C\u00e2nd oamenii \u00eencep s\u0103 arunce termeni precum \u201eRAG\u201d \u0219i \u201evector retrieval\u201d \u00een AI, majoritatea oamenilor normali simt nevoia imediat\u0103 s\u0103 intrebe ce inseamna acesti termeni.<\/p>\n\n\n\n
Sun\u0103 fantezist. Este vorba doar despre a da unui model mare de inteligen\u021b\u0103 artificial\u0103 (OpenAI, Claude, Llama, Deepseek, etc) o bibliografie (documentatie), \u00eenainte de a r\u0103spunde la \u00eentrebarea ta. Dac\u0103 cineva v\u0103 \u00eentreab\u0103: \u201eCine a fost al 3-lea pre\u0219edinte al Romaniei?\u201d<\/p>\n\n\n\n
Asta este RAG:<\/p>\n\n\n\n
Iat\u0103 secretul modelelor mari de inteligen\u021b\u0103 artificial\u0103 (chiar \u0219i GPT-4, Claude etc.): de fapt nu \u201e\u0219tiu\u201d totul. Modelele au o memorie fix\u0103 (pe care au fost antrenate). Dac\u0103 \u00eei \u00eentreba\u021bi despre noile politici ale companiei dvs., notele nepublicate sau orice altceva de ni\u0219\u0103 \u0219au privat, vor spune:<\/p>\n\n\n\n
\u201e\u00cemi pare r\u0103u, nu am auzit niciodat\u0103 de asta, dar iat\u0103 oricum ni\u0219te prostii care sun\u0103 a fi de incredere, dar sunt spuse la \u00eent\u00e2mplare.\u201d<\/em><\/p>\n\n\n\n \u00cen cazul RAG, in loc s\u0103 halucineze, inteligen\u021ba artificial\u0103 preia informa\u021bii reale \u0219i externe \u00eenainte de a v\u0103 r\u0103spunde. G\u00e2ndi\u021bi-v\u0103 la RAG ca la o inteligen\u021b\u0103 artificial\u0103 care \u00ee\u0219i pune ochelari de citit \u0219i verific\u0103 de dou\u0103 ori faptele.<\/p>\n\n\n\n C\u00e2nd AI \u00eencearc\u0103 s\u0103 \u201erecupereze\u201d informa\u021bii utile din bibiografie sau documentatia privata, de unde \u0219tie ce este relevant?<\/p>\n\n\n\n Embeddings = Transformarea lucrurilor \u00een liste de numere care surprind sensul<\/p>\n\n\n\n Ei bine, toate documentele pe care dori\u021bi s\u0103 le c\u0103uta\u021bi sunt \u00eencorporate \u00een vectori, mai pe romaneste spus, transformate in numere. Astfel, c\u00e2nd pui o \u00eentrebare, interogarea ta este transformat\u0103 \u00eentr-un vector (numar). Apoi sistemul caut\u0103 cei mai apropia\u021bi vectori- adic\u0103 cele mai relevante informa\u021bii, chiar dac\u0103 formularea este diferit\u0103.<\/p>\n\n\n\n Daca intrebi \u201eCine a condus Romania dup\u0103 Traian Basescu?<\/p>\n\n\n\n Vector retrieval g\u0103se\u0219te documente despre \u201eKlaus Johanis\u201d f\u0103r\u0103 a fi nevoie s\u0103-l numi\u021bi.<\/em><\/em><\/p>\n\n\n\n F\u0103r\u0103 cuvinte cheie, f\u0103r\u0103 magie- doar potrivire de numere bazat\u0103 pe matematic\u0103.<\/p>\n\n\n\n Iat\u0103 fluxul complet al unui sistem RAG folosind recuperarea vectorilor:<\/p>\n\n\n\n Nu este necesar\u0103 reinstruirea modelelor AI: nu trebuie s\u0103 antrenati un model AI de fiecare dat\u0103 c\u00e2nd actualiza\u021bi documentatia. Doar actualiza\u021bi baza de date de documente. (Mult mai ieftin \u0219i mai rapid.)<\/p>\n\n\n\n Informa\u021bii actualizate: AI poate face referire la documente noi, chiar dac\u0103 acestea nu existau atunci c\u00e2nd modelul a fost antrenat.<\/p>\n\n\n\n AI mai de \u00eencredere: Pute\u021bi urm\u0103ri r\u0103spunsurile \u00eenapoi la documentele surs\u0103 \u00een loc s\u0103 ob\u021bine\u021bi halucinatii \u0219i presupuneri.<\/p>\n\n\n\n Personalizare: Vrei o inteligen\u021b\u0103 artificial\u0103 care \u00ee\u021bi \u201ecunoa\u0219te\u201d afacerea de ni\u0219\u0103 , catalogul de produse, documentele legale sau arhiva? Doar hr\u0103ne\u0219te-l cu documentele potrivite.<\/p>\n\n\n\n sursa: mersi.ai https:\/\/mersi.ai<\/a><\/p>\n","protected":false},"excerpt":{"rendered":" RAG (Retrival Augmentet Generation) bazat pe vectori C\u00e2nd oamenii \u00eencep s\u0103 arunce termeni precum \u201eRAG\u201d \u0219i \u201evector retrieval\u201d \u00een AI, majoritatea oamenilor normali simt nevoia imediat\u0103 s\u0103 intrebe ce inseamna acesti termeni. \u00cen primul r\u00e2nd, ce este \u201eRAG\u201d? RAG = Generare Augmentata de Recuperare. Sun\u0103 fantezist. Este vorba doar despre a da unui model mare\u2026 Read More »<\/a><\/span><\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":[],"categories":[9],"tags":[167,166,168,5,169],"_links":{"self":[{"href":"https:\/\/www.hallocal.com\/articles\/wp-json\/wp\/v2\/posts\/156"}],"collection":[{"href":"https:\/\/www.hallocal.com\/articles\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.hallocal.com\/articles\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.hallocal.com\/articles\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.hallocal.com\/articles\/wp-json\/wp\/v2\/comments?post=156"}],"version-history":[{"count":1,"href":"https:\/\/www.hallocal.com\/articles\/wp-json\/wp\/v2\/posts\/156\/revisions"}],"predecessor-version":[{"id":157,"href":"https:\/\/www.hallocal.com\/articles\/wp-json\/wp\/v2\/posts\/156\/revisions\/157"}],"wp:attachment":[{"href":"https:\/\/www.hallocal.com\/articles\/wp-json\/wp\/v2\/media?parent=156"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.hallocal.com\/articles\/wp-json\/wp\/v2\/categories?post=156"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.hallocal.com\/articles\/wp-json\/wp\/v2\/tags?post=156"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}Remedieri RAG<\/h2>\n\n\n\n
<\/h2>\n\n\n\n
Unde se \u00eencadreaz\u0103 Vector retrieval (recuperarea vectorilor)?<\/h2>\n\n\n\n
\n
Ce inseamna embeddings (\u00eencorporari) ?<\/h2>\n\n\n\n
Pun\u00e2nd totul \u00eempreun\u0103<\/h1>\n\n\n\n
\n
\n
Avantajele uria\u0219e ale RAG + Vector Retrieval<\/h1>\n\n\n\n