Equilibrando precisão e desempenho: como os novos parâmetros do Zilliz Cloud ajudam você a otimizar a busca vetorial
Introdução
Nos últimos anos, grandes modelos de linguagem (LLMs) e bancos de dados vetoriais tornaram-se a espinha dorsal de inúmeras aplicações impulsionadas por IA — de recomendações de e-commerce a reconhecimento facial e sistemas RAG. No entanto, esses diversos casos de uso apresentam um desafio comum: como equilibrar a precisão da busca (taxa de recall) com o desempenho (latência e throughput)?
É exatamente por isso que estamos animados em apresentar dois novos recursos poderosos na versão mais recente do Zilliz Cloud, nosso banco de dados vetorial totalmente gerenciado, construído sobre o Milvus:
O
levelparâmetro - Um ajuste simples, mas poderoso, para refinar a precisão da buscaO
enable_recall_calculationparâmetro - Uma ferramenta integrada para estimar e validar taxas de recall
Essas adições capacitam os desenvolvedores a encontrar o equilíbrio perfeito para seus casos de uso específicos — seja você precise de recomendações extremamente rápidas ou de aplicações de segurança altamente precisas. Neste blog, mostraremos exatamente como aproveitar esses recursos para otimizar suas implementações de busca vetorial.
Diferentes casos de uso, diferentes requisitos
À medida que a busca vetorial se prolifera entre os setores, observamos que os requisitos de recall, latência e consultas por segundo (QPS) variam drasticamente entre as aplicações. Vamos dar uma olhada em dois exemplos contrastantes:
Sistemas de recomendação: velocidade acima da precisão perfeita
Sistemas de recomendação filtram vastas bibliotecas de conteúdo para sugerir itens relevantes com base nas preferências do usuário. Nesses sistemas, a taxa de recall não é a principal prioridade. Embora as recomendações devam ser relevantes, introduzir alguma variedade frequentemente melhora a descoberta e o engajamento do usuário.
Em vez disso, esses sistemas devem lidar com milhares de solicitações simultâneas em tempo real, exigindo:
Alto QPS para atender muitos usuários simultaneamente
Latência muito baixa para experiências de usuário responsivas
Recall moderado (85-95%) com tolerância para algumas correspondências imperfeitas
O impacto comercial de recomendações lentas geralmente supera sugestões ocasionalmente imperfeitas, tornando a otimização de desempenho crítica.
Reconhecimento facial: a precisão é inegociável
Sistemas de reconhecimento facial, especialmente em contextos de segurança, têm requisitos totalmente diferentes. Eles devem identificar com precisão usuários autorizados para evitar tanto falsos positivos (violações de segurança) quanto falsos negativos (usuários legítimos sendo negados).
Esses sistemas precisam de:
Recall muito alto (99%+) para identificação precisa
Tolerância para latência moderada (usuários aceitam um breve atraso de verificação)
Demandas de QPS mais baixas (a verificação é uma tarefa relativamente infrequente)
As consequências de uma identificação incorreta são significativas, tornando a precisão a prioridade inegociável, mesmo ao custo de algum desempenho.
O fio condutor: encontrar seu equilíbrio
Esses exemplos contrastantes destacam por que uma abordagem única para configuração de busca vetorial fica aquém. Cada aplicação se situa em algum ponto desse espectro, exigindo otimização cuidadosa de recall, latência e QPS com base nas necessidades do negócio.
É exatamente por isso que introduzimos os parâmetros level e enable_recall_calculation — para dar aos desenvolvedores as ferramentas para encontrar seu equilíbrio ideal.
Apresentando os recursos de ajuste de precisão do Zilliz Cloud
O parâmetro level: refinando a precisão da busca
O parâmetro level oferece controle simples, mas poderoso, sobre a precisão da busca, com valores variando de 1 a 10:
| Valor do Nível | Ideal Para | Recall Típico | Impacto no Desempenho |
|---|---|---|---|
| Mais baixo (1-3) | Aplicações focadas em desempenho | 90-97% | Impacto mínimo na latência, maior QPS |
| Médio (4-7) | Aplicações equilibradas | 97-99.5% | Impacto moderado na latência, bom QPS |
| Mais alto (8-10) | Aplicações críticas em precisão | 99.5%+ | Maior latência, QPS reduzido |
Estendemos o limite superior de 5 para 10 com base no feedback dos usuários, permitindo uma precisão ainda maior para segurança, controle de riscos e muitos outros cenários que exigem precisão.
