Bayanan rumbun Vector
Wani cikakken bayanan vector da aka sarrafa wanda kawai PostgreSQL ne
Runga sarrafawar 'embeddings' sannan ka bincike su ta hanyar ma'ana, a kan pgvector wanda muke gudanarwa da kuma yi muku 'back up'. Babu wani sabon rumbun bayanai da za ku koya, babu damuwa kan girman 'cluster' — kawai ku kirkiri 'index', ku rubuta 'vectors' naku, sannan ku yi bincike da mabudin API wanda kuke da shi a yanzu.
Maji bayan da kake da kwarewa a kansa, tare da nau'in shafi guda daya karin
Ma'ajiyar bayanai ta vector yawanci tsari ce ta biyu da za a gudanar tare da ta farko: wani gungu daban, wani labarin ajiyar waje daban, wani abu daban da za a farkar da shi. Wannan shine pgvector a cikin PostgreSQL - inji ɗaya, dorewa ɗaya, da ajiye bayanan baya na dare da ci gaba da adana kundin rubutu na gaba da kowa ke samu anan. Alamar ku tebur ce ta al'ajabi tare da alamar HNSW a kanta, don haka tana da sauri don binciken maƙwabcin da ya fi kusa kuma mai ban sha'awa a kowane fanni, wanda shine halayyar da kuke so da karfe uku na safiya.
Fadin samfurinka, ba faɗin da muka zaɓa ba
Yawancin abubuwan da aka sarrafa na vector suna sa ka zaɓi daga ɗan gajeren jerin girma. Naku ba lallai ne ya zama lambar zagaye ba: duk abin da ya fiso daga 1 zuwa 3072 ana karɓa kuma ana riƙe shi daidai, domin an cika wani ɗan ƙaramin vector zuwa faɗin jiki na gaba da sifili — wanda ke canza tazara ta cosine, Euclidean da inner-product da daidai da komai. Model mai girma 1408 yana da matsayi daidai da wanda aka ƙera shi da shi, kuma karatu yana dawo muku da ƙimar 1408 da kuka adana, ba cikar ba.
An haɗa a kowane shirin vector
- Duk wani faɗin haɗawa har zuwa 3072, da nisan cosine, Euclidean ko inner-product
- Metadata da aka adana tare da kowane vector, kuma aka tace a lokacin bincike
- Adana bayanai na kowane dare tare da riƙewa na kwanaki 30 da kuma dawo da bayanan a takamaiman lokaci
- Tsaron matakin-jeri, ta yadda ba za a iya samun damar shiga vectors na wata kungiya ba daga wata kungiyar
Tsare-tsare da farashi
Matakai uku, masu farashi akan abubuwa guda biyu da suka fi cin mu kudi: adadin abin da kake ajiyewa da kuma adadin abin da kake nema.
An ƙaddara farashi bisa ga tsarin da ka yi
Kudin sun dogara ne ga yankin, girman na'ura, bandwidth, da kuma ajiya da ka zaba, don haka babu wani adadi guda na wata-wata da za mu iya rubutawa a nan a gaskiya. Fada mana aikinka kuma za mu ba ka farashi.
Wutarlantarki ta 100% mai sabuntawa ce ta ƙarfafa shi
Kowane sabar da muke sarrafawa, a kan kowane samfuri, wutar lantarki mai sabuntawa ce ke ba shi ikon aiki. Ba a biya diyarsa ba bayan faruwar lamari — an samo shi ta hanyar haka.
- Wutar lantarki mai sabuntawa 100% a duk fadin wuraren ajiyar yanar gizonmu
- Cibiyoyin bayanai da ke aiki a PUE na 1.1 zuwa 1.2
- Shafuka a kan dandalinmu na bai daya sun wuce gwajin Green Web Foundation — wata cibiya ta uku da za ku iya Tabbatar wa kanku
Tambayoyin da mutane ke yi kafin su saya
Wadanne samfuran embedding wannan ke aiki da su?
Duk wacce ta zama. Kuna tura mana lambobin da samfurinku ya samar, don haka OpenAI, Cohere, Voyage, wani samfurin Sentence-Transformers a kan na'urarku ta kanku da duk wani abu daban duk daya suke a gare mu. Kun gaya mana faɗin lokacin da kuka ƙirƙiri ma'amala - 384, 768, 1024, 1536 da 3072 duk sanannu ne, haka ma wani abu da ke tsakani.
Za mu iya canza nisa ko aikin tazara na bango daga baya?
A'a, kuma mun fi son mu fadi hakan fili maimakon barin ka gano kanka. Dukansu biyu suna yanke shawarar inda aka adana vector ɗinka da kuma yadda aka tsara su, don haka canza kowanne daga ciki zai ɓata komai da ke cikin index ɗin ba tare da an sanar da kai ba. Idan ka sauya zuwa wani nau'in embedding model daban, ka ƙirƙiri index na biyu sannan ka rubuta a ciki - abu guda da kowane vector database ke buƙata daga gare ka, kuma saboda dalili guda ɗaya.
Shin akwai wani abu da ya bambanta game da vectors masu fadi sosai?
Eh, kuma yana da kyau a sani. Sama da girma 2,000 ana adana darajojin ne a rabi-daidaito, domin fihirisar HNSW ta PostgreSQL ba ta karɓar shafi mai cike da daidaito da ya fi hakan faɗi. Ga samfuran girma 3072 wannan shi ne hanyar da pgvector kanta ta ba da shawarar kuma tasirin kan tsari ba shi da wani amfani — amma musayar ce ta gaske kuma ya kamata ku ji shi daga wurinmu maimakon ku hango shi.
Shin zan iya samun igiyar haɗin bayanan sirri (database connection string)?
Ba yau ba. Kuna samun damar index ɗinku ta hanyar HTTPS ta amfani da maɓallin API na Zinn®, wanda aka iyakance don haka maɓallin da zai iya bincike ba zai iya yin wani abu daban ba a cikin asusunku. Wannan ne ke ba mu damar auna tambayoyi da gaskiya kuma mu kiyaye sahun mai haya guda ɗaya daga isa ga na wani. Idan haɗin kai tsaye shine abin da kuke buƙata, ku gaya mana - fi son mu gina shi saboda wani ya tambaya maimakon mu yi hasashe.
Ta yaya ake auna shi, kuma zan iya duba lambobin da kaina?
A kan mita biyu: adadin abin da kuke ajiyewa da kuma adadin binciken da kuke gudanarwa a cikin watan kalanda. Dukansu suna kan allon ku yayin da suke faruwa. Ajiya tsari ne da aka buga maimakon girman tebur na zahiri — babban tsarin jeri, da faɗin da aka adjiye, da tsawon ID ɗinku da metadata - daidai don ku iya sake samar da adadin da muke biyan ku kuɗi maimakon ɗaukar maganarmu.