Izinhlobo Eziphezulu Eziyisi-9 Ezihlukile Zokubuyiswa-Isizukulwane Esithuthukisiwe (ama-RAG)

I-Retrieval-Augmented Generation (RAG) iwuhlaka lokufunda lomshini oluhlanganisa izinzuzo zazo zombili amamodeli asuselwe ekubuyiseni nasekelwe esizukulwaneni. Uhlaka lwe-RAG luhlonishwa kakhulu ngekhono lalo lokuphatha inani elikhulu lolwazi futhi lukhiqize izimpendulo ezihambisanayo nezinemba ngokomongo. Isebenzisa imithombo yedatha yangaphandle ngokuthola amadokhumenti afanelekile noma amaqiniso bese iveza impendulo noma okukhiphayo ngokusekelwe olwazini olubuyisiwe kanye nombuzo womsebenzisi. Le nhlanganisela yokubuyisa nokukhiqiza iholela emiphumeleni enolwazi olungcono enembe kakhudlwana futhi ebanzi kunamamodeli ancike kuphela ekukhiqizeni.

Ukuvela kwe-RAG kuholele ezinhlotsheni nezindlela ezahlukahlukene, ngayinye yakhelwe ukubhekana nezinselele ezithile noma ukukhulisa izinzuzo ezithile ezizindeni ezahlukahlukene. Ake sihlole ukuhlukahluka okuyisishiyagalolunye kohlaka lwe-RAG: I-RAG Ejwayelekile, i-Corrective RAG, i-Specculative RAG, i-Fusion RAG, i-Agentic RAG, i-Self RAG, i-Graph RAG, i-Modular RAG, ne-RadioRAG. Ngayinye yalezi zindlela ithuthukisa ngokuhlukile ukusebenza kahle kanye nokunemba kwenqubo yokukhiqiza ethuthukisiwe yokubuyisa.

I-RAG ejwayelekile

I I-RAG ejwayelekile uhlaka luyimodeli eyisisekelo ye-Retrieval-Augmented Generation. Incike enqubweni yezinyathelo ezimbili: Imodeli iqala ikhiphe ulwazi olufanele kudathasethi enkulu yangaphandle, njengesizinda solwazi noma inqolobane yemibhalo, bese ikhiqiza impendulo isebenzisa imodeli yolimi. Amadokhumenti abuyisiwe asebenza njengomongo owengeziwe embuzweni wokufakwayo, ethuthukisa ikhono lemodeli yolimi ukuze kwakhe izimpendulo ezinembile nezifundisayo.

I-RAG evamile iwusizo ikakhulukazi uma umbuzo udinga ulwazi olunembile noluyiqiniso. Isibonelo, ingxenye yokubuyisa idonsa izigaba ezifanele kudathasethi ezinhlelweni zokuphendula imibuzo noma imisebenzi efingqa amadokhumenti amakhulu. Ngesikhathi esifanayo, imodeli yokukhiqiza ihlanganisa ulwazi ekuphumeni okuhambisanayo.

Naphezu kokuhlukahluka kwayo, i-Standard RAG ingase ingabi nasici. Isinyathelo sokubuyisa kwesinye isikhathi siyehluleka ukukhomba amadokhumenti afaneleka kakhulu, okuholela ekuphenduleni okuncane noma okungalungile. Kodwa-ke, ngokuqhubeka nokucwenga izindlela zokubuyisa kanye namamodeli olimi ayisisekelo, i-Standard RAG isengenye yezakhiwo ze-RAG ezisetshenziswa kakhulu ezifundweni nasembonini.

I-RAG yokulungisa

I I-RAG yokulungisa imodeli yakhela phezu kwezisekelo ezijwayelekile ze-RAG kodwa yengeza isendlalelo esiklanyelwe ukulungisa amaphutha angaba khona noma ukungahambisani kwempendulo ekhiqizwayo. Ngemva kwezigaba zokubuyisa nokukhiqiza, kusetshenziswa indlela yokulungisa ukuze kuqinisekiswe ukunemba kokuphumayo okukhiqizwayo. Lokhu kulungiswa kungabandakanya ukubonisana okwengeziwe kwamadokhumenti abuyisiwe, ukulungisa kahle imodeli yolimi, noma ukusebenzisa amaluphu empendulo lapho imodeli izihlola ngokwayo okukhiphayo iqhathanisa nedatha eyiqiniso.

