I-Big Language Models (LLMs) iyatholakala kakhulu embonini yekhompyutha namhlanje. Zihlanganisa izihloko ezintsha, izihloko zebhulogi, izifundo kanye nemodeli eziholela amakhompyutha ethu emkhakheni, njenge-Meta, Huggingface, Microsoft, njll, okuyinto kuthatha thina ukuhlola ubuchwepheshe ezintsha. Thina siphinde ushiye amaphepha amancane, okubhaliweyo ukuze ufake lezi zihlanganisi kanye nokuvakashela ubuchwepheshe ezintsha. I-theme yokuqala etholakalayo kuyoba i-RAG (Retrieval-Augmented Generation). Thina ukwakha a series of articles on the topic we have determined, nge amaphepha ezintathu ahlukahlukene ezisebenzayo futhi zihlanganisa. Kulesi nqaku, thina kuqala inqaku lethu nge definitions and basic information of RAG models. Imodeli ye-Language ye-Big has wahlanganyela zonke izicathulo zokusebenza kwethu. Singatholakala ukuthi ziye zibonise indawo. Kodwa-ke, akuyona isixhobo enhle kakhulu njengokufundisa. It has futhi isizukulwane esikhulu: Kuyinto enhle ekufundeni yayo. Kuyinto enhle idatha esifundiswe ku. A model enikezela ukuqeqeshwa ngoNovemba 2022 ngeke akwazi ukufundisa izindaba, izinhlelo zokusebenza kwezobuchwepheshe njll, okwakhiwa ngoJanuwari 2023. Ngokwesibonelo, imodeli ye-LLM enikezela ukuqeqeshwa futhi wahlanganyela inkonzo ngo-2021. A model enikezela imibuzo mayelana ne-Russia-Ukraine war eqala ngoFebruwari 24, 2022. Kuyinto ngenxa yokuthuthukiswa yayo yasungulwa ngaphambi kwelinye isikhathi. Ngokuvamile, le nkinga akuyona engatholakali, futhi imikhiqizo entsha, uhlelo entsha, yasungulwa. I-RAG (Retrieval-Augmented Generation) uhlelo lithunyelwe ukunikezela ulwazi olusheshayo lapho ufuna. Kwi-mali lwethu, siphindezisa ngokushesha kanye nenkqubo eyenziwe nge-LLM model kanye ne-RAG system, eyodwa ngamunye, ukuze ufunde kubo. Ukufinyelela ulwazi ku-LLM Models Umgangatho yokusebenza we-Big Language Models isekelwe idatha ebonakalayo ngesikhathi sokucwaninga, okungenani, ulwazi olungaphakathi. Abanikezele ukufinyelela idatha ephakeme ngokusebenzisa ezinye izindlela. Ukuze ukunikeza isibonelo, ukuthatha umdlavuza. Uma siqhathanisa ukuxhumana okuzenzakalelayo kwelinye umdlavuza kanye nokufundisa kwabo kuphela e-English, akwazi ukuxhumana igama elilodwa le-Chinese kusuka kumadoda. Kuyinto ngoba sinikeza umfanekiso owenziwe e-English, ngaphandle kwe-Chinese. Ngokunciphisa ukuxhumanisa kwelanga elandelayo, sinikeza futhi umthamo yayo yokufunda emithonjeni ezingaphandle. Ngokusho umfanekiso le-LLM, amamodeli we-LLM zihlanganiswa nezifundo eziyisisekelo kodwa zihlanganisa idatha ezingezansi. Umbala olunye we-LLM amamodeli kuyinto ibhokisi omnyama. Lezi amamodeli akufanele ngokugcwele ukuthi kungenza imizamo etholakalayo. Zihlanganisa izilinganiso zayo kuphela kumemodeli yemathematiki. Ukubiza noma umodeli we-LLM, "Ukuba uthetha lokhu?" Uyakwazi ukucindezela kakhulu. Abanikezela imibuzo ngokusebenzisa ukubuyekezwa noma ucwaningo. I-keyword apha yi-"why?" Singatholakala futhi njenge-bug e-LLM amamodeli. Ukuze uthole okuphumelela okungcono le isakhiwo, bheka isibonelo esekelwe emkhakheni yokwelapha. Uma umsebenzisi uthetha, "Ngingathanda isifo emzimbeni kanye nemibala, kanye ne-post-nasal dropling eside. Yini kufanele uyenze?", imodeli ye-LLM ingatholela, "Umbala emzimbeni kanye nemibala, kanye ne-post-nasal dropling kungabangela i-sinusitis. Nceda uqhagamshelane nomdlavuza. I-acute sinusitis isetshenziswe nge-antibiotics. Ngaphandle kwe-medication, ungasebenzisa ama-nasal sprays njenge-sea water noma i-saline ukunciphisa