AI數(shù)據(jù)系統(tǒng)演進(jìn)：zilliz如何支撐RAG/Agent千萬級(jí)多租與多模態(tài)業(yè)務(wù)落地2025.06郭人通Zilliz合伙人/產(chǎn)品負(fù)責(zé)人AgendaAgenda2|?Copyright2025Zilliz 02多模態(tài)數(shù)據(jù)：統(tǒng)一的數(shù)據(jù)建模 03多租支持：有效應(yīng)對(duì)RAG/Agent等大規(guī)模多租場(chǎng)景 04WebSearch緩存與DeepSearch 05超大規(guī)模數(shù)據(jù)檢索與分析：從向量數(shù)據(jù)庫到向量數(shù)據(jù)湖3|?Copyright2025Zilliz廣泛部署于全球近萬家頭部企業(yè)客戶企業(yè)級(jí)向量數(shù)據(jù)庫解決方案，高性能、高可擴(kuò)展、高度穩(wěn)定從原型構(gòu)建到大規(guī)模生產(chǎn)部署，穩(wěn)定支持百億級(jí)向量規(guī)模300+☆35K+100M+3.0K+forksTheForresterWave?VectorDatabaseProviders,Q32024Zilliz在向量數(shù)據(jù)庫領(lǐng)域深耕八年，陪伴全球客戶深度迭代解決方案。4|?Copyright2025Zilliz5|?Copyright2025Zilliz6|?Copyright2025ZillizCISIONCOPASSDeloitte。ebayInflection中newrelic.accentureOMERSzofato7|?Copyright2025Zillizvivo8|?Copyright2025Zilliz9|?Copyright2025Zilliz數(shù)據(jù)來源:TheDigitizationoftheWorldbyIDC（9|?Copyright2025Zilliz2025年的新增數(shù)據(jù)將是非結(jié)構(gòu)化數(shù)據(jù)10|?Copyright2025ZillizArticle(ID)Title(Text)Title(SparseEmbedding)AuthorVideo(Video/ImageEmbedding)Video(SubscriptEmbedding)Summary(Text)Summary(DenseEmbedding)Summary(SparseEmbedding)DataTypes：VARCHARorINT64VARCHARSPARSE_FLOAT_VECTORJSONVARCHARSPARSE_FLOAT_VECTOR11|?Copyright2025Zilliz12|?Copyright2025Zilliz向量檢索與標(biāo)量過濾全文檢索2.0非結(jié)構(gòu)化數(shù)據(jù)檢索JSON檢索JSON檢索3.013|?Copyright2025Zilliz?稠密向量?JSON?List支持的查詢類型：?語義范圍查詢?表達(dá)式過濾?多向量混合查詢?多路召回與重排序標(biāo)量/JSON…14|?Copyright2025Zilliz標(biāo)量/JSON…15|?Copyright2025Zilliz16|?Copyright2025Zilliz標(biāo)量組/JSON17|?Copyright2025Zilliz數(shù)值/標(biāo)簽檢索非結(jié)構(gòu)化數(shù)據(jù)檢索3.0非結(jié)構(gòu)化數(shù)據(jù)檢索文本檢索文本檢索2025.04在所有數(shù)據(jù)類型上支持索引與加速檢索2025.05JSON索引增強(qiáng)引19|?Copyright2025ZillizDATAISOLATIONSEARCHPERFORMANCEStrongStrongDATAISOLATIONSEARCHPERFORMANCEStrongStrong100+MaxTenantscolectionswithstringentisolationneedstenantDATAISOLATIONStrongtenantDATAISOLATIONStrongSEARCHPERFORMANCEStrongMaxTenantsbalancedisolation&performanceneedsDATAISOLATIONMediumSEARCHPERFORMANCEStrongDATAISOLATIONMediumSEARCHPERFORMANCEStrongMaxTenantsBestfor:Rapidlyscalingtomilionsoftenantswitheficientresourceutilization21|?Copyright