I-RAG yokulungisa iwusizo ikakhulukazi ezizindeni ezinembe kakhulu, njengokuxilongwa kwezokwelapha, iseluleko sezomthetho, noma ucwaningo lwesayensi. Kulezi zindawo, noma yikuphi ukungalungi kungaba nemiphumela ebalulekile; ngakho-ke, isendlalelo esengeziwe sokulungisa sivikela ulwazi olunganembile. Ngokucwenga isigaba sokukhiqiza kanye nokuqinisekisa ukuthi okukhiphayo kuhambisana nemithombo ethembeke kakhulu, I-RAG Yokulungisa ithuthukisa ukwethenjwa kwezimpendulo zemodeli.

I-RAG eqagelayo

I-RAG eqagelayo ithatha indlela ehlukile ngokukhuthaza imodeli ukuthi yenze ukuqagela okufundile noma izimpendulo eziqagelayo lapho idatha ebuyisiwe inganele noma ingaqondakali. Le modeli yakhelwe ukuphatha izimo lapho ulwazi oluphelele lungase lungatholakali khona, nokho isistimu isadinga ukunikeza impendulo ewusizo. Isici sokuqagela sivumela imodeli ukuthi ikhiqize iziphetho ezibambekayo ezisekelwe emaphethini edatha ebuyisiwe kanye nolwazi olubanzi olugxiliswe kumodeli yolimi.

Nakuba izimpendulo zokuqagela ngezinye izikhathi zingase zibe nembe ngokugcwele, zisenganikeza inani ezinqubweni zokwenza izinqumo lapho kungadingeki khona ukuqiniseka okuphelele. Isibonelo, ocwaningweni lokuhlola noma ukubonisana kokuqala kwezezimali, ukumaketha, noma ukuthuthukiswa komkhiqizo, i-Specculative RAG inikeza izixazululo ezingaba khona noma imininingwane ukuze iqondise uphenyo olwengeziwe noma ukucwengwa. Kodwa-ke, enye yezinselelo ezinkulu nge-Specculative RAG iqinisekisa ukuthi abasebenzisi bayayazi imvelo yokuqagela yezimpendulo. Njengoba imodeli yakhelwe ukukhiqiza imibono kuneziphetho eziyiqiniso, imvelo yokuqagela kufanele ivezwe ngokucacile ukuze kugwenywe abasebenzisi abadukisayo.

I-Fusion RAG

I-Fusion RAG imodeli ethuthukisiwe ehlanganisa ulwazi oluvela emithonjeni eminingi noma imibono ukuze udale impendulo ehlanganisiwe. Le ndlela iwusizo ikakhulukazi lapho amasethi edatha ahlukene noma amadokhumenti enikeza ulwazi oluhambisanayo noma oluqhathanisayo. I-Fusion RAG ibuyisa idatha emithonjeni eminingana bese isebenzisa imodeli yokukhiqiza ukuze ihlanganise lokhu okokufaka okuhlukahlukene kokuphumayo okuhlangene, okuzungezwe kahle.

Le modeli inenzuzo ezinqubweni eziyinkimbinkimbi zokwenza izinqumo, ezifana namasu ebhizinisi noma ukwakhiwa kwenqubomgomo, lapho kufanele kucatshangelwe imibono ehlukene namasethi edatha. Ngokuhlanganisa idatha evela emithonjeni ehlukahlukene, i-Fusion RAG iqinisekisa ukuthi okukhiphayo kokugcina kubanzi futhi kunezici eziningi, kubhekwana nokuchema okungaba khona ekuthembeleni kudathasethi eyodwa. Enye yezinselelo ezibalulekile nge-Fusion RAG ingozi yokugcwala ngokweqile kolwazi noma amaphuzu edatha angqubuzanayo. Imodeli idinga ukulinganisa nokuvumelanisa okokufaka okuhlukahlukene ngaphandle kokuphazamisa ukuhambisana noma ukunemba kokuphumayo okukhiqizwayo.

I-RAG ye-Agent

I-RAG ye-Agent yethula ukuzimela ohlakeni lwe-RAG ngokuvumela imodeli ukuthi isebenze ngokuzimela ekunqumeni ukuthi yiluphi ulwazi oludingekayo kanye nendlela yokuluthola. Ngokungafani namamodeli e-RAG avamile, ngokuvamile akhawulelwe ezindleleni zokubuyisa ezichazwe kusengaphambili, i-Agentic RAG ihlanganisa ingxenye yokwenza izinqumo evumela isistimu ukuthi ibone imithombo eyengeziwe, ibeke phambili izinhlobo ezihlukene zolwazi, noma iqale imibuzo emisha ngokusekelwe kokufakwayo komsebenzisi.

