titles
listlengths
1
2
url
stringclasses
222 values
location
stringlengths
59
197
vector
listlengths
384
384
tag
stringclasses
8 values
text
stringlengths
1
961
sections
listlengths
1
3
id
int64
0
18.8k
[ "Explore - Qdrant", "Context search" ]
/documentation/concepts/explore/
html > body > div:nth-of-type(1) > section > div > div > div:nth-of-type(2) > article > p:nth-of-type(40)
[ -0.0028162840753793716, 0.04600208252668381, 0.03242453560233116, 0.06413580477237701, -0.0660071149468422, -0.07343345135450363, 0.11492646485567093, 0.059214998036623, 0.01834903471171856, -0.0051994649693369865, 0.015808872878551483, 0.030076848343014717, 0.02832612209022045, -0.0061395...
p
$$
[ "documentation", "documentation/concepts", "documentation/concepts/explore" ]
7,000
[ "Explore - Qdrant", "Context search" ]
/documentation/concepts/explore/
html > body > div:nth-of-type(1) > section > div > div > div:nth-of-type(2) > article > p:nth-of-type(40)
[ 0.05739787966012955, 0.0771348774433136, -0.046567607671022415, -0.05271374061703682, 0.056607626378536224, 0.08698386698961258, 0.052172522991895676, 0.06096440553665161, 0.01727299764752388, -0.017193978652358055, 0.000010108810784004163, -0.06308406591415405, 0.07994973659515381, 0.0028...
p
\text{context score} = \sum \min(s(v^+_i) - s(v^-_i), 0.0)
[ "documentation", "documentation/concepts", "documentation/concepts/explore" ]
7,001
[ "Explore - Qdrant", "Context search" ]
/documentation/concepts/explore/
html > body > div:nth-of-type(1) > section > div > div > div:nth-of-type(2) > article > p:nth-of-type(40)
[ -0.0028162840753793716, 0.04600208252668381, 0.03242453560233116, 0.06413580477237701, -0.0660071149468422, -0.07343345135450363, 0.11492646485567093, 0.059214998036623, 0.01834903471171856, -0.0051994649693369865, 0.015808872878551483, 0.030076848343014717, 0.02832612209022045, -0.0061395...
p
$$
[ "documentation", "documentation/concepts", "documentation/concepts/explore" ]
7,002
[ "Explore - Qdrant", "Context search" ]
/documentation/concepts/explore/
html > body > div:nth-of-type(1) > section > div > div > div:nth-of-type(2) > article > p:nth-of-type(41)
[ -0.02868049591779709, 0.019762221723794937, -0.06747742742300034, -0.08569545298814774, -0.001752396347001195, 0.09424692392349243, -0.03540996462106705, -0.04215187579393387, 0.0360620841383934, -0.011647721752524376, 0.04643014445900917, 0.033124011009931564, 0.05475998669862747, 0.05363...
p
Where $v^+_i$ and $v^-_i$ are the positive and negative examples of each pair, and $s(v)$ is the similarity function.
[ "documentation", "documentation/concepts", "documentation/concepts/explore" ]
7,003
[ "Explore - Qdrant", "Context search" ]
/documentation/concepts/explore/
html > body > div:nth-of-type(1) > section > div > div > div:nth-of-type(2) > article > p:nth-of-type(42)
[ -0.016355641186237335, 0.00762534886598587, -0.009050583466887474, 0.0035829832777380943, 0.058938127011060715, -0.012950923293828964, -0.01973547600209713, -0.02662811428308487, 0.07600216567516327, -0.0654485896229744, 0.026526523754000664, -0.010834446176886559, 0.09623891115188599, -0....
p
Using this kind of search, you can expect the output to not necessarily be around a single point, but rather, to be any point that isn’t closer to a negative example, which creates a constrained diverse result.
[ "documentation", "documentation/concepts", "documentation/concepts/explore" ]
7,004
[ "Explore - Qdrant", "Context search" ]
/documentation/concepts/explore/
html > body > div:nth-of-type(1) > section > div > div > div:nth-of-type(2) > article > p:nth-of-type(42)
[ -0.10629843920469284, -0.05808241292834282, -0.03347235918045044, 0.034534867852926254, 0.054074596613645554, 0.009514844045042992, 0.017439192160964012, 0.07399658858776093, 0.0008232271648012102, -0.1152774915099144, 0.018329406157135963, 0.06397642195224762, 0.07372768968343735, 0.06382...
p
So, even when the API is not called recommend, recommendation systems can also use this approach and adapt it for their specific use-cases.
