Riba:
- Ƙirƙirar abubuwan haɗin gwiwa da kwararar bayanai na mataimakan RAG na ƙarshen-zuwa-ƙarshen
- Haɗa bayanan tushe da yawa (wiki, tikiti, PDF, bayanai) zuwa mataimaki guda ɗaya
- Yi yanke shawara na gine-gine don haɓakawa, caching da latency
A cikin raka'o'in da suka gabata, mun koyi sassan ɗaya bayan ɗaya: haɗawa, bayanai na vector, chunking, maidowa. Yanzu bari mu haɗa waɗannan kuma mu gina ƙarshen-zuwa-ƙarshen gine-gine na mataimaki wanda ke magana da bayanan kamfanin ku. Manufar shine a sami ma'aikaci ya tambayi, "Mene ne manufar barin mu?" Tsarin da mutane za su iya yin tambayoyi, amsoshin sun dogara ne akan ainihin takardun ciki, ƙididdiga da kuma haɗa tushen bayanai da yawa. Wannan rukunin yana aiwatar da dukkan gine-gine, kwararar bayanai da yanke shawara matakin samarwa.
Abubuwan Ƙarshe-zuwa-Ƙarshe
Mataimakin RAG na kamfani ya ƙunshi layi biyu daban-daban. Layin indexing (offline) yana shirya bayanai; Layin tambaya (kan layi) yana amsa tambayar.
Abubuwan haɗin layi na fihirisa:
- Masu haɗawa: Masu haɗawa waɗanda ke cire bayanai daga tushe - wiki, tsarin tikiti, ajiyar fayil, bayanai, imel.
- Daidaitawa: Canza nau'i daban-daban (PDF, HTML, DOCX) don tsaftace rubutu; gogewar kai/ƙafa.
- Chunking + metadata: chunking da yiwa alama (tushen, kwanan wata, iko).
- Shigar da + lodi: Rubutun ɓarna da metadata cikin ma'ajin bayanai na vector.
Abubuwan da ake buƙatar bututun mai:
- Gabatar da tambaya: Sake rubutawa, rarrabawa.
- Maidowa: Haɗaɗɗen bincike + tace metadata + sake yin matsayi.
- Ƙirƙirar gaggawa: Sanya mahallin + tambaya + umarni a cikin samfuri.
- Generation: Amsa mai tushe (na yau da kullun) daga ƙirar + tushe.
- Bayan aiwatarwa: Tsarin ƙira, duba tsaro, shiga.
Tukwici: A zahiri keɓance layin ƙididdiga daga layin tambaya. Fihirisa yana jinkiri kuma lokaci-lokaci (yana gudana cikin batches na dare); Ya kamata layin bincike ya zama haske da gaggawa. Hada layin biyu yana tilasta aiki mai nauyi yayin da mai amfani ke jira.
Ganin Gudun Bayanai
[INDEXING - offline] Albarkatun → Daidaita → Chunk+Metadata → Embed → Vector DB (wiki, tikiti, PDF, DB)[QUERY - online] Tambayar mai amfani → Gabatar da aiwatarwa → Maidowa (matattarar+matattarar+sakewa) → Gaggawa (maganin amfani
Haɗa Bayanan Maɓuɓɓuka Masu Mahimmanci
A cikin kamfanoni na gaske, amsar ba ta tsaya a wuri ɗaya ba. "Yadda za a mayar da kuɗi ga abokin ciniki?" Ana iya samun amsar tambayar duka a cikin labarin taimako (tsari), a cikin tarihin tikiti (misali na gaske), da kuma a cikin manufofin PDF (dokokin). Ya kamata mataimaki ya bincika dukansu a cikin tafkin daya.
Mahimmin batu: lokacin da ake haɗa albarkatu a cikin kantin sayar da kaya ɗaya, kowane shard dole ne ya ɗauki metadata 'source_tour'. Don haka kuna iya bincika su duka kuma ku tace su idan ya cancanta, kamar "kawo manufofin hukuma kawai". Hakanan, tushe daban-daban suna da matakan dogaro daban-daban: manufofin hukuma> labarin taimako> bayanin tikitin ma'aikaci. Kuna iya ƙididdige wannan fifiko a cikin sake yin matsayi ko faɗakarwa.
Source
Nau'in abun ciki
amana
Sabunta mita
Manufar PDF
hukuma hukuma
babba
kowane wata
Taimako labarin
Tsari
matsakaici-high
mako-mako
Tarihin tikiti
samfurin gaske
matsakaici
Ci gaba
wiki
Gaɗaɗɗen bayanin kula / halin yanzu
Mai canzawa
Ci gaba
Scalability, Cache da Latency
Batutuwa uku sun yi fice wajen samarwa. Latency: Ƙwarewar tana raguwa lokacin da mai amfani ya jira fiye da daƙiƙa 2. Magani: nuna amsar a cikin sigar yawo - ana zuba shi akan allon kamar yadda ƙirar ke rubutawa. Cache: Don tambayoyin da ake yawan yi da maimaitawa, cache duka yana ƙara sauri kuma yana rage farashi. Sikeli: Yayin da mai amfani ke ƙaruwa, ya zama dole a sami damar yin ƙima da ƙira da ƙira a kwance.
Dokar babban yatsan hannu a gefen farashi: mataki mafi tsada shine yawanci adadin alamun da ke zuwa mafi girma samfurin. Sabili da haka, rage mahallin zuwa sassa 4 masu kyau ta hanyar sake fasalin yana inganta duka inganci da farashi. Zane na gama gari shine a yi amfani da ƙaramin ƙarami/samfuri don sassauƙan rarrabuwa ko kewayawa, da mafi ƙarfi samfuri don amsa ta ƙarshe (misali claude-opus-4-8).