Lokhu kuziphatha kokuzimela kwenza i-Agentic RAG isebenziseke ikakhulukazi ezindaweni eziguqukayo lapho ulwazi oludingekayo lungase luguquke khona, noma inqubo yokubuyisa idinga ukuzivumelanisa nezimo ezintsha. Izibonelo zohlelo lwayo zingatholakala ezinhlelweni zocwaningo ezizimele, ama-bot esevisi yamakhasimende, nabasizi abahlakaniphile abadinga ukuphatha imibuzo eguqukayo noma engalindelekile. Inselele eyodwa nge-Agentic RAG iqinisekisa ukuthi ukubuyisa okuzenzakalelayo kanye nezinqubo zokukhiqiza zihambisana nezinjongo zomsebenzisi. Amasistimu azimele ngokweqile angase aphambuke kakhulu kumsebenzi ohlosiwe noma anikeze ulwazi olungabalulekile embuzweni wangempela.

I-self-RAG

I-self-RAG iwukuhluka okubonakala kakhudlwana kwemodeli egcizelela ikhono lesistimu lokuhlola ukusebenza kwayo. Ku-Self-RAG, imodeli ikhiqiza izimpendulo ngokusekelwe kudatha ebuyisiwe futhi ihlola ikhwalithi yezimpendulo zayo. Lokhu kuzihlola kungenzeka ngokusebenzisa izihibe zempendulo zangaphakathi, lapho imodeli ihlola ukuvumelana kokukhiphayo kumadokhumenti abuyisiwe, noma ngezindlela zempendulo zangaphandle, ezifana nezilinganiso zomsebenzisi noma izilungiso.

I-Self-RAG ibaluleke kakhulu ezinhlelweni zemfundo nokuqeqesha, lapho ukuthuthukiswa okuqhubekayo nokunemba kubalulekile. Isibonelo, kumasistimu aklanyelwe ukusiza ngokufundisa noma ukufunda okuzenzakalelayo, i-self-RAG ivumela imodeli ukuthi ikhombe izindawo lapho izimpendulo zayo zingashoda khona futhi ilungise ukubuyisa noma amasu okukhiqiza ngokufanele.

Inselele enkulu nge-Self RAG ukuthi ikhono lemodeli lokuzihlola lincike ekunembeni nasekupheleleni kwamadokhumenti abuyisiwe. Uma inqubo yokuthola ibuyisela idatha engaphelele noma engalungile, izindlela zokuzihlola zingase ziqinise lokhu kungalungile.

Igrafu ye-RAG

Igrafu ye-RAG ihlanganisa izakhiwo zedatha ezisekelwe kugrafu enqubweni yokuthola, okuvumela imodeli ukuthi ithole futhi ihlele ulwazi ngokusekelwe ebudlelwaneni bebhizinisi. Iwusizo ikakhulukazi ezimweni lapho ukwakheka kwedatha kubalulekile ukuze kuqondwe, njengamagrafu olwazi, amanethiwekhi omphakathi, noma izinhlelo zokusebenza zewebhu ze-semantic.

Ngokusebenzisa amagrafu, imodeli ingakwazi ukubuyisa ulwazi olungalodwa kanye nokuxhumana kwalo. Isibonelo, kumongo wezomthetho, i-Graph RAG ingathola umthetho wamacala afanelekile kanye nezandulela ezixhuma lezo zimo, inikeze ukuqonda okune-nuanced okwengeziwe kwesihloko.

Igrafu i-RAG ihamba phambili ezizindeni ezidinga ukuqonda okujulile kobudlelwane, njengocwaningo lwebhayoloji, lapho ukuqonda ubudlelwano phakathi kwezakhi zofuzo, amaprotheni, nezifo kubalulekile. Enye yezinselelo ezinkulu ngeGraph RAG iqinisekisa ukuthi izakhiwo zegrafu ziyabuyekezwa futhi zigcinwe ngokunembile, njengoba amagrafu aphelelwe yisikhathi noma angaphelele angaholela ezimpendulweni ezingalungile noma ezingaphelele.

I-RAG ye-Modular

I-RAG ye-Modular ithatha indlela evumelana nezimo futhi eyenzeka ngokwezifiso ngokwephula ukubuyisa kanye nezingxenye zokukhiqiza zibe amamojula ahlukene, alungiselelwe ngokuzimela. Imojuli ngayinye ingacushwa kahle noma ishintshwe kuye ngomsebenzi othile. Isibonelo, izinjini zokubuyisa ezihlukene zingase zisetshenziselwe amasethi edatha ahlukene noma izizinda, kuyilapho imodeli ekhiqizayo ingase yakhelwe izinhlobo ezithile zezimpendulo (isb., okuyiqiniso, okuqagelayo, noma ubuciko).

