03bd \u03c3\u03cd\u03c3\u03c4\u03b7\u03bc\u03b1 \u03ba\u03b1\u03b9 \u03c0\u03c1\u03bf\u03c7\u03c9\u03c1\u03ae\u03c3\u03c4\u03b5 \u03bc\u03cc\u03bd\u03bf \u03cc\u03c4\u03b1\u03bd \u03b7 \u03b2\u03b5\u03bb\u03c4\u03af\u03c9\u03c3\u03b7 \u03c0\u03b1\u03c1\u03b1\u03bc\u03ad\u03bd\u03b5\u03b9 \u03c3\u03c4\u03b1\u03b8\u03b5\u03c1\u03ae \u03c3\u03b5 \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03ac queries.<\/p>\n<\/li>\n<\/ol>\n<\/div>\n<h2 id=\"symperasma\">\u0397 \u03c3\u03c5\u03bd\u03ac\u03c6\u03b5\u03b9\u03b1 \u03b5\u03af\u03bd\u03b1\u03b9 \u03c0\u03c1\u03bf\u03ca\u03cc\u03bd, \u03cc\u03c7\u03b9 \u03bc\u03cc\u03bd\u03bf \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf<\/h2>\n<p>\u03a4\u03bf Sentence Transformers 6.0 \u03bc\u03b5\u03b9\u03ce\u03bd\u03b5\u03b9 \u03c4\u03b7\u03bd \u03c4\u03b5\u03c7\u03bd\u03b9\u03ba\u03ae \u03c4\u03c1\u03b9\u03b2\u03ae \u03b3\u03b9\u03b1 multi-vector training \u03ba\u03b1\u03b9 retrieval. \u0397 \u03c0\u03c1\u03bf\u03c3\u03b2\u03b1\u03c3\u03b9\u03bc\u03cc\u03c4\u03b7\u03c4\u03b1 \u03c4\u03bf\u03c5 \u03b5\u03c1\u03b3\u03b1\u03bb\u03b5\u03af\u03bf\u03c5 \u03cc\u03bc\u03c9\u03c2 \u03b4\u03b5\u03bd \u03b1\u03c6\u03b1\u03b9\u03c1\u03b5\u03af \u03c4\u03b7\u03bd \u03b5\u03c5\u03b8\u03cd\u03bd\u03b7 \u03c4\u03bf\u03c5 \u03c3\u03c7\u03b5\u03b4\u03b9\u03b1\u03c3\u03bc\u03bf\u03cd. \u0397 \u03bf\u03bc\u03ac\u03b4\u03b1 \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03bf\u03c1\u03af\u03c3\u03b5\u03b9 \u03c4\u03b9 \u03c3\u03b7\u03bc\u03b1\u03af\u03bd\u03b5\u03b9 \u03c3\u03c9\u03c3\u03c4\u03ae \u03b1\u03c0\u03ac\u03bd\u03c4\u03b7\u03c3\u03b7, \u03bd\u03b1 \u03b5\u03c0\u03b9\u03bb\u03ad\u03be\u03b5\u03b9 \u03b1\u03bd\u03c4\u03b9\u03c0\u03c1\u03bf\u03c3\u03c9\u03c0\u03b5\u03c5\u03c4\u03b9\u03ba\u03ac \u03b4\u03b5\u03b4\u03bf\u03bc\u03ad\u03bd\u03b1, \u03bd\u03b1 \u03b4\u03b9\u03b1\u03c4\u03b7\u03c1\u03ae\u03c3\u03b5\u03b9 \u03b4\u03cd\u03c3\u03ba\u03bf\u03bb\u03bf evaluation \u03ba\u03b1\u03b9 \u03bd\u03b1 \u03b3\u03bd\u03c9\u03c1\u03af\u03b6\u03b5\u03b9 \u03c0\u03cc\u03c3\u03bf \u03ba\u03bf\u03c3\u03c4\u03af\u03b6\u03b5\u03b9 \u03bf index \u03c3\u03c4\u03b7\u03bd \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03ae \u03c5\u03c0\u03bf\u03b4\u03bf\u03bc\u03ae.