Tsanaki: Kar a saita firikwensin kamar "yi sau ɗaya, manta da shi". Ana canza takaddun, sharewa, ƙarawa. Ƙaddamar da dabarun sake-fidi: gano takardun da aka canza kuma a sake sarrafa su kawai. Fihirisar datti tana samar da amsa mai bayyana halin yanzu amma ba daidai ba.
Gine-gine mai rauni / Ƙarfin Gine-gine
Rauni (rubutu ɗaya, duk abin da ya gauraye):
Lokacin da mai amfani ya yi tambaya: karanta takaddun a wannan lokacin, yanke su, saka su, bincika su, amsa su. # Matsala: ana maimaita duk abubuwan da aka rubuta akan kowace tambaya; dakikoki na jinkirtawa, # babu rabuwar tushe, babu tacewa, babu wartsakewa.
Mai ƙarfi (raga bututu + metadata + cache + yawo):
Fihirisa: batch yana gudana da daddare, sabunta takaddun da aka canza. Tambaya: layi mai nauyi — pre-processing → haɗaɗɗen dawo da + tace → rerank → hanzari → samfuri (streaming) → ambato → log. Tambayoyin da ake yawan yi da kuma tushen ana adana su.
Mini Cases guda uku
Case 1 - Layin ruɗani, jinkiri mai nauyi. Wani farawa ya rubuta rubutun da ke sake sarrafa PDFs tare da kowace tambaya; Kowace amsa ta ɗauki matsakaicin daƙiƙa 11. Lokacin da aka raba layin firikwensin kuma an tura bayanan a baya zuwa ma'ajiyar vector, lokacin tambaya ya ragu zuwa daƙiƙa 1.3 kuma tare da yawo, "kalmar farko" ta bayyana a cikin 400 ms.
Hali na 2 - Abubuwan albarkatu da yawa, fifiko mara kyau. Mataimakin tallafi ya ba da ma'auni daidai da manufar PDF da tsoffin bayanan tikiti; Misalin wani lokaci yana gabatar da ƙimar ma'aikaci mara daidai daga shekaru biyu da suka gabata a matsayin ƙa'idar hukuma. Lokacin da tushen_tour metadata da "yi la'akari da manufofin hukuma idan akwai rikici" aka ƙara umarni zuwa ga gaggawa, kurakuran fifikon karya sun ragu da kashi 89%.
Case 3 - Matsakaicin Matsala. Mataimakin HR yana aiki tare da fihirisar da ba a sabunta ta tsawon watanni 3 ba; Manufar izinin barin ya canza, amma mataimaki yana faɗin zamanin da. Lokacin da aka shigar da sabuntawa na yau da kullun, wanda ke gano fayilolin da aka canza, ƙimar amsawar yanzu ta ƙaru daga 70% zuwa 99%.
Kuskuren gama gari
- Haɗin haɗawa da layukan tambaya: Ana yin aiki mai nauyi yayin da mai amfani ke jira; jinkiri ya fashe.
- Rashin sanya nau'in tushe a cikin metadata: Babu fifiko da tacewa; Majiyar da ba a amince da ita ba ta zama hukuma.
- Ba kafa dabarar wartsakewa ba: Fihirisar ta zama tawul; Ana samar da amsoshi marasa kuskure waɗanda suka bayyana na yanzu.
- Tsallake yawo: Mai amfani yana kallon allo mara komai; Jinkirin da aka gane ya zama babba.
- Yin amfani da samfurin mafi girma a kowane mataki: Farashin yana ƙaruwa ba dole ba; Bar tuƙi zuwa ƙaramin samfurin.
A takaice
- Mataimakin RAG na kamfani ya ƙunshi layi biyu daban-daban: ƙididdigar layi da kuma tambayar kan layi; raba su jiki.
- Fihirisa = mai haɗawa + daidaita + chunk / metadata + saka / loda; tambaya = pre-tsari + dawo da + faɗakarwa + ƙirƙira + bayan aiwatarwa.
- Ana haɗa bayanan tushen da yawa zuwa wurin ajiya guda ɗaya, amma tushen_type metadata da fifikon amana ana kiyaye su.
- Yawo da cache don latency, mahallin mahallin da zaɓin ƙira don farashi suna da mahimmanci.
- Ba tare da sake yin gyare-gyare ba, fihirisar ta zama tawul; Sake aiwatar da canza takardu akai-akai.
Aikin aikace-aikace
Zana zanen zane na mataimaki don ƙungiyar ku. (1) Gano aƙalla tushen bayanai na ainihi guda uku kuma rubuta buƙatun mai haɗawa, sabunta mita, da matakin amincewa ga kowane. (2) Zana firikwensin da layukan tambaya daban tare da zane-zanen kibiya. (3) "A ina zan rage jinkiri da farashi a cikin wannan mataimaki?" Rubuta aƙalla takamaiman yanke shawara guda biyu ga tambayar. (4) Bayyana dabarun wartsakewa a cikin jumla ɗaya: wace hanya za a sake fiddawa kuma sau nawa?
jerin abubuwan dubawa
- [ ] Zan iya zana firikwensin da layukan tambaya daban kuma tare da madaidaitan sassan.
- [ ] Zan iya haɗa bayanan tushe da yawa tare da tushen_type da fifikon amincewa.
- [ ] Zan iya yanke shawarar yawo/cache don jinkiri da zaɓin ƙira don farashi.
- [ ] Na san dalilin da ya sa dabarun sake firikwensin ke da mahimmanci.
- [ ] Ina tuna cewa mataki mafi tsada a cikin gine-gine na yawanci shine alamar da ke zuwa mafi girma samfurin.