[ "documentation", "documentation/concepts", "documentation/concepts/explore" ]
7,005
[ "Explore - Qdrant", "Context search" ]
/documentation/concepts/explore/
html > body > div:nth-of-type(1) > section > div > div > div:nth-of-type(2) > article > p:nth-of-type(36)
[ 0.009657991118729115, -0.044360946863889694, 0.005418023094534874, 0.06800470501184464, -0.05134975165128708, -0.01160994078963995, 0.020690567791461945, 0.02639947272837162, -0.024253476411104202, -0.0060655102133750916, 0.057527314871549606, 0.018698764964938164, 0.05972754582762718, 0.0...
p
Example:
[ "documentation", "documentation/concepts", "documentation/concepts/explore" ]
7,006
[ "Explore - Qdrant", "Context search" ]
/documentation/concepts/explore/
html > body > div:nth-of-type(1) > section > div > div > div:nth-of-type(2) > article > aside:nth-of-type(5) > ul > li:nth-of-type(1)
[ -0.05477146431803703, 0.10724236071109772, -0.01764717884361744, 0.03811099752783775, 0.042935293167829514, 0.03820614516735077, 0.0008718445897102356, -0.06730823218822479, 0.03002895973622799, -0.08695068210363388, 0.03169797360897064, 0.02395111694931984, 0.12125977128744125, -0.0728668...
li
When providing ids as examples, they will be excluded from the results.
[ "documentation", "documentation/concepts", "documentation/concepts/explore" ]
7,007
[ "Explore - Qdrant", "Context search" ]
/documentation/concepts/explore/
html > body > div:nth-of-type(1) > section > div > div > div:nth-of-type(2) > article > aside:nth-of-type(5) > ul > li:nth-of-type(2)
[ 0.06362029910087585, -0.03314540162682533, 0.01409970223903656, -0.053697045892477036, -0.03016839176416397, 0.025520069524645805, -0.06738575547933578, 0.05720176175236702, 0.08540219068527222, 0.03358956798911095, -0.0061394586227834225, 0.11556925624608994, 0.026898033916950226, 0.06651...
li
Score is always in descending order (larger is better), regardless of the metric used.
[ "documentation", "documentation/concepts", "documentation/concepts/explore" ]
7,008
[ "Explore - Qdrant", "Context search" ]
/documentation/concepts/explore/
html > body > div:nth-of-type(1) > section > div > div > div:nth-of-type(2) > article > aside:nth-of-type(5) > ul > li:nth-of-type(3)
[ 0.05482957512140274, -0.01652119867503643, -0.06105709820985794, 0.004029820207506418, -0.01469194795936346, 0.01832425966858864, 0.013383130542933941, 0.048527371138334274, 0.07761535048484802, 0.02501688338816166, -0.0250812117010355, -0.0019830737728625536, 0.019029738381505013, 0.05136...
li
Best possible score is 0.0, and it is normal that many points get this score.
[ "documentation", "documentation/concepts", "documentation/concepts/explore" ]
7,009
[ "Filtrable HNSW - Qdrant" ]
/articles/filtrable-hnsw/
html > body > div:nth-of-type(1) > div:nth-of-type(1) > div > section > article > ul:nth-of-type(1) > li:nth-of-type(1)
[ 0.0409557968378067, 0.01968824490904808, -0.026446374133229256, 0.020037280395627022, 0.018771959468722343, -0.0022932016290724277, 0.06403844058513641, -0.018167899921536446, -0.0274504367262125, -0.03519021347165108, -0.031034929677844048, -0.027398349717259407, 0.007306049577891827, -0....