Observe que uma taxa de recall mais alta nem sempre é boa para seus casos de uso. Por exemplo, se level=3 ou 5 satisfaz as necessidades, aumentar ainda mais o nível levaria apenas a uso desnecessário de recursos e maior latência.
O parâmetro enable_recall_calculation: Validando sua precisão
Embora o parâmetro level ajuste a precisão da pesquisa, como você verifica se está atingindo sua taxa de recall alvo? É aí que entra enable_recall_calculation.
Quando ativado durante uma operação de pesquisa, este parâmetro:
Estima a taxa de recall real da sua configuração atual
Retorna esse valor junto com seus resultados de pesquisa
Permite decisões orientadas por dados sobre mudanças de configuração
Este cálculo único ajuda você a validar se suas configurações atuais atendem aos seus requisitos de precisão sem exigir ferramentas externas de benchmark.
Implementação prática: um guia passo a passo
Vamos percorrer um exemplo real de otimização de uma implementação de pesquisa vetorial com esses novos recursos.
Pré-requisitos
Uma instância do Zilliz Cloud (inscreva-se aqui se ainda não o fez)
PyMilvus 2.5.4 ou posterior
Seu conjunto de dados de teste (usaremos o conjunto de dados cohere-1M neste exemplo)
Etapa 1: Defina seus requisitos
Antes de mergulhar no ajuste de parâmetros, defina claramente seus requisitos:
Taxa de recall alvo: 99.5%
Latência máxima aceitável: 5ms
Outras considerações: QPS, utilização de recursos
Padrão de consulta: pesquisa vetorial pura
TopK: 16,000
Etapa 2: Importe dados de teste
Para testes precisos, você deve usar dados que se assemelhem bastante ao seu ambiente de produção. Neste exemplo, usaremos o conjunto de dados cohere-1M e o importaremos para o Zilliz Cloud via VectorDBBench.
Etapa 3: Estimando a taxa de recall de linha de base
Vamos primeiro verificar como as configurações padrão (level=1) se comportam:
search_params = {
"params": {
"level": 1,
"enable_recall_calculation": True
}
}
res = client.search(
collection_name = "ZillizCloudVectorDBBench",
# test data
data = [
[0.22252834,0.26758388,0.3414864,0.31775144,0.25819996,-0.06176423,0.60313016,-0.31930527,-0.05070293,0.80085576,-0.7066278,-0.14704825,0.07324219,-0.051405758,0.24823247,0.20365287,-0.005265507,0.24754052,0.06302843,-0.24397966,-0.2805658,0.543768,0.018544307,0.14154078,0.03093845,-0.25058296,-0.61569184,0.08389459,-0.27519965,0.2121497,0.26527727,-0.03291658,-0.2631627,0.026973603,-0.22165383,0.38862047,0.012616088,-0.066382475,-0.013436819,-0.59001106,-0.08682751,0.13704056,0.08583454,0.0802483,0.01096984,0.20214474,0.11156094,0.5482859,-0.0807617,-0.16539982,-0.29261217,0.08269717,0.03385099,-0.48874223,-0.013168857,0.01616468,-0.6270225,0.13169415,1.0166928,0.6573267,0.40487188,0.2235163,-0.68331105,-0.24911362,-0.1763628,0.34692895,0.077760294,0.96388775,-0.10841275,0.3977706,0.08965021,0.29019687,0.106024966,0.11854912,-0.070255764,-0.24960922,-0.5354312,-0.70186704,0.25364435,0.43369204,0.42516047,0.15078346,-0.35151976,0.4886603,0.37026608,0.39485633,-0.046821307,0.18807457,0.13850202,-0.15630293,0.1469321,0.13219471,-0.31201053,