<\/p>\n<p>\u0397 \u03af\u03b4\u03b9\u03b1 \u03b1\u03c1\u03c7\u03ae \u03b9\u03c3\u03c7\u03cd\u03b5\u03b9 \u03c3\u03b5 \u03cc\u03bb\u03bf \u03c4\u03bf RAG pipeline. \u0397 <a href=\"https:\/\/twodots.gr\/ai-mnimi-giati-anaktisi-apotygchanei-prin-xekinisei\/\">\u03b1\u03bd\u03ac\u03ba\u03c4\u03b7\u03c3\u03b7 \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03b1\u03c0\u03bf\u03c4\u03cd\u03c7\u03b5\u03b9 \u03c0\u03c1\u03b9\u03bd \u03ba\u03b1\u03bd \u03be\u03b5\u03ba\u03b9\u03bd\u03ae\u03c3\u03b5\u03b9<\/a> \u03cc\u03c4\u03b1\u03bd \u03c4\u03bf \u03c3\u03cd\u03c3\u03c4\u03b7\u03bc\u03b1 \u03b4\u03b5\u03bd \u03bf\u03c1\u03af\u03b6\u03b5\u03b9 \u03c3\u03c9\u03c3\u03c4\u03ac \u03c4\u03b9 \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03b8\u03c5\u03bc\u03ac\u03c4\u03b1\u03b9, \u03c0\u03bf\u03cd \u03bd\u03b1 \u03c8\u03ac\u03be\u03b5\u03b9 \u03ae \u03c0\u03bf\u03b9\u03b1 \u03ad\u03ba\u03b4\u03bf\u03c3\u03b7 \u03bd\u03b1 \u03b5\u03bc\u03c0\u03b9\u03c3\u03c4\u03b5\u03c5\u03c4\u03b5\u03af. \u0388\u03bd\u03b1 \u03ba\u03b1\u03bb\u03cd\u03c4\u03b5\u03c1\u03bf retriever \u03b5\u03bd\u03b9\u03c3\u03c7\u03cd\u03b5\u03b9 \u03bc\u03b9\u03b1 \u03ba\u03b1\u03b8\u03b1\u03c1\u03ae \u03b1\u03c1\u03c7\u03b9\u03c4\u03b5\u03ba\u03c4\u03bf\u03bd\u03b9\u03ba\u03ae \u03b3\u03bd\u03ce\u03c3\u03b7\u03c2\u00b7 \u03b4\u03b5\u03bd \u03b8\u03b5\u03c1\u03b1\u03c0\u03b5\u03cd\u03b5\u03b9 \u03b1\u03c3\u03c5\u03bd\u03ad\u03c0\u03b5\u03b9\u03b1 \u03c0\u03bf\u03c5 \u03c0\u03b1\u03c1\u03b1\u03bc\u03ad\u03bd\u03b5\u03b9 \u03bc\u03ad\u03c3\u03b1 \u03c3\u03c4\u03b1 \u03b4\u03b5\u03b4\u03bf\u03bc\u03ad\u03bd\u03b1.<\/p>\n<p>\u03a4\u03bf \u03ce\u03c1\u03b9\u03bc\u03bf \u03c3\u03c5\u03bc\u03c0\u03ad\u03c1\u03b1\u03c3\u03bc\u03b1 \u03b4\u03b5\u03bd \u03b5\u03af\u03bd\u03b1\u03b9 \u03cc\u03c4\u03b9 \u03ba\u03ac\u03b8\u03b5 \u03b5\u03c0\u03b9\u03c7\u03b5\u03af\u03c1\u03b7\u03c3\u03b7 \u03c7\u03c1\u03b5\u03b9\u03ac\u03b6\u03b5\u03c4\u03b1\u03b9 multi-vector embeddings. \u0395\u03af\u03bd\u03b1\u03b9 \u03cc\u03c4\u03b9 \u03b7 \u03b5\u03c4\u03b1\u03b9\u03c1\u03b9\u03ba\u03ae \u03b1\u03bd\u03b1\u03b6\u03ae\u03c4\u03b7\u03c3\u03b7 \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03b1\u03bd\u03c4\u03b9\u03bc\u03b5\u03c4\u03c9\u03c0\u03af\u03b6\u03b5\u03c4\u03b1\u03b9 \u03c9\u03c2 \u03c0\u03c1\u03bf\u03ca\u03cc\u03bd \u03bc\u03b5 \u03c3\u03c5\u03b3\u03ba\u03b5\u03ba\u03c1\u03b9\u03bc\u03ad\u03bd\u03bf\u03c5\u03c2 \u03c7\u03c1\u03ae\u03c3\u03c4\u03b5\u03c2, relevance