2025Zilliz22|?Copyright2025Zilliz用戶活躍度與數(shù)據(jù)冷熱度%Largenumberofusers,butmostofthemarenotactive.Typicalscenarios:personalknowledgebase,conversationco大規(guī)模網(wǎng)頁檢索/企業(yè)級(jí)知識(shí)庫Largevolumeofdata,butmostofitisnotfrequentlyaccessed.Typicalscenarios:customerservicesupport,companyknowledgebase,websearch,etc.23|?Copyright2025Zilliz多層存儲(chǔ)向量數(shù)據(jù)庫解決方案MemoryBasedDecreasingcharacteristicradiusDecreasingcharacteristicradiusS324|?Copyright2025Zilliz25|?Copyright2025Zilliz多租方案總結(jié)隔離性與成本：?重保用戶：基于Collection（千）/DataBase（百）/Cluster（十）實(shí)現(xiàn)物理多租隔離?一般用戶：基于PartitionKey實(shí)現(xiàn)邏輯多租隔離（百萬～千萬）?從一般租戶到重保租戶的升級(jí)機(jī)制活躍度/冷熱度與成本：?最活躍用戶/高熱度數(shù)據(jù)：內(nèi)存?一般活躍用戶/熱數(shù)據(jù)：本地NVMe?非活躍用戶/冷數(shù)據(jù)：對(duì)象存儲(chǔ)?主動(dòng)與自動(dòng)化數(shù)據(jù)放置策略優(yōu)化26|?Copyright2025ZillizWebSearch緩存與DeepSearch/zilliztech/deep-searcherZilliz’sDeepSearcher27|?Copyright2025ZillizDeepResearchFramework28|?Copyright2025Zilliz2929|?Copyright2025ZillizHybridIterativeDeepSearch30|?Copyright2025Zillizoriginal-小孩子發(fā)燒吃布洛芬還是對(duì)乙酰氨基酚rewritten–布洛芬和對(duì)乙酰氨基酚在兒童退燒方面的區(qū)別是什么？?jī)和貌悸宸业倪m用年齡和注意事項(xiàng)？?jī)和脤?duì)乙酰氨基酚的推薦劑量和副作用？rewritten–適用于兒童的治療呼吸道感BGE’sNewWork--O1Embedder:LetRetrieversThinkBeforeAction31|?Copyright2025ZillizIterativeSearchwithPositiveandNegativeSamples“Aframewithabluebusandamotorcycle”3232|?Copyright2025ZillizIterativeSearchwithPositiveandNegativeSamplesHowitworksviamovingthequerypointintheembeddingspacebasedonpositiveandnegativeexamplesYoucanalsousethesequery/positive/negativeembeddingstotuneyourembeddingmodelstiteration22nditeration多層存儲(chǔ)向量數(shù)據(jù)庫解決方案MemoryBasedDecreasingcharacteristicradiusDecreasingcharacteristicradiusS333|?Copyright2025Zilliz34|?Copyright2025Zilliz緩存與DeepSearch總結(jié)緩存的填充方式：?離線：軟實(shí)時(shí)場(chǎng)景的預(yù)爬取與預(yù)填充?在線：由在線查詢觸發(fā)的實(shí)時(shí)填充?構(gòu)建多層存儲(chǔ)緩存進(jìn)一步降低成本，提升緩存數(shù)據(jù)規(guī)模緩存系統(tǒng)內(nèi)小模型的工作：?補(bǔ)充Key，提高內(nèi)容的命中覆蓋?查詢語句改寫?打標(biāo)反饋與迭代搜索?緩存命中/缺失判斷增強(qiáng)Agent與大模型的工作：?根據(jù)緩存命中結(jié)果的充分性、實(shí)效性判斷是否需要補(bǔ)充Web查詢?緩存查詢與API查詢的輪次與比?持續(xù)的將Web查詢結(jié)果進(jìn)行緩存填充35|?Copyright202