li
Home
[ "articles", "articles/filtrable-hnsw" ]
7,010
[ "Filtrable HNSW - Qdrant" ]
/articles/filtrable-hnsw/
html > body > div:nth-of-type(1) > div:nth-of-type(1) > div > section > article > ul:nth-of-type(1) > li:nth-of-type(2)
[ -0.1359460949897766, -0.018250174820423126, 0.01159171387553215, -0.05307161435484886, 0.02118542790412903, -0.08700745552778244, -0.006286499090492725, -0.011478456668555737, 0.02877473458647728, -0.03753074258565903, 0.028667796403169632, -0.06563723087310791, 0.017261670902371407, 0.044...
li
/
[ "articles", "articles/filtrable-hnsw" ]
7,011
[ "Filtrable HNSW - Qdrant" ]
/articles/filtrable-hnsw/
html > body > div:nth-of-type(1) > div:nth-of-type(1) > div > section > article > ul:nth-of-type(1) > li:nth-of-type(3)
[ -0.007825803011655807, 0.07636253535747528, -0.012029952369630337, 0.11870371550321579, -0.02357262559235096, 0.07180462777614594, -0.005251061171293259, 0.0708870068192482, -0.012350792065262794, 0.0705045759677887, 0.03871509060263634, 0.07755763083696365, -0.004780877847224474, 0.057188...
li
Articles
[ "articles", "articles/filtrable-hnsw" ]
7,012
[ "Filtrable HNSW - Qdrant" ]
/articles/filtrable-hnsw/
html > body > div:nth-of-type(1) > div:nth-of-type(1) > div > section > article > ul:nth-of-type(1) > li:nth-of-type(2)
[ -0.1359460949897766, -0.018250174820423126, 0.01159171387553215, -0.05307161435484886, 0.02118542790412903, -0.08700745552778244, -0.006286499090492725, -0.011478456668555737, 0.02877473458647728, -0.03753074258565903, 0.028667796403169632, -0.06563723087310791, 0.017261670902371407, 0.044...
li
/
[ "articles", "articles/filtrable-hnsw" ]
7,013
[ "Filtrable HNSW - Qdrant" ]
/articles/filtrable-hnsw/
html > body > div:nth-of-type(1) > div:nth-of-type(1) > div > section > article > ul:nth-of-type(1) > li:nth-of-type(5)
[ -0.07745631039142609, 0.006582682486623526, -0.0028027277439832687, 0.04357463866472244, 0.01641889102756977, 0.047957643866539, 0.07221507281064987, 0.0343962125480175, -0.12438082695007324, -0.08718494325876236, 0.008216324262320995, -0.050789669156074524, -0.05436446890234947, 0.0738753...
li
Filtrable HNSW
[ "articles", "articles/filtrable-hnsw" ]
7,014
[ "Filtrable HNSW - Qdrant" ]
/articles/filtrable-hnsw/
html > body > div:nth-of-type(1) > div:nth-of-type(1) > div > section > article > h1
[ -0.07745631039142609, 0.006582682486623526, -0.0028027277439832687, 0.04357463866472244, 0.01641889102756977, 0.047957643866539, 0.07221507281064987, 0.0343962125480175, -0.12438082695007324, -0.08718494325876236, 0.008216324262320995, -0.050789669156074524, -0.05436446890234947, 0.0738753...
h1
Filtrable HNSW
[ "articles", "articles/filtrable-hnsw" ]
7,015
[ "Filtrable HNSW - Qdrant", "Filtrable HNSW" ]
/articles/filtrable-hnsw/
html > body > div:nth-of-type(1) > div:nth-of-type(1) > div > section > article > p:nth-of-type(1)
[ -0.12239910662174225, -0.08692429214715958, -0.06637804210186005, -0.04906853660941124, 0.04147666320204735, -0.024038132280111313, 0.0293487049639225, -0.020791111513972282, 0.01123040821403265, -0.06526719033718109, -0.03302404284477234, -0.013492883183062077, -0.0015734749613329768, 0.0...
p
If you need to find some similar objects in vector space, provided e.g. by embeddings or matching NN, you can choose among a variety of libraries: Annoy, FAISS or NMSLib.