0.22278261,-0.23063508,0.42379102,0.66762435,-0.11903275,0.22101034,-0.10102379,0.21675844,-0.2571628,-0.26546052,0.6185626,0.241754,-0.15792729,-0.37045696,0.23996626,-0.27011713,-0.11512929,0.31577003,0.11118326,-0.76019096,-0.1682175,0.5329204,0.3614348,-0.2348311,0.108711846,0.068240434,-1.3124024,0.20385003,0.6049856,-0.2924265,-0.07293733,0.5181119,0.15742214,0.7537572,-0.656809,0.57201636,0.09775318,0.5414663,-0.53804034,-0.0802571,0.62367904,0.023681821,0.6950041,0.3207407,0.36089638,0.53875273,-0.7288189,-0.12956187,0.15076943,0.057313107,0.41065657,-0.00928412,0.090415776,0.18091775,-0.010793066,-0.010142676,-0.20986095,-0.3740831,0.25086942,-0.31494206,0.17761512,0.04850758,-0.06098805,0.7605751,0.3038707,0.5178377,-0.5769507,0.80365956,0.22879237,-0.34868854,-0.2688102,-0.20910782,0.3392469,0.2990533,0.5502763,0.6561665,0.04177933,-0.45408615,-0.055697974,0.05596695,-0.22720425,0.63778013,-0.1921504,-0.16227654,-0.053658817,0.04536426,-0.28570235,0.30350783,0.5217574,0.0025516534,0.10135456,0.5973671,0.09276529,0.7803261,0.45648357,-0.21722879,-0.3496141,0.18574907,0.1729008,0.6754883,-0.5101994,0.16308193,-0.32053986,-0.0013728795,0.0371755,-0.114131995,0.19870742,0.3973309,0.17016156,0.016581284,0.3074155,0.32889378,-0.6682561,0.36933577,0.45571226,-0.19315217,-0.5065343,0.15996625,0.026897952,0.046443015,0.2667398,-0.18946062,-0.4283052,-0.44281873,-0.24062063,0.41703427,-0.30064407,0.35975343,0.31060407,0.18125875,0.14912511,0.10962614,-0.06901708,-0.2846222,-0.027887726,0.037055127,0.031954445,0.56672156,-0.0863331,0.1497875,-0.1635759,-0.25121027,0.6303942,0.17385906,0.4313834,0.15800661,-0.6267578,-0.03539913,0.32520285,0.42759246,0.24401832,0.2115575,-0.8652025,0.13317755,-0.5719402,0.17294376,-0.12595764,0.34818307,0.24802469,0.05904272,0.1538172,-0.57994705,0.2582915,0.45511153,-0.44164076,-0.074042775,0.04943926,-0.1648779,0.3280813,0.5601293,-0.0018850226,-0.140464,-0.07845455,0.44026145,0.56197315,0.1462102,-0.18595229,0.014953136,0.46956787,-0.14819877,0.1859354,0.019512085,-0.01712815,0.5366789,0.7835224,-0.7423546,0.6503855,0.44647282,0.3631722,-0.66614413,0.10151727,-0.1348695,0.32992417,0.10387001,-0.26746857,-0.33413792,-0.5662058,0.36110422,0.7741211,-0.039930806,-0.15249825,0.09454683,0.4891987,0.0062028184,0.06745152,0.55465925,-0.06739082,0.5588079,-0.43696547,0.555966,0.56702,0.056295626,-0.62005293,-0.3722073,0.21030217,-0.017268468,0.95288086,0.51696795,-0.25066343,-0.3169728,0.42543235,0.31396082,0.17551036,0.3922707,0.07407632,0.91187936,0.38888615,-0.12070266,0.011815081,-0.45720986,0.04727247,0.62094855,-0.45443076,0.16062841,-0.40287957,-0.55417335,-0.3830013,0.055438586,0.1718703,-0.6422826,0.22917171,-0.5290951,0.1585279,0.07934802,0.50577295,-0.035466444,0.088082545,0.5693788,0.11773129,0.1821725,0.41347143,-0.2278498,0.50422746,0.29794943,-0.9369089,0.47065943,0.28594327,-0.6866015,-0.63375616,-0.15243755,-0.46409172,-0.4630497,-0.25025153,0.6375835,0.54886156,0.19831929,-0.03725618,0.20592122,0.36338213,0.31409082,-0.05410012,0.14887711,0.09740482,0.05067692,0.14775206,0.28025