InoteveraGaidhi rinotevera
Multi-Query Attention
Mutauro AI
Nhungamiro yehunyanzvi
Kunyorazve mubvunzo uye kudzoreredza kwemibvunzo yakawanda inyanzvi dzeRAG dzinoshandura mubvunzo wemushandisi usati watsvaga, nekuudzokorora, kuutyora kuita mibvunzo midiki, kana kugadzira akati wandei nekubatanidza zvabuda.
Teerera kugwaro iriAI-yakarondedzerwa · ~3 min
Izvo zvine basa nekuti vashandisi vanowanzo bvunza isina kujeka, yekukurukurirana kana yakawanda-chikamu mibvunzo ine mazwi asingaenderane nemagwaro, uye mubvunzo uri nani unogona kutaura ndima dzakakodzera kutsvaga kumwechete chaiko kungapotsa.
Hunhu hwekutsvagisa hunoenderana zvakanyanya nemubvunzo. Vashandisi vanonyora mibvunzo mipfupi, isinganzwisisike kana yekukurukurirana, nepo zvinyorwa zvichishandisa iwo ega mazwi. Sanduro yemubvunzo inovhara gap iyo yekutsvaga isati yaitika. Iyo yakapusa fomu ndeyekunyorazve: LLM inoshandura meseji yemushandisi kuita mubvunzo wakajeka wekutsvaga. Izvi zvakakosha mu-multi-turn chat, uko kunotevera se "uye muna 2023?" hazvireve chinhu pasina nhoroondo, saka sisitimu inodzoreredza nhaurirano kuita mubvunzo wakazvimirira. Kuparara kunopatsanura mubvunzo wakaoma kuita mibvunzo midiki inogona kutorwa yakaparadzana, inobatsira pakuenzanisa kana mibvunzo yakawanda-hop. Nhanho-shure yekukurudzira, yakatsanangurwa neGoogle DeepMind vaongorori muna 2023, inobvunza mubvunzo wakawanda kutanga kutora kumashure misimboti. Multi-query kudzoreredza inoburitsa akati wandei emubvunzo, inomhanyisa tsvakiridzo yega yega, uye inobatanidza mhinduro. LangChain yakasimudzira izvi neMultiQueryRetriever yayo. RAG-Fusion inosanganisa mibvunzo yakawanda ine reciprocal rank fusion (RRF), nzira yakatsanangurwa naCormack, Clarke naBuettcher muna 2009, iyo inokora gwaro rega rega nekupfupisa 1/(k + chinzvimbo) pane zvese zvawanikwa. Mapepa anoonekwa epamusoro mumazita akati wandei anokwira kumusoro asingade zvibodzwa zvinofananidzwa. Imwe nzira yakabatana, HyDE (Hypothetical Document Embeddings, Gao uye vaanoshanda navo, 2022), ine modhi inonyora mhinduro yekufungidzira uye inomisikidza iyo panzvimbo yemubvunzo, nekuti chinyorwa chakaita semhinduro chinowanzogara padyo nendima dzemhinduro chaidzo munzvimbo yekumisikidza. Mafungiro asiri iwo: mizhinji mibvunzo haisi nani nguva dzose. Musiyano wega wega unowedzera latency uye mutengo, uye akasiyana akagadzirwa zvisina kunaka anogona kukukurwa kubva muchinangwa chemushandisi, achidhonza magwaro asina basa. Kunyora zvakare kunogona kudonhedza zvipingaidzo zvakakosha semazuva kana mazita echigadzirwa. Yera kurangarira uye pindura mhando pamibvunzo chaiyo usati watora matanho aya.
Mumubatsiri wekutaura, iyo yekutevera "ko nezvevashandi venguva pfupi?" inonyorwazve kumubvunzo wakamira wakadai se "gwaro rezororo revabereki revashandi venguva pfupi" vachishandisa hurukuro yekutanga.