[ "articles", "articles/filtrable-hnsw" ]
7,016
[ "Filtrable HNSW - Qdrant", "Filtrable HNSW" ]
/articles/filtrable-hnsw/
html > body > div:nth-of-type(1) > div:nth-of-type(1) > div > section > article > p:nth-of-type(1)
[ 0.02924797683954239, -0.07865840941667557, -0.0073363096453249454, -0.0880231186747551, 0.07312949001789093, -0.05289999023079872, -0.10742831975221634, -0.07068334519863129, -0.03217742592096329, 0.031062351539731026, 0.08522619307041168, 0.0043744589202106, 0.014687493443489075, -0.00057...
p
All of them will give you a fast approximate neighbors search within almost any space.
[ "articles", "articles/filtrable-hnsw" ]
7,017
[ "Filtrable HNSW - Qdrant", "Filtrable HNSW" ]
/articles/filtrable-hnsw/
html > body > div:nth-of-type(1) > div:nth-of-type(1) > div > section > article > p:nth-of-type(2)
[ -0.016665145754814148, 0.02192862145602703, -0.05050179734826088, -0.01245387364178896, 0.0033371003810316324, 0.024550115689635277, -0.04730670154094696, -0.046389561146497726, -0.0439378097653389, -0.02572600543498993, 0.011295231059193611, -0.027754278853535652, 0.01955348253250122, 0.0...
p
But what if you need to introduce some constraints in your search?
[ "articles", "articles/filtrable-hnsw" ]
7,018
[ "Filtrable HNSW - Qdrant", "Filtrable HNSW" ]
/articles/filtrable-hnsw/
html > body > div:nth-of-type(1) > div:nth-of-type(1) > div > section > article > p:nth-of-type(2)
[ 0.0026111272163689137, -0.004465417005121708, -0.05219857767224312, -0.038773179054260254, 0.04782692342996597, 0.03183384984731674, 0.027293171733617783, 0.023508915677666664, -0.007780537474900484, -0.1168031245470047, 0.05689743533730507, -0.004613224416971207, 0.0239365566521883, 0.000...
p
For example, you want search only for products in some category or select the most similar customer of a particular brand.
[ "articles", "articles/filtrable-hnsw" ]
7,019
[ "Filtrable HNSW - Qdrant", "Filtrable HNSW" ]
/articles/filtrable-hnsw/
html > body > div:nth-of-type(1) > div:nth-of-type(1) > div > section > article > p:nth-of-type(2)
[ -0.03195617347955704, 0.042831409722566605, -0.03374258428812027, 0.026215633377432823, 0.005030016414821148, 0.00018501032900530845, -0.045997317880392075, -0.010988772846758366, -0.005877416580915451, -0.012207954190671444, 0.0660615786910057, -0.052612531930208206, -0.0029552329797297716,...
p
I did not find any simple solutions for this.
[ "articles", "articles/filtrable-hnsw" ]
7,020
[ "Filtrable HNSW - Qdrant", "Filtrable HNSW" ]
/articles/filtrable-hnsw/
html > body > div:nth-of-type(1) > div:nth-of-type(1) > div > section > article > p:nth-of-type(2)
[ -0.046246711164712906, 0.04988311976194382, 0.018693117424845695, -0.004963013809174299, 0.07539281249046326, 0.05151832476258278, -0.02717060223221779, 0.03994698449969292, -0.08225211501121521, -0.03625159338116646, -0.01704748533666134, 0.010026239790022373, 0.07622344046831131, -0.0240...
p
There are several discussions like this, but they only suggest to iterate over top search results and apply conditions consequently after the search.
[ "articles", "articles/filtrable-hnsw" ]
7,021
[ "Filtrable HNSW - Qdrant", "Filtrable HNSW" ]
/articles/filtrable-hnsw/
html > body > div:nth-of-type(1) > div:nth-of-type(1) > div > section > article > p:nth-of-type(3)
[ -0.08632374554872513, 0.05279216542840004, -0.010401824489235878, -0.029569311067461967, 0.03984856233000755, -0.010663049295544624, -0.0376572422683239, -0.05825084075331688, -0.07316607981920242, -0.005246800370514393, -0.06112644076347351, 0.039846453815698624, -0.007451614364981651, 0....
p
Let’s see if we could somehow modify any of ANN algorithms to be able to apply constrains during the search itself.
