向量检索时，topk的值在<=255 和 >255时检索出来的数据不一致

[yujiefangfei]

召回100个

召回255个

召回300个

召回数量大于255之后，就少了部分数据，并且是分数较高的
index_type="GPU_IVF_FLAT"
metric_type="IP"

[yhmo]

调用search的时候设置了哪些参数？

[yujiefangfei]

{"collectionName":"image_vector","consistencyLevel":"EVENTUALLY","expr":"","gracefulTime":5000,"guaranteeTimestamp":1,"ignoreGrowing":false,"metricType":"IP","nQ":1,"outFields":["itemId"],"params":"{"nprobe":50}","partitionNames":[],"plType":"None","roundDecimal":-1,"topK":300,"travelTimestamp":0,"vectorFieldName":"vectors","vectors":[[0.017668035,-0.011638614,0.024306279,-8.159827E-4,-0.011512056,-0.04670389,0.005944621,-0.09777747,0.030088648,0.034337875,-0.04002155,-0.042296417,-0.04761632,-0.04723062,0.040329184,0.014475837,0.027402177,0.0076565985,-5.322728E-4,-0.022145819,0.05934978,0.019097492,0.02796327,0.058844022,0.008410808,0.017110607,-0.008204461,0.03606606,0.0038372192,-0.008156777,0.0012446017,0.037977755,0.019522985,-0.030655377,-0.06208211,0.017627055,0.038120735,0.02414759,-0.01910181,0.0230862,0.016248414,0.01722224,0.03413119,0.0046254937,0.077597894,0.0075641144,0.0068683825,0.022819256,-0.0030075945,0.01233097,-0.0069075027,-0.014545523,0.016139522,-0.013299041,0.004876295,-0.023910781,0.0315721,0.03842436,-0.015659466,0.023017297,0.02240377,-0.0062213186,-0.012235059,-0.0019695985,-0.050301652,0.023416827,-0.0025450897,-0.051943112,0.01868323,-0.023431327,-0.021281824,-0.018026534,-0.022435231,-0.009036911,-0.02637883,-0.008635563,0.041199394,0.0021149025,-0.077329256,-0.03276411,-0.0037962478,-0.0050173174,0.029868484,0.010562843,0.027329389,-0.00644208,0.01987515,-0.013023183,0.030968143,-0.023360347,0.02956031,-0.045063663,-0.15289041,0.0029526749,-0.012282223,-0.029609561,0.02356125,0.057649855,0.03379,0.021536812,0.037736733,-0.024432527,0.050838746,-0.06716974,-0.013910005,0.014322128,0.008502053,-0.004088294,-0.042971656,0.03705169,0.016230943,-0.033402015,0.0040355693,-0.017054195,0.05262078,0.029554145,-0.015264607,-0.011354771,-0.019721113,0.01843935,0.025166927,-0.017649624,0.0072383103,-0.052367028,-5.436469E-4,-0.03172969,-0.02289103,-0.013947928,0.014396659,0.028095711,0.014803691,0.027969087,0.021011084,0.5172036,0.024488872,0.008355814,0.022979895,0.0052811513,0.022175645,0.040546816,0.032257486,-0.046046928,0.02158695,0.024567218,-0.028188871,0.04988444,-0.013852378,-0.008409448,-0.00762551,0.0060273497,-0.004855542,0.02785248,0.015851768,0.004643827,0.015163031,-0.0037883816,0.073196284,0.0054740226,0.013217066,-0.013885582,-0.019097479,0.02234225,-0.017221127,-0.0036323778,0.028212178,0.0068881577,0.012323303,-0.039319776,0.031155871,0.0042256885,-0.004152213,-0.01351092,-0.044099256,-0.020591626,-0.02273089,-0.011398132,-0.022189641,-0.051650327,0.066636495,0.010160067,0.04712277,-0.014591111,-0.033193354,0.015979145,-0.0391487,0.0070581874,0.010210665,-0.07038592,0.02387716,-0.06863544,0.017961815,-0.055901762,1.6715514E-4,0.02786131,-0.0129243955,-0.030125296,0.05901195,0.0031134845,-0.033100933,0.014692775,-0.029201558,-0.006460085,0.052052222,0.048771434,0.0063536256,0.022612762,-0.0043455046,-0.02834659,-0.02173262,0.058052603,0.03890874,5.968382E-4,0.023382327,-0.029084703,-0.012004477,0.02048661,-0.012613855,0.025050618,-0.074945934,-0.008408048,-0.03793687,-0.030774103,-0.03756076,0.01703788,0.030099723,0.0075263376,-0.00829761,0.02324529,0.013191792,0.03262075,0.042174026,-0.004738727,0.022584906,0.0136569245,-0.025416268,0.007692907,-0.0050902166,0.0013539789,0.014187741,-0.020548131,0.012419947,-0.022472449,0.017208442,0.013050185,-0.0010343939,0.03025407,-0.004761997,-0.04150773,-0.004658568,-0.026445033,-0.006435253,-0.0019200109,-0.011933866,-0.033691514,-0.009900709,0.017700229,-0.0026