Mubvunzo wekuenzanisa mitengo yevapi vemakore maviri unoiswa mukutsvaga kwakasiyana kwemupi wega wega, uye mibairo inosanganiswa isati yapindurwa.
Bot yekutsigira inogadzira zvirevo zvina zve "app yangu inoramba ichindiburitsa", kusanganisira tekinoroji mazwi sekupera kweseshini uye tokeni yekuzorodza, uye inosanganisa mhedzisiro nekudzoreredza chinzvimbo fusion.
Mubatsiri wekutsvagisa anoshandisa nhanho yekudzokera kumashure kutanga kutsvaga iyo general musimboti kuseri kwemubvunzo wakamanikana, obva atora zvakati wandei, achipa iyo modhi zvese kumashure uye zvakati.
Sezvo agentic RAG ichipararira, shanduko yemubvunzo iri kuwedzera kubatwa nemuenzaniso uchisarudza chekutsvaga, ichidzokorodza kana mhinduro dzichiita sedzisina kusimba, kwete nedanho rekunyora patsva. Izvo zvinoita kuti iwo epasi mazano, kuora, mutsauko uye fusion, zvinyanye kukosha kunyangwe zvazvinoita zvishoma. Kunatsurudzwa mumamodheru ekumisikidza uye kutsvaga kwakasanganiswa kunoderedza kusawirirana kwemazwi asi hazvibvise kukosha kwekujekesa mibvunzo isina kujeka kana yekutaura. Tarisira kuti zvikwata zvishandise nzira idzi zvakasarudzwa, zvichikonzerwa kana mubvunzo uchiita sewakaoma kana kuti kutora kwekutanga kunenge kusina kusimba, kuenzanisa kunaka nemutengo.
Tsamba yemahara
Nhatu dzakasimbiswa dzeAI nyaya mangwanani ega ega pakati pevhiki, dzakanyorwa neChirungu chakajeka. Yemahara zvachose, hapana ads.
Imwe email zuva rega rega revhiki. Bvisa kunyoresa kamwe chete. Hatimbotengese kana kugovera kero yako.
Zviedze iwe pachako
Mhinduro yepakarepo pamhinduro yega yega, uye chitupa chinogovaniswa chine ID inobatika kana wapedza kosi.
Tsigira dzidzo yemahara yeAI. AI Understanding ndeye 501(c)(3) isingabatsiri - hapana ads, hapana paywall, zvachose. Ita chipo
Kunyorazve mubvunzo uye kudzoreredza kwemibvunzo yakawanda inyanzvi dzeRAG dzinoshandura mubvunzo wemushandisi usati watsvaga, nekuudzokorora, kuutyora kuita mibvunzo midiki, kana kugadzira akati wandei nekubatanidza zvabuda. Izvo zvine basa nekuti vashandisi vanowanzo bvunza isina kujeka, yekukurukurirana kana yakawanda-chikamu mibvunzo ine mazwi asingaenderane nemagwaro, uye mubvunzo uri nani unogona kutaura ndima dzakakodzera kutsvaga kumwechete chaiko kungapotsa.
Kutevera senge "uye muna 2023?" inoda nhoroondo yenhaurirano kuti ive mubvunzo unobatsira wega.
Decomposition inotyora mibvunzo yakaoma kana yakawanda-chikamu kuita mibvunzo midiki inodzoserwa zvakasiyana.
Nhanho-shure yekukurudzira, kubva kuGoogle DeepMind vaongorori, inotora zvakajairika kumashure zvisati zvatsanangurwa.
RRF inoverengera 1/(k + chinzvimbo) pamusoro pechinyorwa chega chega chine gwaro, magwaro emubairo akaiswa kumusoro mune akati wandei.
Nekuti RRF inoshandisa zvinzvimbo chete, inogona kusanganisa zvinyorwa zvine zvibodzwa zvisina kuenzaniswa.
Ramba uchidzidza
Mamwe madhairekitori akasarudzirwa nyaya iyi
InoteveraGaidhi rinotevera
Multi-Query Attention
Mutauro AI