[ "articles", "articles/filtrable-hnsw" ]
7,022
[ "Filtrable HNSW - Qdrant", "Filtrable HNSW" ]
/articles/filtrable-hnsw/
html > body > div:nth-of-type(1) > div:nth-of-type(1) > div > section > article > p:nth-of-type(4)
[ -0.018721288070082664, 0.06274110078811646, 0.07361727207899094, 0.0031202281825244427, 0.10567832738161087, -0.07954519242048264, 0.037800662219524384, -0.028395364060997963, 0.05511120706796646, 0.0787733793258667, -0.040071357041597366, -0.017356080934405327, 0.025772644206881523, -0.00...
p
Annoy builds tree index over random projections.
[ "articles", "articles/filtrable-hnsw" ]
7,023
[ "Filtrable HNSW - Qdrant", "Filtrable HNSW" ]
/articles/filtrable-hnsw/
html > body > div:nth-of-type(1) > div:nth-of-type(1) > div > section > article > p:nth-of-type(4)
[ 0.016904208809137344, 0.003410707227885723, 0.008426331914961338, 0.06263643503189087, -0.006322857458144426, -0.019853422418236732, -0.011344456113874912, -0.025062937289476395, 0.03823400288820267, 0.013087138533592224, -0.012259322218596935, 0.027834486216306686, 0.07311394810676575, 0....
p
Tree index implies that we will meet same problem that appears in relational databases:
[ "articles", "articles/filtrable-hnsw" ]
7,024
[ "Filtrable HNSW - Qdrant", "Filtrable HNSW" ]
/articles/filtrable-hnsw/
html > body > div:nth-of-type(1) > div:nth-of-type(1) > div > section > article > p:nth-of-type(4)
[ -0.0004409191315062344, -0.002707963576540351, -0.012278420850634575, 0.09062810242176056, 0.02289339154958725, 0.007393927779048681, -0.007292606867849827, -0.04735274612903595, 0.04227106273174286, -0.06974738836288452, -0.019393164664506912, 0.043105702847242355, 0.026112467050552368, -...
p
if field indexes were built independently, then it is possible to use only one of them at a time.
[ "articles", "articles/filtrable-hnsw" ]
7,025
[ "Filtrable HNSW - Qdrant", "Filtrable HNSW" ]
/articles/filtrable-hnsw/
html > body > div:nth-of-type(1) > div:nth-of-type(1) > div > section > article > p:nth-of-type(4)
[ 0.0017164601013064384, 0.0723380371928215, -0.06514989584684372, 0.0010266265599057078, -0.04819748178124428, 0.016272371634840965, -0.030189594253897667, -0.08215322345495224, -0.06415464729070663, -0.02141186222434044, -0.007343923673033714, -0.04468386992812157, 0.02462097629904747, 0.0...
p
Since nobody solved this problem before, it seems that there is no easy approach.
[ "articles", "articles/filtrable-hnsw" ]
7,026
[ "Filtrable HNSW - Qdrant", "Filtrable HNSW" ]
/articles/filtrable-hnsw/
html > body > div:nth-of-type(1) > div:nth-of-type(1) > div > section > article > p:nth-of-type(5)
[ -0.005156395491212606, -0.055662862956523895, -0.0634143278002739, -0.04248246178030968, 0.027371974661946297, -0.06837382167577744, -0.06750848144292831, 0.007576151750981808, -0.06974572688341141, -0.0036120496224611998, -0.07032656669616699, 0.039879973977804184, 0.05870778486132622, 0....
p
There is another algorithm which shows top results on the benchmark.
[ "articles", "articles/filtrable-hnsw" ]
7,027
[ "Filtrable HNSW - Qdrant", "Filtrable HNSW" ]
/articles/filtrable-hnsw/
html > body > div:nth-of-type(1) > div:nth-of-type(1) > div > section > article > p:nth-of-type(5)
[ 0.022488908842206, -0.05016845092177391, -0.034109629690647125, 0.009803409688174725, -0.004845573101192713, -0.06968412548303604, -0.0443698912858963, -0.0072596254758536816, -0.03789523243904114, 0.02947608195245266, 0.05092121288180351, 0.003503324929624796, -0.017964782193303108, 0.047...
p
It is called HNSW which stands for Hierarchical Navigable Small World.