Le modularity ivumela i-Modular RAG ukuthi ivumelane nezimo kakhulu, iyenze ilungele izinhlelo zokusebenza ezahlukahlukene. Isibonelo, ohlelweni oluyingxube losekelo lwekhasimende, imojula eyodwa ingase igxile ekubuyiseni ulwazi kumanuwali yobuchwepheshe, kanti enye ingase ibuyise ama-FAQ. Imojula yokukhiqiza izobe isihlanganisa impendulo ngohlobo oluthile lombuzo, iqinisekise ukuthi imibuzo yobuchwepheshe ithola izimpendulo ezinemininingwane, eziyiqiniso. Ngesikhathi esifanayo, imibuzo evamile ihlangatshezwana nezimpendulo ezibanzi, ezisebenziseka kalula. Inzuzo eyinhloko ye-Modular RAG ilele ekuguquguqukeni kwayo, okwenza abasebenzisi benze ngendlela oyifisayo ingxenye yesistimu ngayinye ukuze ihambisane nezidingo zabo ezithile. Kodwa-ke, ukuqinisekisa ukuthi amamojula ahlukahlukene asebenza ndawonye ngaphandle komthungo kungaba inselele, ikakhulukazi lapho usebenza namasistimu okubuyisa akhethekile noma uhlanganisa amamodeli ahlukahlukene okukhiqiza.

I-RadioRAG

I-RadioRAG ukuqaliswa okukhethekile kwe-RAG ethuthukiswe ukubhekana nezinselele zokuhlanganisa ulwazi lwesikhathi sangempela, oluqondene nesizinda kuma-LLM e-radiology. Ama-LLM endabuko, nakuba enamandla, ngokuvamile akhawulwa idatha yawo yokuqeqeshwa emile, engaholela ekuphenduleni okudlulelwe yisikhathi noma okungalungile, ikakhulukazi emikhakheni eguquguqukayo njengomuthi. I-RadioRAG inciphisa lo mkhawulo ngokuthola ulwazi lwakamuva oluvela emithonjeni egunyaziwe ye-radiological ngesikhathi sangempela, ithuthukisa ukunemba nokufaneleka kwezimpendulo zemodeli. Ngokungafani nezinhlelo ze-RAG zangaphambilini ezithembele ekuhlanganisweni kwangaphambili, kolwazi olumile, i-RadioRAG idonsa ngenkuthalo idatha kusuka kudathabhethi yolwazi lwe-radiology eku-inthanethi, iyivumele ukuthi iphendule ngolwazi oluqondene nomongo, lwesikhathi sangempela.

I-RadioRAG ihlolwe ngokuqinile kusetshenziswa idathasethi ezinikele, i-RadioQA, eyakhiwe imibuzo ye-radiologic evela emikhakheni ehlukahlukene ehlukahlukene, okuhlanganisa ukuthwebula izithombe zamabele kanye ne-radiology yezimo eziphuthumayo. Ngokuthola ulwazi olunembile lwe-radiological ngesikhathi sangempela, i-RadioRAG ithuthukisa amakhono okuxilonga ama-LLM, ikakhulukazi ezimeni lapho ulwazi lwezokwelapha oluningiliziwe kanye nolwamanje lubalulekile. Ukusebenza kwayo kuwo wonke ama-LLM amaningi, njenge-GPT-3.5-turbo, GPT-4, namanye, kuthuthukise kakhulu ukunemba kokuxilonga, ngamamodeli athile athola izinzuzo zokunemba ezifika ku-54%. Le miphumela igcizelela amandla e-RadioRAG okuguqula ukuhlonzwa kwezokwelapha okusizwa yi-AI ngokunikeza ama-LLM ngokufinyelela okuguquguqukayo kudatha ethembekile, enegunya, okuholela emininingwaneni eminingi enolwazi nenembile ye-radiological.

Isiphetho

Ukwehluka ngakunye Kokubuyiswa-Okwengeziwe Kwesizukulwane kufeza injongo ehlukile, kuhlinzekela izidingo nezinselele ezahlukene ezizindeni ezahlukahlukene. I-RAG evamile ihlala iyisisekelo sezinhlelo zokusebenza eziningi. Ngokuphambene, amamodeli akhethekile afana ne-Corrective RAG, i-Specculative RAG, i-Fusion RAG, i-Agentic RAG, i-Self RAG, i-Graph RAG, i-Modular RAG, ne-RadioRAG inikeza izithuthukisi ezihambisana nezidingo ezithile. Njengoba la mamodeli athuthuka, angakwazi ukuguqula izimboni ngokunikeza ulwazi olunembe kakhudlwana, olunokuqonda, nolwazi oluhambisana nezimo, ngokuqhubekayo ukuvala igebe phakathi kokubuyiswa kwedatha kanye nokwenziwa kwezinqumo ezihlakaniphile.

Imithombo