475,-0.34377113,-0.27423778,-0.354568,-0.20043556,0.3899774,-0.19812085,-0.36292082,-0.18255037,0.07038237,0.642794,-0.060884897,0.2948623,0.68766963,0.6928454,-0.3849391,0.17996079,0.12549743,0.10299729,-0.25861496,-0.09246836,-0.32353002,-0.01378604,-0.095313616,-0.04558251,0.20014873,-0.4066689,-0.08052519,-0.4618455,0.37693843,0.45283204,-0.114583425,0.050728872,0.13196129,-0.1941961,-0.11727777,3.9586966,0.05150596,0.11701303,0.5739518,0.07567582,0.48247826,0.25156844,0.38180268,0.12796494,0.009531708,-0.04081167,-0.30954623,-0.035167653,0.43064785,0.24091315,-0.11113215,0.027972942,0.3501582,0.54151994,0.14281327,-0.6420307,0.48611403,0.5221564,0.47878447,0.8510151,0.5528693,0.27463847,0.7548287,0.760392,0.4057206,0.5247366,0.6815815,0.46189928,-0.0665814,0.29575244,-0.13240056,0.44400057,-0.26975283,-0.15510945,0.15475176,-0.46221858,0.054507546,0.14640503,0.66453534,0.19300742,-0.3626597,-0.16279799,0.3795997,0.122737944,-0.20419496,0.18285695,0.027228406,-0.22584598,-0.16478994,0.28747237,0.53937024,0.44095138,0.6340223,-0.41380823,0.38367343,0.39497304,-0.043954037,0.38885015,-0.33315817,-0.4766579,0.17371525,-0.23392603,0.7948543,0.3054392,-0.72041094,0.2532946,0.415873,0.80443436,-0.34634262,-0.4886025,0.30351955,-0.049782824,-0.47253707,-0.11401102,-0.096243046,0.19083612,-0.34427363,-0.24545296,0.5773733,0.16357873,0.38620606,0.39995435,-0.65907687,0.6957725,0.24120355,0.34054404,-0.039899644,0.80393964,0.06337182,0.14144897,0.117613785,-0.019442292,-3.7490542,-0.38971332,0.14894387,-0.61240107,0.19039957,0.23817067,0.022639165,0.015894404,-0.6198486,0.14320132,0.041371442,-0.30882874,-0.30676636,1.0463533,-0.034157425,0.31748047,0.4891939,0.5333419,-0.3289819,0.14962271,0.2807266,0.35519713,0.4001028,-0.18559772,-0.7066097,0.14664957,-0.565848,0.013109448,-0.18452193,-0.07372118,0.28156808,-0.36035228,0.8867393,-0.16306667,-0.04191513,0.4594507,0.43135175,-0.091903865,-0.042651527,-0.32555583,-0.19054003,-0.06525034,0.16911364,0.04686202,-0.038171988,0.1336097,0.3761719,-0.050084345,-0.2679286,0.64759475,0.7107872,0.2074471,-0.27312976,0.40090975,0.5491712,-0.10747743,0.74496686,0.18130445,-0.09431538,0.19524746,-0.21418755,-0.12488151,0.15227054,-0.3852693,-0.7784234,-0.14571632,0.041122317,-0.16407914,0.03949264,-0.1925929,0.32901394,0.12069722,0.23391949,-0.16763128,-0.12962814,0.5088096,0.21486548,-0.20993523,0.603585,0.24633685,-0.14029086,-0.27401388,-0.49189645,-0.10249644,2.3196032,-0.12417316,2.1448603,0.058190174,-0.6551869,0.6827868,0.6356786,0.7710372,0.3722568,0.8363127,0.3799041,0.26085538,-0.20764771,0.512162,0.08349497,-0.15835808,0.5738307,-0.66654295,0.18993358,0.32188657,0.0764867,0.64592606,-0.2310478,0.18350935,-0.3915338,0.028645294,-0.101273224,0.8696747,-0.50792813,-0.39119712,-0.30162883,0.7319297,0.71813834,0.39383802,-0.012138247,0.3298783,0.23386809,4.5470805,-0.049004212,0.107414484,0.052308656,0.2678271,-0.15366946,0.5438965,-0.47809094,-0.14649442,0.022792917,0.1358324,-0.4503206,0.57014287,-0.13368002,0.23805767,-0.22125027,-0.08700341,0.045676652,0.16678812,-0.27974084,0.45245