[ "articles", "articles/filtrable-hnsw" ]
7,028
[ "Filtrable HNSW - Qdrant", "Filtrable HNSW" ]
/articles/filtrable-hnsw/
html > body > div:nth-of-type(1) > div:nth-of-type(1) > div > section > article > p:nth-of-type(6)
[ -0.18616661429405212, 0.06341461837291718, 0.048204898834228516, 0.030101805925369263, 0.03691430017352104, -0.002034332137554884, -0.08506818860769272, 0.03667370229959488, -0.03777735307812691, 0.009153028950095177, 0.04700353741645813, 0.13900712132453918, -0.025565482676029205, -0.0213...
p
The original paper is well written and very easy to read, so I will only give the main idea here.
[ "articles", "articles/filtrable-hnsw" ]
7,029
[ "Filtrable HNSW - Qdrant", "Filtrable HNSW" ]
/articles/filtrable-hnsw/
html > body > div:nth-of-type(1) > div:nth-of-type(1) > div > section > article > p:nth-of-type(6)
[ 0.027832865715026855, -0.005757729057222605, 0.04038406163454056, -0.08632301539182663, 0.02964942157268524, 0.021447518840432167, -0.022597789764404297, -0.008548754267394543, -0.02908470295369625, 0.038144636899232864, -0.05849184840917587, 0.00011469983292045072, 0.08670098334550858, 0....
p
We need to build a navigation graph among all indexed points so that the greedy search on this graph will lead us to the nearest point.
[ "articles", "articles/filtrable-hnsw" ]
7,030
[ "Filtrable HNSW - Qdrant", "Filtrable HNSW" ]
/articles/filtrable-hnsw/
html > body > div:nth-of-type(1) > div:nth-of-type(1) > div > section > article > p:nth-of-type(6)
[ -0.047501929104328156, -0.00827273353934288, -0.04112234339118004, 0.015122664161026478, -0.037727728486061096, 0.06742314994335175, 0.010475441813468933, -0.049180466681718826, 0.02197856456041336, 0.002401354257017374, 0.06615936756134033, 0.03176311030983925, 0.012703408487141132, 0.026...
p
This graph is constructed by sequentially adding points that are connected by a fixed number of edges to previously added points.
[ "articles", "articles/filtrable-hnsw" ]
7,031
[ "Filtrable HNSW - Qdrant", "Filtrable HNSW" ]
/articles/filtrable-hnsw/
html > body > div:nth-of-type(1) > div:nth-of-type(1) > div > section > article > p:nth-of-type(6)
[ 0.05434658005833626, 0.019356589764356613, -0.021739592775702477, -0.0488661527633667, 0.0849737599492073, 0.02562405727803707, 0.001537567237392068, 0.03841210529208183, 0.007861658930778503, 0.021296195685863495, 0.052588511258363724, -0.03382859751582146, 0.09062068909406662, -0.0098226...
p
In the resulting graph, the number of edges at each point does not exceed a given threshold $m$ and always contains the nearest considered points.
[ "articles", "articles/filtrable-hnsw" ]
7,032
[ "Filtrable HNSW - Qdrant" ]
/articles/filtrable-hnsw/
html > body > div:nth-of-type(1) > div:nth-of-type(1) > div > section > article > h3
[ -0.04967004805803299, 0.10167185962200165, -0.02625580132007599, -0.03325432538986206, -0.003427248913794756, 0.018231453374028206, -0.008029435761272907, -0.014361977577209473, -0.006055202800780535, 0.03208411857485771, -0.03234126418828964, 0.06615316867828369, -0.05360320955514908, -0....
h3
How can we modify it?
[ "articles", "articles/filtrable-hnsw" ]
7,033
[ "Filtrable HNSW - Qdrant", "How can we modify it?" ]
/articles/filtrable-hnsw/
html > body > div:nth-of-type(1) > div:nth-of-type(1) > div > section > article > p:nth-of-type(8)
[ -0.06520695239305496, 0.0970831885933876, 0.05631935968995094, -0.051794786006212234, 0.07158005237579346, 0.014108465984463692, -0.027371535077691078, -0.0607922300696373, -0.048524439334869385, 0.023081421852111816, -0.020466912537813187, -0.02489827200770378, 0.05802883952856064, -0.003...
p
What if we simply apply the filter criteria to the nodes of this graph and use in the greedy search only those that meet these criteria?