definition, baselines, SLOs \u03ba\u03b1\u03b9 \u03c3\u03c5\u03bd\u03b5\u03c7\u03ae \u03b1\u03be\u03b9\u03bf\u03bb\u03cc\u03b3\u03b7\u03c3\u03b7. \u038c\u03c4\u03b1\u03bd \u03b7 \u03bb\u03b5\u03c0\u03c4\u03bf\u03bc\u03ad\u03c1\u03b5\u03b9\u03b1 \u03b5\u03bd\u03cc\u03c2 token \u03b1\u03bb\u03bb\u03ac\u03b6\u03b5\u03b9 \u03c4\u03b7\u03bd \u03b1\u03c0\u03ac\u03bd\u03c4\u03b7\u03c3\u03b7, \u03c4\u03bf late interaction \u03b1\u03be\u03af\u03b6\u03b5\u03b9 \u03bd\u03b1 \u03b4\u03bf\u03ba\u03b9\u03bc\u03b1\u03c3\u03c4\u03b5\u03af. \u038c\u03c4\u03b1\u03bd \u03b4\u03b5\u03bd \u03b1\u03bb\u03bb\u03ac\u03b6\u03b5\u03b9, \u03b7 \u03b1\u03c0\u03bb\u03bf\u03cd\u03c3\u03c4\u03b5\u03c1\u03b7 \u03bb\u03cd\u03c3\u03b7 \u03ba\u03b5\u03c1\u03b4\u03af\u03b6\u03b5\u03b9.<\/p>\n<section class=\"td-service-cta\">\n<div class=\"td-service-cta-content\">\n<p class=\"td-service-cta-kicker\">RAG \u03ba\u03b1\u03b9 \u03b5\u03c4\u03b1\u03b9\u03c1\u03b9\u03ba\u03ae \u03b1\u03bd\u03b1\u03b6\u03ae\u03c4\u03b7\u03c3\u03b7 \u03bc\u03b5 \u03bc\u03b5\u03c4\u03c1\u03ae\u03c3\u03b9\u03bc\u03b7 \u03b1\u03be\u03af\u03b1<\/p>\n<p class=\"td-service-cta-title\">\u03a3\u03c7\u03b5\u03b4\u03b9\u03ac\u03c3\u03c4\u03b5 retrieval \u03b3\u03cd\u03c1\u03c9 \u03b1\u03c0\u03cc \u03c4\u03b1 \u03b4\u03b9\u03ba\u03ac \u03c3\u03b1\u03c2 documents \u03ba\u03b1\u03b9 queries<\/p>\n<p>\u0397 TWO DOTS \u03c7\u03b1\u03c1\u03c4\u03bf\u03b3\u03c1\u03b1\u03c6\u03b5\u03af corpus, relevance labels, baselines, data flows \u03ba\u03b1\u03b9 \u03bb\u03b5\u03b9\u03c4\u03bf\u03c5\u03c1\u03b3\u03b9\u03ba\u03ac \u03cc\u03c1\u03b9\u03b1 \u03c0\u03c1\u03b9\u03bd \u03c3\u03c5\u03bd\u03b4\u03ad\u03c3\u03b5\u03b9 \u03b5\u03c4\u03b1\u03b9\u03c1\u03b9\u03ba\u03ae \u03b3\u03bd\u03ce\u03c3\u03b7 \u03bc\u03b5 AI assistants \u03ae \u03b1\u03c5\u03c4\u03bf\u03bc\u03b1\u03c4\u03b9\u03c3\u03bc\u03bf\u03cd\u03c2. \u0388\u03c4\u03c3\u03b9 \u03b7 \u03b5\u03c0\u03b9\u03bb\u03bf\u03b3\u03ae dense, hybrid \u03ae multi-vector retrieval \u03b2\u03b1\u03c3\u03af\u03b6\u03b5\u03c4\u03b1\u03b9 \u03c3\u03b5 \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03ac search failures \u03ba\u03b1\u03b9 \u03cc\u03c7\u03b9 \u03c3\u03b5 \u03ad\u03bd\u03b1 \u03b3\u03b5\u03bd\u03b9\u03ba\u03cc leaderboard.