427,-0.2107062,0.6667994,0.036875203,0.54632777,0.20104687,0.5349449,0.06913179,-0.086024776,0.76876926,0.16203642,5.155099,0.2797164,0.21450946,-0.17529553,-0.038863413,0.5156995,0.08603405,-0.516439,-0.35604522,0.10131945,0.008194211,0.084706515,-0.34049395,0.21572115,-0.83385843,-0.046860088,-0.48247585,0.023293016,0.22008015,-0.5305121,0.5061096,0.0183293,0.1326365,0.22057603,-0.43027383,-0.3885953,0.1500542,-0.1449458,-0.38747045,0.2789606,0.27069542,-0.37978157,-0.58541,0.5139468,-0.60586643,-0.5236463,0.22003366,0.15764758,0.3512009,0.13694952,0.7772281,0.28431293,0.113065295,0.14233269,-0.047996823,0.0024461043,0.06218189,-0.28065726,-0.2061346,-0.36278206,0.24291486,-0.0869041,0.7448049,0.36513415,0.61559093,0.42820337,0.41123256,-0.32082868,-0.10876272,-0.028618973,0.6750199,-0.048880983,-0.12521495,0.1926665,0.6695621,0.21937566,0.46856737,0.30544627,0.2650348,-0.11578811,-0.15696093,-0.047148716,0.19283816,0.12149068,-0.03274016,0.021503512,0.008024155,0.19709297,0.15727529,0.14134975,-0.16997191,-0.063695885,-0.39591065,-0.11891319,-0.04673462,0.16978487,-0.09345571,0.11924938,0.13301763,-0.2266567,0.4164705,0.3571622,0.09038913,0.18044233,0.09119875,-0.23754075,0.45051736,0.35435763,0.20957275,0.5704436,-0.36682,0.26963162,0.15532929,-0.24306794,0.17486432,0.39116114,0.12234816,0.21448524,-0.019066956,-0.09756305,0.4465544,0.3394048,-0.7088385,-0.5032021,0.03529406]
],
limit = 16000,
search_params = search_params
)
print(f"recall: {res.recalls}")
A saída é:
recall: [0.9886875152587891]
Com um recall de 98,87%, estamos abaixo da nossa meta de 99,5%. Vamos medir a latência também (opcional):
search_params = {
"params": {
"level": 1,
# "enable_recall_calculation": True
}
}
start_time = time.perf_counter()
client.search(collection_name = "ZillizCloudVectorDBBench",
data = data,
limit = limit,
search_params = search_params
)
end_time = time.perf_counter()
elapsed_time = end_time - start_time
print(f"latency at level=1: {elapsed_time:.2f} seconds")
E a saída é:
latency at level=1: 0.03 seconds
Etapa 4: Encontrar o valor ideal de Level
Agora, vamos testar sistematicamente diferentes valores de level para encontrar a configuração ideal:
recall at level=1: [0.9886875152587891]
latency at level=1: 0.03 seconds
...
recall at level=6: [0.9947500228881836]
latency at level=6: 0.04 seconds
recall at level=7: [0.9961249828338623]
latency at level=7: 0.04 seconds
...
recall at level=10: [1.0]
latency at level=10: 0.06 seconds
Os experimentos mostram que level=7 atende à nossa meta de recall (99,61% > 99,5%) mantendo a latência dentro de limites aceitáveis.
Etapa 5: Monitorar o desempenho do sistema (opcional)
O Zilliz Cloud oferece um conjunto abrangente de métricas do sistema para ajudar você a monitorar o desempenho de forma eficaz. Ao ajustar o parâmetro level e aplicar novas configurações, o Zilliz Cloud fornece acompanhamento em tempo real de métricas-chave como QPS e latência. Isso permite avaliar o impacto dos ajustes de parâmetros no desempenho geral do sistema, possibilitando uma tomada de decisão mais informada.
