I-Vector Search: I-Reranker Algorithm Showdown

Ukusesha kweVekhtha kuletha ngempumelelo ukufana kwe-semantic ukuze kubuyiswe isizukulwane esithuthukisiwe, kodwa akwenzi kahle ngokusesha kwamagama angukhiye amafushane noma amagama okusesha angaphandle kwesizinda. Ukwengeza ukubuyiswa kwe-vector ngokusesha kwamagama angukhiye njenge-BM25 nokuhlanganisa imiphumela ne-reranker kuba indlela evamile yokuthola okungcono kakhulu kuyo yomibili imihlaba.

Ama-Reranker amamodeli e-ML athatha isethi yemiphumela yosesho futhi ayihlele kabusha ukuze kuthuthukiswe ukuhambisana. Bahlola umbuzo obhangqwe nomphumela ngamunye wekhandidethi ngemininingwane, obiza kakhulu ngokwekhompyutha kodwa okhiqiza imiphumela enembe kakhulu kunezindlela zokubuyisa ezilula zizodwa. Lokhu kungenziwa njengesigaba sesibili phezu kokusesha okukodwa (khipha imiphumela eyi-100 ekusesheni kwevekhtha, bese ucela ohlela kabusha ukuthi akhombe okungu-10 okuphezulu) noma, kaningi, ukuhlanganisa imiphumela evela ezinhlotsheni ezihlukene zokusesha; kulokhu, usesho lwe-vector nokusesha igama elingukhiye.

Kodwa kuhle kangakanani abahlaziyi be-off-the-shelf? Ukuze ngithole, ngihlole ama-renkers ayisithupha embhalweni ovela ku- I-ViDoRe benchmark , usebenzisa i-Gemini Flash ukukhipha umbhalo ezithombeni. Imininingwane ngamasethi edatha ingatholakala esigabeni 3.1 sokuthi iphepha laseColPali . Ngokuphawulekayo, imithombo ye-TabFQuAD kanye ne-Shift Project ingesiFulentshi; ezinye zingesiNgisi.

Sihlole lezi renkers:

• I-Reciprocal Rank Fusion (RRF), ifomula yokuhlanganisa imiphumela evela emithonjeni eminingi ngaphandle kokwazi lutho mayelana nemibuzo noma imibhalo; kuncike ngokuphelele ekuhleleni okuhlobene ngaphakathi komthombo ngamunye. I-RRF isetshenziswa ku-Elastic ne-LlamaIndex, phakathi kwamanye amaphrojekthi.

• I-Cohere Rerank v3 futhi I-Jina Reranker v2 , cishe amamodeli aphethwe athandwa kakhulu.

• I-BGE-reranker-v2-m3 , imodeli yomthombo ovulekile enamaphuzu aphezulu kakhulu (ilayisensi ye-Apache).

• I-Voyage Renk-2 kanye ne-reank-2-lite, esanda kukhishwa ( ngoSepthemba ) yinkampani eqinile.

Abahlaziyi banikezwe imiphumela ephezulu engama-20 evela kokubili i-DPR ne-BM25, futhi i-NDCG@5 ehlaziywe yahlolwa.

Emiphumeleni, usesho lwevekhtha eluhlaza (okushumekiwe olusuka kumodeli ye-bge-m3) lubhalwe ukuthi dpr (ukubuyisa isiqephu esiminyene). I-BGE-m3 ikhethelwe ukubala okushumekiwe ngoba yilokho okwasetshenziswa ababhali be-ColPali njengesisekelo.

Nansi idatha emayelana nokuhambisana (NDCG@5):

Futhi nakhu ukuthi bashesha kangakanani ekuhleleni kabusha usesho kudathasethi ye-arxiv; I-latency ilingana nobude bedokhumenti. Lokhu ukubambezeleka kwe-graphing, ngakho okuphansi kungcono. Imodeli ye-bge ezibambele yona yasetshenziswa kuzo zombili i-NVIDIA 3090 kusetshenziswa ikhodi elula kakhulu ephakanyiswe isuka ikhadi lemodeli ye-Hugging Face .

Okokugcina, nakhu ukuthi kubiza malini ngemodeli ngayinye ukuphinda umise ukusesha okucishe kube ngu-3,000 kuwo wonke amadathasethi ayisithupha. Izintengo ezihlangene ngokusesha ngakunye (nezinkokhelo ezengeziwe zamadokhumenti amade), kanti amanye amanani ethokheni ngayinye.

Ukuhlaziya

• Wonke amamodeli enza cishe kahle kudathasethi yesiFulentshi njengoba enza kweyesiNgisi.

• I-Cohere ibiza kakhulu futhi inikezela kancane (kodwa ngokungaguquki) ukuhambisana okubi kakhulu kunezinye iziqalisi ze-ML - kodwa ishesha izikhathi ezi-3 kunezinsizakalo ezilandelayo ezishesha kakhulu. Ngaphezu kwalokho, Imikhawulo yezinga elijwayelekile lika-Cohere yizona eziphana kakhulu.

• I-Voyage rerank-2 iyinkosi yokubeka kabusha ukuhlobana kuwo wonke amasethi wedatha, ukuze uthole okunye okuthinta ukubambezeleka. Ngokuphawulekayo, ukuphela kwemodeli engenzi okubi kakhulu kune-DPR iyodwa kudathasethi ye-arxiv, ebonakala ikhohlisa kakhulu.

• I-Voyage reranker-2-lite ne-jina reranker v2 afana kakhulu, anesivinini esifanayo, asingathwe ngentengo efanayo, futhi asondele ekuhlobaneni okufanayo (ngonqenqema oluncane lwe-Voyage). Kodwa Umkhawulo wesilinganiso ojwayelekile we-Voyage kabili igama , futhi nge-Voyage uthola "impela" Python iklayenti esikhundleni sokwenza izicelo ezingaphekiwe ze-http.

• I-BGE-reranker-v2-m3 iyimodeli engasindi (ngaphansi kwamapharamitha angu-600M) kangangokuthi ngisho nakumthengi omdala we-GPU iyashesha.

Isiphetho

I-RRF yengeza inani elincane noma elingekho ezimeni zokucinga ezixubile; engxenyeni yamadathasethi, lenze kabi kakhulu kune-BM25 noma i-DPR iyodwa. Ngokuphambene, bonke abahlaziyi abasekelwe ku-ML abahloliwe balethe ukuthuthuka okunengqondo ngaphezu kwevekhtha emsulwa noma ukusesha kwegama elingukhiye, i-Voyage rerank-2 ibeka ibha ukuze ihlobane.

Ukuhwebelana kusekhona: ukunemba okuphakeme okuvela ku-Voyage relank-2, ukucutshungulwa okusheshayo okuvela ku-Cohere, noma ukusebenza okuqinile okumaphakathi okuvela kuJina noma kumodeli ye-Voyage lite. Ngisho ne-reranker yomthombo ovulekile ye-BGE, ngenkathi ilandela izinketho zokuthengisa, yengeza inani elibalulekile emaqenjini akhetha ukuzibamba.

Njengoba amamodeli esisekelo eqhubeka nokuthuthuka, singalindela ukusebenza okungcono nakakhulu. Kodwa abahlaziyi be-ML banamuhla sebevuthwe ngokwanele ukuthi bangasebenza ngokuzethemba kukho konke okuqukethwe ngezilimi eziningi.

Ngu-Jonathan Ellis, DataStax