<\/p>\n<div class=\"td-service-cta-actions\"><a class=\"td-service-cta-button\" href=\"https:\/\/twodots.gr\/aftomatismoi-epicheiriseon-ai\/\">\u0394\u03b5\u03af\u03c4\u03b5 \u0391\u03c5\u03c4\u03bf\u03bc\u03b1\u03c4\u03b9\u03c3\u03bc\u03bf\u03cd\u03c2 \u0395\u03c0\u03b9\u03c7\u03b5\u03b9\u03c1\u03ae\u03c3\u03b5\u03c9\u03bd &amp; AI<\/a><\/div>\n<\/div>\n<\/section>\n<section id=\"sychnes-erotiseis\" class=\"td-faq-section\">\n<div class=\"td-faq\">\n<p class=\"td-faq-heading\">\u0427\u0435\u0441\u0442\u043e \u0437\u0430\u0434\u0430\u0432\u0430\u043d\u0438 \u0432\u044a\u043f\u0440\u043e\u0441\u0438<\/p>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">\u03a4\u03b9 \u03b5\u03af\u03bd\u03b1\u03b9 \u03c4\u03b1 multi-vector embeddings;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u0395\u03af\u03bd\u03b1\u03b9 \u03b1\u03bd\u03b1\u03c0\u03b1\u03c1\u03b1\u03c3\u03c4\u03ac\u03c3\u03b5\u03b9\u03c2 \u03c0\u03bf\u03c5 \u03ba\u03c1\u03b1\u03c4\u03bf\u03cd\u03bd \u03c3\u03c5\u03bd\u03ae\u03b8\u03c9\u03c2 \u03ad\u03bd\u03b1 vector \u03b1\u03bd\u03ac token \u03ba\u03b1\u03b9 \u03c3\u03c5\u03b3\u03ba\u03c1\u03af\u03bd\u03bf\u03c5\u03bd query \u03ba\u03b1\u03b9 document \u03bc\u03b5 late interaction, \u03b1\u03bd\u03c4\u03af \u03bd\u03b1 \u03c3\u03c5\u03bc\u03c0\u03b9\u03ad\u03b6\u03bf\u03c5\u03bd \u03bf\u03bb\u03cc\u03ba\u03bb\u03b7\u03c1\u03bf \u03c4\u03bf \u03ba\u03b5\u03af\u03bc\u03b5\u03bd\u03bf \u03c3\u03b5 \u03ad\u03bd\u03b1 \u03bc\u03cc\u03bd\u03bf vector.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">\u03a0\u03bf\u03cd \u03b4\u03b9\u03b1\u03c6\u03ad\u03c1\u03bf\u03c5\u03bd \u03b1\u03c0\u03cc \u03c4\u03b1 dense embeddings;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u03a4\u03b1 dense embeddings \u03c7\u03c1\u03b7\u03c3\u03b9\u03bc\u03bf\u03c0\u03bf\u03b9\u03bf\u03cd\u03bd \u03c3\u03c5\u03bd\u03ae\u03b8\u03c9\u03c2 \u03ad\u03bd\u03b1 vector \u03b1\u03bd\u03ac \u03ba\u03b5\u03af\u03bc\u03b5\u03bd\u03bf \u03ba\u03b1\u03b9 \u03bc\u03b9\u03ba\u03c1\u03cc\u03c4\u03b5\u03c1\u03bf index. \u03a4\u03b1 multi-vector embeddings \u03b4\u03b9\u03b1\u03c4\u03b7\u03c1\u03bf\u03cd\u03bd \u03bb\u03b5\u03c0\u03c4\u03bf\u03bc\u03b5\u03c1\u03ad\u03c3\u03c4\u03b5\u03c1\u03b1 token-level signals, \u03b1\u03bb\u03bb\u03ac \u03b1\u03c0\u03b1\u03b9\u03c4\u03bf\u03cd\u03bd \u03c0\u03b5\u03c1\u03b9\u03c3\u03c3\u03cc\u03c4\u03b5\u03c1\u03b7 \u03b1\u03c0\u03bf\u03b8\u03ae\u03ba\u03b5\u03c5\u03c3\u03b7 \u03ba\u03b1\u03b9 \u03c3\u03cd\u03bd\u03b8\u03b5\u03c4\u03bf serving.