[ "articles", "articles/filtrable-hnsw" ]
7,034
[ "Filtrable HNSW - Qdrant", "How can we modify it?" ]
/articles/filtrable-hnsw/
html > body > div:nth-of-type(1) > div:nth-of-type(1) > div > section > article > p:nth-of-type(8)
[ -0.07866837084293365, 0.09166298806667328, -0.003391209291294217, -0.03335307911038399, 0.051422733813524246, -0.0879276692867279, 0.014796915464103222, -0.09158496558666229, -0.07561233639717102, 0.03880320116877556, 0.017551934346556664, 0.07040631026029587, 0.04223400354385376, -0.03458...
p
It turns out that even with this naive modification algorithm can cover some use cases.
[ "articles", "articles/filtrable-hnsw" ]
7,035
[ "Filtrable HNSW - Qdrant", "How can we modify it?" ]
/articles/filtrable-hnsw/
html > body > div:nth-of-type(1) > div:nth-of-type(1) > div > section > article > p:nth-of-type(9)
[ 0.01985347829759121, 0.02472173422574997, -0.011660506017506123, -0.040245383977890015, 0.06404677778482437, 0.005929110571742058, -0.011945109814405441, -0.06492995470762253, 0.038066282868385315, -0.0014885789714753628, 0.0279369093477726, -0.026355644688010216, 0.07082466781139374, 0.08...
p
One such case is if your criteria do not correlate with vector semantics.
[ "articles", "articles/filtrable-hnsw" ]
7,036
[ "Filtrable HNSW - Qdrant", "How can we modify it?" ]
/articles/filtrable-hnsw/
html > body > div:nth-of-type(1) > div:nth-of-type(1) > div > section > article > p:nth-of-type(9)
[ -0.008591226302087307, 0.0953616350889206, -0.03745152801275253, 0.03279396519064903, 0.08991096168756485, -0.02700282819569111, -0.0001344041811535135, -0.04059961065649986, -0.05939484015107155, -0.060928069055080414, 0.012653781101107597, 0.041734613478183746, 0.05516444146633148, 0.015...
p
For example, you use a vector search for clothing names and want to filter out some sizes.
[ "articles", "articles/filtrable-hnsw" ]
7,037
[ "Filtrable HNSW - Qdrant", "How can we modify it?" ]
/articles/filtrable-hnsw/
html > body > div:nth-of-type(1) > div:nth-of-type(1) > div > section > article > p:nth-of-type(9)
[ 0.027813801541924477, 0.060446128249168396, 0.017121590673923492, -0.00484272837638855, 0.07893149554729462, -0.060354556888341904, 0.003257166361436248, -0.18201622366905212, -0.03652265667915344, -0.022537758573889732, 0.06107194721698761, -0.0007882665959186852, 0.06365660578012466, -0....
p
In this case, the nodes will be uniformly filtered out from the entire cluster structure.
[ "articles", "articles/filtrable-hnsw" ]
7,038
[ "Filtrable HNSW - Qdrant", "How can we modify it?" ]
/articles/filtrable-hnsw/
html > body > div:nth-of-type(1) > div:nth-of-type(1) > div > section > article > p:nth-of-type(9)
[ -0.07201773673295975, -0.060715436935424805, 0.05013895779848099, 0.06497064977884293, 0.006854315288364887, 0.04354215785861015, 0.04232371225953102, 0.005985861178487539, 0.042576201260089874, 0.026073280721902847, -0.03722991794347763, 0.026027889922261238, 0.0006020691362209618, 0.0501...
p
Therefore, the theoretical conclusions obtained in the Percolation theory become applicable:
[ "articles", "articles/filtrable-hnsw" ]
7,039
[ "Filtrable HNSW - Qdrant", "How can we modify it?" ]
/articles/filtrable-hnsw/
html > body > div:nth-of-type(1) > div:nth-of-type(1) > div > section > article > blockquote > p
[ -0.053589824587106705, -0.05823251232504845, 0.019478468224406242, 0.04242650419473648, 0.0011238554725423455, 0.007832735776901245, 0.006185384467244148, 0.02893655188381672, 0.043331485241651535, 0.016868535429239273, -0.034812163561582565, 0.07548559457063675, 0.0704532265663147, 0.0172...
p
Percolation is related to the robustness of the graph (called also network).