847946,-0.034204405,0.019556466,0.07175781,-0.01989661,-0.030586908,0.033149667,-0.0056692255,-0.0058995937,-0.0026554721,-0.0076181088,0.031657096,0.0034733564,0.014260821,0.0057822233,-0.027190493,-0.01925845,-0.047812454,-0.010480365,-0.028384112,-0.01901553,0.035757188,0.02117641,0.006059255,-0.017436916,0.015792022,0.005173357,0.015606501,0.033343866,0.004002309,0.037286133,-0.015747515,0.028561335,0.03234052,0.049641117,-0.0022899138,-0.027704313,0.027959535,-0.013309888,-0.008865395,-0.030613516,-0.0063104113,0.015824126,0.024398226,0.009954333,-0.033575766,-0.009599956,-0.030173887,0.024239145,-0.05634807,-0.038456045,0.0035270508,0.024057304,0.012369527,0.0070071653,-0.011279669,0.5174007,-0.02690695,0.0052805943,0.009783851,0.011689357,0.054027386,0.0028496308,0.011519891,0.040626355,0.039220437,-0.0012240659,-0.022301232,0.02936027,0.023665851,0.028187696,0.020992355,0.00949348,-0.12665568,0.009248812,0.03536245,0.039241917,0.03054757,0.019003617,0.029330065,-0.0050591473,-0.056128282,-0.0026003753,-0.039258413,0.0036918677,-0.0017430717,0.038119778,0.018399425,0.056148022,-6.7933206E-4,0.0115155075,0.015150226,0.006618172,0.0068444624,7.085489E-4,0.00821716,-0.03035542,0.07096207,-0.0021532723,-0.038746584,-0.027487256,-0.0022355393,-0.008440537,-0.0153884925,-0.0027719967,0.0063202037,-0.044214644,-0.037376773,0.027540145,-0.035442263,0.036101002,-0.021967642,0.04384717,0.034554575,0.007707845,-0.02494999,0.023765748,-0.0057392647,0.010823522,-0.0474743,0.015241399,0.02554453,-0.013836896,-0.008772597,-0.004970209,-0.031223826,-0.01700619,-0.013480958,0.022796769,-0.0036950468,-0.043071218,0.049778562,0.02583067,0.011173396,-0.04399071,0.0024797115,-0.026669323,-0.051202476,-0.009828243,0.009753357,-0.012874793,-0.008741036,0.027752494,0.07329693,-0.008993937,-0.05123615,0.041206192,-6.7676004E-4,-0.021655088,-0.026744574,0.011379195,0.0034902417,-0.008179322,0.020669319,0.006858075,0.019483095,-0.03647249,0.032377474,-0.01403606,-0.040062565,0.002599681,-0.04708046,0.016308738,-0.02281171,-0.01930235,0.008961924,0.0113032665,-0.017498275,-0.02269865,-0.0047564832,-0.025777439,0.0010427106,-0.018787382,0.012619558,-0.031724308,0.005315729,0.027676048,0.047386445,0.009265821,-0.029162142,-0.013598345,-0.009299134,-0.04823087,-0.0056503974,-0.0024649845,5.138708E-4,0.015265762,-0.037234575,-0.009478332,0.034773182,-0.034984276,-0.04055323,0.0041441345,0.0559374,-0.04646459,-0.022517709,0.017486839,0.037722524,0.0088474285,-0.025380386,-0.053030733,-0.0059654233,0.0076552914,-0.018016063,0.017573303,-0.043401983,-0.0116437115,0.030270964,-0.042263243,-0.030951822,-0.01583371,-0.047209047,0.041586302,-0.07996849,0.014103748,-0.015006323,-0.024220144,0.001366454,0.004096906,0.0018538237,0.03446947,-0.005179594,-0.030328942,0.05671469,0.029299632,-0.085199796,0.019967102,-0.043227397,-0.01325435,0.035183642,0.034116834,0.039243646,-0.009908469,0.0021833647,0.006454289,0.017808946,-0.012496346,0.095684536,-0.01804116,-0.026245957,0.060204584,-0.010393611,-0.031913023,0.018770011,0.01723801,0.03154286,-0.051888432,0.0367296,0.038808692,0.008600572,-0.029877158,0.027740197,0.01510123,-0.018043062,0.022769246,0.013899986]]}

[yujiefangfei]

创建索引参数如下：
index_params.add_index(
index_name=name + "_ivf_flat",
field_name="vectors",
index_type="GPU_IVF_FLAT",
metric_type="IP",
dimension=512,
params={"nlist": 512}
)

[yhmo]

参数设置都挺常规的，看起来没有什么可疑之处。
已知的问题里并没有关于CPU IVF_FLAT索引结果不一致的问题，所以可能是GPU索引的问题。

有可能相关的issues:
#40573
#36607

这些关于GPU_IVF_FLAT索引结果有误的问题是在2.4的中期版本2.4.12 ~ 2.4.14发现的，所以如果你的milvus版本是2.4中早期版本，可以升级到2.4的最后一个版本。
如果你的milvus版本已经是最新的2.5，还是去报一个issue吧，我们找相关工程师看看：https://github.com/milvus-io/milvus/issues

你也可以改用CPU IVF_FLAT看看是否有这种问题。