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">\u0393\u03b9\u03b1\u03c4\u03af \u03b5\u03af\u03bd\u03b1\u03b9 \u03c3\u03b7\u03bc\u03b1\u03bd\u03c4\u03b9\u03ba\u03cc \u03c4\u03bf domain finetuning;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u0395\u03c0\u03b5\u03b9\u03b4\u03ae \u03bc\u03b1\u03b8\u03b1\u03af\u03bd\u03b5\u03b9 \u03c4\u03b7 \u03b4\u03b9\u03ba\u03ae \u03c3\u03b1\u03c2 \u03ad\u03bd\u03bd\u03bf\u03b9\u03b1 \u03c4\u03b7\u03c2 \u03c3\u03c5\u03bd\u03ac\u03c6\u03b5\u03b9\u03b1\u03c2 \u03b1\u03c0\u03cc \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03ac query\u2013document pairs, \u03bf\u03c1\u03bf\u03bb\u03bf\u03b3\u03af\u03b1 \u03ba\u03b1\u03b9 patterns \u03c7\u03c1\u03b7\u03c3\u03c4\u03ce\u03bd, \u03b1\u03bd\u03c4\u03af \u03bd\u03b1 \u03b2\u03b1\u03c3\u03af\u03b6\u03b5\u03c4\u03b1\u03b9 \u03bc\u03cc\u03bd\u03bf \u03c3\u03b5 \u03b3\u03b5\u03bd\u03b9\u03ba\u03ae \u03c3\u03b7\u03bc\u03b1\u03c3\u03b9\u03bf\u03bb\u03bf\u03b3\u03b9\u03ba\u03ae \u03bf\u03bc\u03bf\u03b9\u03cc\u03c4\u03b7\u03c4\u03b1.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">\u03a4\u03b9 \u03c1\u03cc\u03bb\u03bf \u03c0\u03b1\u03af\u03b6\u03b5\u03b9 \u03c4\u03bf document truncation;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u039a\u03ac\u03b8\u03b5 \u03c0\u03bb\u03b7\u03c1\u03bf\u03c6\u03bf\u03c1\u03af\u03b1 \u03bc\u03b5\u03c4\u03ac \u03c4\u03bf length cap \u03bc\u03ad\u03bd\u03b5\u03b9 \u03b1\u03cc\u03c1\u03b1\u03c4\u03b7 \u03c3\u03c4\u03bf retrieval. \u03a3\u03c4\u03bf \u03c3\u03c5\u03b3\u03ba\u03b5\u03ba\u03c1\u03b9\u03bc\u03ad\u03bd\u03bf MIRIAD experiment, \u03b7 \u03c0\u03b7\u03b3\u03ae \u03bc\u03ad\u03c4\u03c1\u03b7\u03c3\u03b5 \u03b1\u03c0\u03ce\u03bb\u03b5\u03b9\u03b1 \u03ad\u03c9\u03c2 0,24 NDCG@10 \u03b1\u03c0\u03cc truncation \u03c3\u03b5 multi-vector \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03b1.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">\u03a7\u03c1\u03b5\u03b9\u03ac\u03b6\u03bf\u03bd\u03c4\u03b1\u03b9 \u03ad\u03bd\u03b1 \u03b5\u03ba\u03b1\u03c4\u03bf\u03bc\u03bc\u03cd\u03c1\u03b9\u03bf training pairs;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u038c\u03c7\u03b9 \u03c5\u03c0\u03bf\u03c7\u03c1\u03b5\u03c9\u03c4\u03b9\u03ba\u03ac. \u03a3\u03c4\u03b7 \u03c3\u03c5\u03b3\u03ba\u03b5\u03ba\u03c1\u03b9\u03bc\u03ad\u03bd\u03b7 \u03b4\u03bf\u03ba\u03b9\u03bc\u03ae, 100.000 pairs \u03ba\u03b1\u03b9 75 \u03bb\u03b5\u03c0\u03c4\u03ac training \u03b2\u03c1\u03ad\u03b8\u03b7\u03ba\u03b1\u03bd \u03b5\u03bd\u03c4\u03cc\u03c2 0,012 NDCG@10 \u03b1\u03c0\u03cc \u03c4\u03bf \u03c0\u03bb\u03ae\u03c1\u03b5\u03c2 run, \u03b1\u03bb\u03bb\u03ac \u03ba\u03ac\u03b8\u03b5 domain \u03c7\u03c1\u03b5\u03b9\u03ac\u03b6\u03b5\u03c4\u03b1\u03b9 \u03b4\u03b9\u03ba\u03cc \u03c4\u03bf\u03c5 scaling experiment.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">\u0393\u03b9\u03b1\u03c4\u03af \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03ba\u03c1\u03b1\u03c4\u03ae\u03c3\u03c9 \u03c4\u03bf BM25 \u03c9\u03c2 baseline;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u0395\u03af\u03bd\u03b1\u03b9 \u03c6\u03b8\u03b7\u03bd\u03cc, \u03b4\u03b9\u03b1\u03b2\u03ac\u03b6\u03b5\u03b9 \u03cc\u03bb\u03bf \u03c4\u03bf \u03ba\u03b5\u03af\u03bc\u03b5\u03bd\u03bf \u03ba\u03b1\u03b9 \u03c0\u03b1\u03c1\u03b1\u03bc\u03ad\u03bd\u03b5\u03b9 \u03b9\u03c3\u03c7\u03c5\u03c1\u03cc \u03cc\u03c4\u03b1\u03bd \u03c5\u03c0\u03ac\u03c1\u03c7\u03b5\u03b9 \u03c5\u03c8\u03b7\u03bb\u03cc lexical overlap. \u03a7\u03c9\u03c1\u03af\u03c2 \u03b1\u03c5\u03c4\u03cc \u03b4\u03b5\u03bd \u03b3\u03bd\u03c9\u03c1\u03af\u03b6\u03b5\u03c4\u03b5 \u03b1\u03bd \u03c4\u03bf \u03bd\u03ad\u03bf retrieval stack \u03b4\u03b9\u03ba\u03b1\u03b9\u03bf\u03bb\u03bf\u03b3\u03b5\u03af \u03c4\u03b7\u03bd \u03b5\u03c0\u03b9\u03c0\u03bb\u03ad\u03bf\u03bd \u03c0\u03bf\u03bb\u03c5\u03c0\u03bb\u03bf\u03ba\u03cc\u03c4\u03b7\u03c4\u03b1.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">\u03a0\u03cc\u03c3\u03bf \u03bc\u03b5\u03b3\u03ac\u03bb\u03bf\u03c2 \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03b3\u03af\u03bd\u03b5\u03b9 \u03bf multi-vector index;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u0395\u03be\u03b1\u03c1\u03c4\u03ac\u03c4\u03b1\u03b9 \u03ba\u03c5\u03c1\u03af\u03c9\u03c2 \u03b1\u03c0\u03cc \u03c4\u03bf \u03bc\u03ae\u03ba\u03bf\u03c2 \u03c4\u03c9\u03bd documents \u03ba\u03b1\u03b9 \u03c4\u03b7 \u03c3\u03c5\u03bc\u03c0\u03af\u03b5\u03c3\u03b7. \u03a3\u03c4\u03bf \u03c0\u03b5\u03af\u03c1\u03b1\u03bc\u03b1 \u03ae\u03c4\u03b1\u03bd \u03c0\u03b5\u03c1\u03af\u03c0\u03bf\u03c5 45 GB raw FP16 \u03ba\u03b1\u03b9 3,37 GB \u03bc\u03b5 1-bit PLAID \u03b3\u03b9\u03b1 \u03cc\u03bb\u03b1 \u03c4\u03b1 vectors.