i-sinus." Konke kubonakala ngokuvamile kuze kube manje. Kodwa uma siphinde imodeli, "Why did you give that answer?" ngemva lokhu impendulo, izimo zihlanganisa. Umzekelo imodeli iboniswe lokhu impendulo kuyinto ngoba izici "umdlavuza wesibeletho" futhi "ukudlulisela naso" ngokuvamile zihlanganisa nge- word "sinusitis" ku-data ukuqeqeshwa. Lezi amamodeli akufanele ukugcina ulwazi ememori yayo njengoba isampula sokuvamile. Imibuzo yemathemathemikhali kubalulekile kumodeli ye-LLM. Njengoba akufanele ukugcina ulwazi emoyeni yayo ngokusekelwe emakhasini, inikeza imibuzo futhi inikeza abantu abaninzi ngokushesha, kodwa uma abantu abanolwazi kakhulu abanolwazi akufundisa imodeli, "Ukuye wabhala lokhu?", imodeli akufuna ukukhiqiza inkinobho yokuxhumana. Ngingathanda ukuthi sibe nokufuna okuphakeme kwama-LLM amamodeli. Ngoku, singathola uhlelo le-RAG kanye nemisombululo yayo kwezi zimo. RAG: Ukuhlanganiswa kwe-LLM ne-Retrieval Systems I-RAG Systems inikeza izinzuzo kunezinhlelo ezisekelwe kumamodeli ye-LLM enhle. Enye kuhlanganise ukuthi zihlanganisa ulwazi olusebenzayo, akuyona ulwazi olusekelwe njenge-LLM models. Ngokuvamile, zihlanganisa izinhlelo ze-external ngaphandle kokufaka ku-data eyenziwa ku-trained. I-r yokuqala e-RAG inikeza ukubuyekeza. I-role of the retrieval component kuyinto ukubuyekeza imisebenzi yokufaka. I-generators iyinhlangano yesibili esikhulu, futhi umsebenzi yayo kuyinto ukukhiqiza impendulo esiyingqayizivele esekelwe ku-data retrieval returns. Kule nqaku, siza kubhalwe ngokushesha ku-RAG isisombululo sokusebenza: I-Retrieval i-scans idatha evela kumadivayisi ezingenalutho futhi i-recovering idokhumenti esiyingqayizivele kwebhizinisi lomsebenzisi ngokuhambisa kwama-pieces amancane ebizwa ngokuthi "chunks." Yenza i-vectorize lezi zihlanganisa kanye ne-user's question, bese ivela inguqulo enhle ngokuvamile ngokuhambisana phakathi kwabo. Lesi umsebenzi yokwenza inguqulo iyatholakala yi-"Generation." Sitholela ngokuphathelene kulesi sihloko elilandelayo kulesi sethu. Lezi zihlanganisi zokusebenza kwezinhlelo zokusebenza ze-RAG zenza isakhiwo esihle kakhulu. Lezi zihlanganisa ukuthi zihlanganisa ulwazi se-static njenge-LLM amamodeli. Imodeli ye-RAG inikeza izinzuzo emikhulu, ikakhulukazi eziningana nezinkampani ezininzi ezidingekayo zebhizinisi. Ngokwesibonelo, ungahambisa ukwenza imodeli ye-doktor e-medical. Njengoba imodeli yakho uzobonakalisa indawo enhle, akuyona ulwazi oluthile noma olungaphakathi. I-medical world, njenge-IT sector, ikhiqizwa ngosuku zonke, futhi izifundo ezintsha zithunyelwe. Ngakho-ke, imodeli yakho iyatholakala ukwamukela ngisho izifundo ezintsha. Ngaphandle kwalokho, unemibuzo yokuvimbela ubomi womuntu nge-model ebuthile. Kulezi zimo, izinhlelo ezisekelwe yi-RAG zithunyelwe inkinga ulwazi oluthile ngokuhambisana nezithombe ezingenalutho. I-different fundamental phakathi kwezinhlelo zokusebenza ze-RAG kanye nezinhlelo zokusebenza ze-LLM kuyinto i-RAG core philosophy: "Hlola ulwazi, ukufinyelela lapho ufuna!" Nangona ama-small-language amamodeli zihlanganisa ulwazi emangalisayo kanye nokukhiqiza imibuzo ngemuva kokufundwa, izinhlelo zokusebenza ze-RAG ukufinyelela ulwazi ngokufunda futhi ukucindezeleka ngaphandle lapho zinobunzima, ngokuvumelana ne-philosophy yabo. Njengoba umntu uthola inthanethi, lokhu kuvimbela enye ingxaki kakhulu ye-LLM amamodeli amancane: ukunakekelwa kwememori yayo. Ukuze uthole okwengeziwe ukubonisa point yethu, singakwazi ukuguqulwa ngezinhlelo ezimbili. 
 
