Uchungechunge lwe-AI Interview 8: Uyini i-RAG? Kungani kucabangwa ukwenza iphrojekthi ye-RAG?

Uyini i-RAG?

I-RAG igcwele ukuthi Retrieval-Augmented Generation, okusho ukukhiqiza okuthuthukisiwe ngokusesha ngesiZulu.

Kalula nje, ubuchwepheshe bokunikeza i-modeli yolimi enkulu "incwadi yokubhekisela engavulwa noma nini".

Ungacabanga nge-modeli yolimi enkulu njengomfundi ophambili onomkhumbulo omuhle kanye nolwazi oluningi. Kodwa lo mfundi unezinkinga ezimbili ezemvelo:

1. Umkhawulo wolwazi: Ulwazi alufunda lufinyelela kuphela kudatha yesikhathi sokuqeqeshwa. Izehlakalo ezenzeka ngemva kuka-2023 akazi lutho ngazo.
2. Angase aqambe amanga: Lapho ebuzwa umbuzo angawazi, akasho ukuthi "angazi", kodwa uqambe impendulo ezwakalayo (lokhu kuthiwa yi-AI hallucination).

I-RAG ixazulula lezi zinkinga ezimbili. Inqubo yayo ilula, inezinyathelo ezintathu:

1. Ukusesha: Uma ubuza umbuzo, uhlelo luya ku "knowledge base" yangaphandle (njengayo yonke imibhalo yenkampani yakho, i-Wikipedia yakamuva, noma imithetho eminingi), bese lusesha ngokushesha futhi lithole izingxenyana ezihambisana kakhulu. Lokhu kufana nokwenza umfundi avule incwadi ukuze afune impendulo.
2. Ukuthuthukisa: Uhlelo luhlanganisa "umbuzo owubuzile" kanye ne "izingxenyana ezitholiwe" ndawonye, lakhe i-“prompt” ethuthukisiwe. Lokhu kufana nokunikeza umfundi izinto zokubhekisela.
3. Ukukhiqiza: I-modeli yolimi enkulu isekela leli prompt elithuthukisiwe ukukhiqiza impendulo yokugcina. Ayisaxhomeki olwazini oludala olusemkhumbulweni wayo, kodwa ibheka kakhulu kulokho okunikeziwe. Lokhu kufana nokuthi umfundi aphendule ngokubheka incwadi, hhayi ngokucabanga nje.

Isifaniso esilula:
- I-LLM evamile: "Ngingayilungisa kanjani ibhayisikili yami yohlobo oluthi XX?" → I-modeli iphendula ngenkumbulo, okungenzeka iyiphelelwe yisikhathi noma inamaphutha.
- I-RAG: "Ngingayilungisa kanjani ibhayisikili yami yohlobo oluthi XX?" → Kucela ukusesha incwadi yokulungisa yakamuva → Bese ikhiqiza: "Ngokwesahluko 3 sencwadi yokulungisa ka-2024, kufanele uqale ngokuthi..."

Kungani kucabangwa ukwenza iphrojekthi ye-RAG?

Ukwenza iphrojekthi ye-RAG, empeleni kuwukusebenzisa amandla nemikhawulo ye-LLM, ukukhulula amandla ayikhona. Izizathu eziyinhloko:

1. Ukuxazulula izinkinga "ukuguga kolwazi" kanye "nokuqamba amanga"
2. Isisusa: Ufuna i-LLM iphendule imibuzo ngezehlakalo zakamuva, idatha yangaphakathi, nemibhalo eyimfihlo, futhi impendulo ibe nezinkomba ezitholakalayo.

3. Inani: Uhlelo lwezempilo olusebenzisa i-RAG lungacaphuna amajenali ezokwelapha amasha ukuphendula “izimpawu zohlobo olusha lwe-COVID”, hhayi ukunikeza ulwazi oluphelelwe yisikhathi lwango-2021, futhi lunikeze umthombo owengeziwe, kunciphisa ubungozi bokukhuluma ngokungenacala.

4. Ukwenza i-AI ikwazi ukusebenza "nedatha eyimfihlo", futhi ivikeleke

5. Isisusa: Inkampani ngayinye ine-knowledge base yayo (izinkontileka, ikhodi, amarekhodi okusiza abathengi, njll). Le datha ayikwazi ukuqeqeshwa kabusha noma ukuthuthukisa imodeli (kubiza kakhulu, kunzima, kunobungozi bokweba idatha).

6. Inani: Nge-RAG, ungakha umsizi wemibuzo wangaphakathi wenkampani. Lapho umsebenzi ebuza umbuzo, i-AI ithole ulwazi olufanele emibhalweni eyimfihlo yenkampani ukuze iphendule. Idatha eyimfihlo ihlala ngaphakathi enkampanini, ayithunyelwa kubenzi bamodeli ukuze baqeqeshe, kusetshenziswa amandla okuqonda e-LLM, kugcinwa nokuphepha kwedatha.

7. Ukunciphisa izindleko, ukukhulisa ukusebenza kahle

8. Isisusa: Ukuqeqesha kabusha noma ukuthuthukisa i-modeli enkulu ukuze ifunde ulwazi olusha kufana nokufunda umtapo wezincwadi wonke kabusha, kudinga izindleko ezinkulu zamakhompyutha.

9. Inani: I-RAG icishe ingadingi ukuqeqeshwa, idinga nje ukwakha uhlelo lokusesha. Izindleko zingaba i-1% yalezo zokuthuthukisa, noma ngaphansi. Futhi uma i-knowledge base isibuyekeziwe, imiphumela yokusesha ibuyekeziwe ngokuzenzakalelayo, ngaphandle kokuqeqesha kabusha imodeli, kwenzeka “ukubuyekeza ngesikhathi sangempela”.

10. Ukwenza i-AI yazi ukuthi 'iyazi noma ayazi'

11. Isisusa: Ufuna imodeli ikwazi ukuhlukanisa imingcele yolwazi lwayo.
12. Inani: Uhlelo lwe-RAG lungabeka umthetho: uma kungatholakali imibhalo efanele, impendulo ibe “Uxolo, angitholanga ulwazi oluhambisana nalombuzo ku-knowledge base, sicela uqinisekise umbuzo wakho.” Lo mthetho wokungaphumeleli ekucaphuneni wenza ukusebenza kwe-AI kuthembeke futhi kube sobala.

Ukufingqa:

Kucatshangwa ukwenza iphrojekthi ye-RAG ngoba sifuna amandla okukhuluma nokuqonda kwe-LLM, kodwa futhi sifuna ukwenza i-LLM ibe 'neqiniso, ithembeke, ihambisane nesikhathi, futhi iqonde umsebenzi wangasese'. I-RAG ifana nokufaka isiteringi esinembayo ukuze siqondise injini engathi sína (LLM) kanye nemephu yokuqondisa evuselelekayo (uhlelo lokusesha), futhi ingomunye wezindlela eziyisisekelo zokwenza i-LLM isebenze emikhakheni ebalulekile njengezamabhizinisi, ezokwelapha, ezomthetho, nakwezezimali.