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">\u03a0\u03cc\u03c4\u03b5 \u03b1\u03be\u03af\u03b6\u03b5\u03b9 \u03b3\u03b9\u03b1 \u03b5\u03c4\u03b1\u03b9\u03c1\u03b9\u03ba\u03cc RAG;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u038c\u03c4\u03b1\u03bd \u03bc\u03b5\u03b3\u03ac\u03bb\u03b1 \u03ae \u03b5\u03be\u03b5\u03b9\u03b4\u03b9\u03ba\u03b5\u03c5\u03bc\u03ad\u03bd\u03b1 documents \u03c7\u03ac\u03bd\u03bf\u03c5\u03bd \u03ba\u03c1\u03af\u03c3\u03b9\u03bc\u03b5\u03c2 \u03bb\u03b5\u03c0\u03c4\u03bf\u03bc\u03ad\u03c1\u03b5\u03b9\u03b5\u03c2 \u03bc\u03b5 dense \u03ae lexical retrieval \u03ba\u03b1\u03b9 \u03b7 \u03bc\u03b5\u03c4\u03c1\u03b7\u03bc\u03ad\u03bd\u03b7 \u03ac\u03bd\u03bf\u03b4\u03bf\u03c2 \u03c0\u03bf\u03b9\u03cc\u03c4\u03b7\u03c4\u03b1\u03c2 \u03b4\u03b9\u03ba\u03b1\u03b9\u03bf\u03bb\u03bf\u03b3\u03b5\u03af storage, latency \u03ba\u03b1\u03b9 \u03bb\u03b5\u03b9\u03c4\u03bf\u03c5\u03c1\u03b3\u03b9\u03ba\u03cc \u03ba\u03cc\u03c3\u03c4\u03bf\u03c2.<\/p>\n<\/div>\n<\/details>\n<\/div>\n<\/section>\n<div class=\"td-source-list\">\n<p id=\"piges\" class=\"td-source-list-title\">\u0418\u0437\u0442\u043e\u0447\u043d\u0438\u0446\u0438<\/p>\n<ul>\n<li><a href=\"https:\/\/huggingface.co\/blog\/train-multi-vector-encoder\" target=\"_blank\" rel=\"noopener\">Hugging Face \u2014 Training and Finetuning Multi-Vector Embedding Models with Sentence Transformers<\/a><\/li>\n<li><a href=\"https:\/\/huggingface.co\/blog\/multi-vector-encoder\" target=\"_blank\" rel=\"noopener\">Hugging Face \u2014 Multi-Vector (Late Interaction) Embedding Models with Sentence Transformers<\/a><\/li>\n<li><a href=\"https:\/\/www.sbert.net\/docs\/multi_vector_encoder\/training_overview.html\" target=\"_blank\" rel=\"noopener\">Sentence Transformers \u2014 Multi-Vector Encoder Training Overview<\/a><\/li>\n<li><a href=\"https:\/\/www.sbert.net\/docs\/migration_guide.html\" target=\"_blank\" rel=\"noopener\">Sentence Transformers \u2014 Migration Guide for v6<\/a><\/li>\n<li><a href=\"https:\/\/arxiv.org\/abs\/2506.06091\" target=\"_blank\" rel=\"noopener\">MIRIAD \u2014 Augmenting LLMs with millions of medical query-response pairs<\/a><\/li>\n<li><a href=\"https:\/\/github.com\/lightonai\/fast-plaid\" target=\"_blank\" rel=\"noopener\">LightOn \u2014 fast-plaid, high-performance multi-vector search engine<\/a><\/li>\n<\/ul>\n<\/div>","protected":false},"excerpt":{"rendered":"<p>\u03a4\u03b1 multi-vector embeddings \u03b2\u03b5\u03bb\u03c4\u03b9\u03ce\u03bd\u03bf\u03c5\u03bd \u03c4\u03b7\u03bd \u03b5\u03c4\u03b1\u03b9\u03c1\u03b9\u03ba\u03ae \u03b1\u03bd\u03b1\u03b6\u03ae\u03c4\u03b7\u03c3\u03b7 \u03cc\u03c4\u03b1\u03bd domain data, \u03bc\u03ae\u03ba\u03bf\u03c2 \u03b5\u03b3\u03b3\u03c1\u03ac\u03c6\u03c9\u03bd, evaluation \u03ba\u03b1\u03b9 index \u03bc\u03b5\u03c4\u03c1\u03ce\u03bd\u03c4\u03b1\u03b9 \u03bc\u03b1\u03b6\u03af \u03c3\u03b5 \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03ac queries.<\/p>","protected":false},"author":1,"featured_media":105753,"comment_status":"open","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"_gspb_post_css":"","content-type":"","footnotes":""},"categories":[199],"tags":[7477,20807,8233,19677,20808],"class_list":["post-98135","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-techniti-noimosyni","tag-machine-learning","tag-multi-vector-embeddings","tag-rag","tag-semantic-search","tag-sentence-transformers"],"blocksy_meta":{"styles_descriptor":{"styles":{"desktop":"","tablet":"","mobile":""},"google_fonts":[],"version":8}},"_links":{"self":[{"href":"https:\/\/twodots.gr\/bg\/wp-json\/wp\/v2\/posts\/98135","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/twodots.gr\/bg\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/twodots.gr\/bg\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/twodots.gr\/bg\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/twodots.gr\/bg\/wp-json\/wp\/v2\/comments?post=98135"}],"version-history":[{"count":1,"href":"https:\/\/twodots.gr\/bg\/wp-json\/wp\/v2\/posts\/98135\/revisions"}],"predecessor-version":[{"id":105754,"href":"https:\/\/twodots.gr\/bg\/wp-json\/wp\/v2\/posts\/98135\/revisions\/105754"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/twodots.gr\/bg\/wp-json\/wp\/v2\/media\/105753"}],"wp:attachment":[{"href":"https:\/\/twodots.gr\/bg\/wp-json\/wp\/v2\/media?parent=98135"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/twodots.gr\/bg\/wp-json\/wp\/v2\/categories?post=98135"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/twodots.gr\/bg\/wp-json\/wp\/v2\/tags?post=98135"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}