 
 
 
 
 Umhlahlandlela: Usayizi: "Ukuhlwa kwe-Inflation ye-United States ngo-December 2024?" LLM: "Ngokusho idatha ngoDisemba we-2022, kwaba i-6.5%." (Hhayi impendulo olusheshayo) Waze: 
 
 
 Retrieves December 2024 data from a reliable source or database (World Bank, Trading Economics, etc.). LLM uses this data and responds, "According to Trading Economics, inflation in the United States for December 2024 is announced as 2.9%." Scenario: User: "What is the United States' December 2024 inflation rate?" LLM: "According to December 2022 data, it was 6.5%." (Not an up-to-date answer) RAG: 
 
 
 Yenza idatha ngo-December 2024 kusuka ku-source noma database enhle (i-World Bank, i-Trade Economics, njll). I-LLM isetshenziselwa idatha etholakalayo, "Ngokusho i-Trade Economics, isivinini se-Inflation e-United States ngo-December 2024 iyatholakala ku-2.9%." Sicela ukuguqulwa ngokushesha okuhlobene nokufaka ku-akhawunti ye-imeyili elilandelayo. 
 
 
 
 
 
 
 
 
 
 Features 
 LLM (static model) 
 RAG (retrieval-augmented generation) 
 
 
 
 
 Information 
 Limited to training data 
 Can pull real-time information from external sources 
 
 
 
 
 Current Level 
 Low 
 High 
 
 
 
 
 Transparency 
 The source of the decision cannot be disclosed (black box) 
 The source can be cited Ukucaciswa Imininingwane ku-Training Data Ungathola ulwazi isikhathi real kusuka ku-sources external Isikhathi eside Ngaphansi Ukuhlehlela Ukuhlobisa Umthombo we-decision ayikho (Black Box) Isilinganiso kungenziwa Ngokuvumelana, ukuze uxhumane ngokushesha, ngenkathi amamodeli ye-LLM zihlanganisa idatha abalandeli, amamodeli we-RAG asekelwe kuphela amamodeli ye-LLM esifundeni futhi zihlanganisa izazi zokufunda, kodwa futhi zihlanganisa ulwazi esithakazelisayo ezingenalutho. Lokhu kuncike ukuthi iyatholakala njalo. Lokhu kuhlanganisa inqaku lokuqala kulesi sethu. Kwi-artikel esilandelayo, siza kuhlukanisa kwezinhlobonhlobo ezingaphezu kwezobuchwepheshe. Ukuze ama-friends abalandela ukufinyelela noma ukubuyekeza ukusebenza kwalo umsebenzi, zihlanganisa izixhumanisi ezijwayelekile ezivela nge-Python kanye nezibhayisikobho ezihambisana emakhasini yami ye-GitHub ek Ngibuyekeze wena ku-artikeli elilandelayo ye-Series. Ukubuyekezwa kwe-Yurduzeven Ukungena ngemvume "Sihlola futhi ukunakekelwa umphumela wokufunda we-AI ye-RAG ye-Facebook ngo-2020, okuvumela kakhulu ukubuyekeza inkinobho se-artikeli." Ukulungiselela Github I-Multi-Model RAG Chatbot Project I-PDF ye-RAG Chatbot Project

Part of HackerNoon's growing list of open-source research papers, promoting free access to academic material.

This story contains new, firsthand information uncovered by the writer.

Read My Stories

Lo msindo ukhiqizwa ngolimi lokuqala lwendaba!

Izinhlelo ze-RAG zihlukanise izinzuzo ze-model ye-language: Nazi indlela

About Author

IMIBONO

HANG TAGS

LESI SIHLOKO SETHULWE NGAPHAKATHI

Related Stories

Zulu Network: Moving The Bitcoin Economy Forward With a Two-Tiered Bitcoin Layer 2 Architecture

The Billionaires F*cked Up

Zulu Network: Moving The Bitcoin Economy Forward With a Two-Tiered Bitcoin Layer 2 Architecture

The Billionaires F*cked Up

Light-Mode

Classic

Newspaper

Minty

Dark-Mode

Neon Noir

Minty

HN StartUps