Etapa 6: Validar com carga de trabalho de produção
Para validação em produção, recomendamos:
Executar testes com vetores de consulta representativos
Monitorar o painel do Zilliz Cloud para:
Latência média
QPS
Utilização de recursos
Etapa 7: Realizar benchmarking abrangente (opcional)
Para uma avaliação mais rigorosa, usamos o VectorDBBench, uma ferramenta de benchmarking de código aberto, para testar várias configurações com 1.000 operações de busca:
| level | Latência média(ms) | QPS | Recall do VectorDBBench | Recall do Zilliz Cloud |
|---|---|---|---|---|
| 1 | 3.1 | 1266 | 0.9519 | 0.953 |
| 2 | 3.2 | 1080 | 0.9644 | 0.9669 |
| 3 | 3.4 | 972 | 0.9728 | 0.9755 |
| 4 | 3.6 | 814 | 0.9816 | 0.9846 |
| 5 | 3.9 | 704 | 0.9871 | 0.99 |
| 6 | 4.4 | 549 | 0.9916 | 0.995 |
| 7 | 5 | 448 | 0.9936 | 0.9971 |
| 8 | 5.4 | 375 | 0.9945 | 0.9983 |
| 9 | 5.9 | 340 | 0.9952 | 0.9991 |
| 10 | 6.3 | 296 | 0.9958 | 0.9995 |
Tabela 1: Resultados dos testes com base em 1.000 operações de busca do VDBBench
Esses resultados de benchmarking confirmam que:
As estimativas de
enable_recall_calculationcorrespondem de perto ao desempenho real de recallUm recall mais alto vem ao custo de maior latência e QPS reduzido
Para nossos requisitos-alvo,
level=7oferece o equilíbrio ideal
É importante observar que melhorias no recall geralmente vêm ao custo de maior latência e QPS reduzido. Se o recall atual atende às suas necessidades de negócio, mas o QPS está abaixo do ideal, você pode considerar aumentar os recursos de CU ou adicionar réplicas para melhorar a taxa de transferência.
Etapa 8: Confirmar a configuração ideal
Com base nos resultados experimentais, a configuração ideal neste exemplo é level=7, que oferece 99,6% de recall mantendo uma latência aceitável.
Conclusão
Encontrar o equilíbrio perfeito entre precisão de busca e desempenho há muito tempo é um desafio para os usuários de bancos de dados vetoriais. Com os novos parâmetros level e enable_recall_calculation do Zilliz Cloud, os desenvolvedores agora têm ferramentas poderosas para otimizar suas implementações de busca vetorial de acordo com seus requisitos específicos.
Seja criando sistemas de recomendação que priorizam velocidade, aplicações de segurança que exigem alta precisão, ou qualquer coisa entre esses dois extremos, esses recursos permitem que você:
Ajuste com precisão a exatidão da busca para atender aos seus requisitos
Valide as taxas reais de recall com estimativa integrada
Tome decisões baseadas em dados sobre compensações de configuração
Estamos comprometidos em continuar aprimorando o Zilliz Cloud com recursos que tornam a busca vetorial mais poderosa, flexível e fácil de otimizar. Esses novos parâmetros são apenas o começo da nossa jornada para capacitar desenvolvedores a criar aplicações excepcionais impulsionadas por IA.
Pronto para começar a otimizar suas buscas vetoriais? Cadastre-se no Zilliz Cloud hoje mesmo ou faça login na sua conta existente para explorar esses novos recursos.
Ken ZhangKen Zhang is a Senior Product Manager at Zilliz, leading the development of the Milvus vector database by setting its strategic direction and key features. Prior to Zilliz, he served as a kernel engineer at SAP HANA and enhanced his product management skills at PingCAP. Ken holds a master's degree from Fudan University and has over eight years of experience specializing in database development and big data infrastructure management.
Chris GaoChris Gao is a senior software engineer at Zilliz, focusing on the vector search engine optimization in the Milvus vector database and Zilliz Cloud. Prior to Zilliz, he served as a software engineer at Bytedance and optimized the execution of Spark ETL jobs. Chris holds a master's degree from Shanghai Jiao Tong University.
Continue lendo
A Few Notes from Databricks Data + AI Summit 2026: Why the Data Layer Matters Again
James Luan shares notes from Databricks Data + AI Summit 2026 on why production AI is pushing the data layer back to the center of infrastructure.
My Wife Wanted Dior. I Spent $600 on Claude Code to Vibe-Code a 2M-Line Database Instead.
Write tests, not code reviews. How a test-first workflow with 6 parallel Claude Code sessions turns a 2M-line C++ codebase into a daily shipping pipeline.
Top 10 Context Engineering Techniques You Should Know for Production RAG
A practical guide to context engineering for production LLM systems, covering RAG, context processing, memory, agents, and multimodal context.