[ "articles", "articles/filtrable-hnsw" ]
7,040
[ "Filtrable HNSW - Qdrant", "How can we modify it?" ]
/articles/filtrable-hnsw/
html > body > div:nth-of-type(1) > div:nth-of-type(1) > div > section > article > blockquote > p
[ 0.03417377546429634, 0.04718221351504326, 0.02704467996954918, 0.0057425289414823055, 0.003630884224548936, -0.0498410202562809, 0.04710153117775917, -0.04389113187789917, -0.037311941385269165, 0.05230983719229698, -0.015592124313116074, 0.0004325520421843976, 0.06162614747881889, -0.0731...
p
Given a random graph of $n$ nodes and an average degree $\langle k\rangle$ .
[ "articles", "articles/filtrable-hnsw" ]
7,041
[ "Filtrable HNSW - Qdrant", "How can we modify it?" ]
/articles/filtrable-hnsw/
html > body > div:nth-of-type(1) > div:nth-of-type(1) > div > section > article > blockquote > p
[ 0.018104134127497673, 0.020390238612890244, 0.06692363321781158, 0.05185464397072792, 0.09273649007081985, -0.09419357031583786, 0.06679341197013855, -0.08142220228910446, 0.005193978548049927, 0.07261122763156891, -0.03025275655090809, 0.010227417573332787, 0.06628546118736267, -0.0413141...
p
Next we remove randomly a fraction $1-p$ of nodes and leave only a fraction $p$.
[ "articles", "articles/filtrable-hnsw" ]
7,042
[ "Filtrable HNSW - Qdrant", "How can we modify it?" ]
/articles/filtrable-hnsw/
html > body > div:nth-of-type(1) > div:nth-of-type(1) > div > section > article > blockquote > p
[ 0.018296465277671814, -0.037929635494947433, 0.0507374107837677, 0.013885010033845901, 0.0441313311457634, -0.04918393865227699, 0.04002504050731659, 0.05303569510579109, -0.0009183306246995926, 0.027024637907743454, -0.06663864850997925, 0.058176442980766296, 0.004445991013199091, 0.00224...
p
There exists a critical percolation threshold $ pc = \frac{1}{\langle k\rangle} $ below which the network becomes fragmented while above $pc$ a giant connected component exists.
[ "articles", "articles/filtrable-hnsw" ]
7,043
[ "Filtrable HNSW - Qdrant", "How can we modify it?" ]
/articles/filtrable-hnsw/
html > body > div:nth-of-type(1) > div:nth-of-type(1) > div > section > article > p:nth-of-type(10)
[ -0.027103865519165993, -0.015213063918054104, 0.04913616552948952, 0.1559027135372162, 0.05182405933737755, -0.054165586829185486, 0.06475216150283813, -0.004086376167833805, -0.003024244448170066, 0.037002626806497574, 0.05028431490063667, 0.007901565171778202, 0.04524631425738335, 0.0293...
p
This statement also confirmed by experiments:
[ "articles", "articles/filtrable-hnsw" ]
7,044
[ "Filtrable HNSW - Qdrant", "How can we modify it?" ]
/articles/filtrable-hnsw/
html > body > div:nth-of-type(1) > div:nth-of-type(1) > div > section > article > figure:nth-of-type(1) > figcaption > p
[ 0.035305123776197433, -0.04281964525580406, 0.002909719478338957, 0.004442739300429821, -0.045174021273851395, 0.02760118432343006, 0.028648173436522484, -0.03696576878428459, 0.0031875898130238056, -0.009833778254687786, 0.01625887304544449, 0.041592247784137726, 0.03166469559073448, 0.03...
p
Dependency of connectivity to the number of edges
[ "articles", "articles/filtrable-hnsw" ]
7,045
[ "Filtrable HNSW - Qdrant", "How can we modify it?" ]
/articles/filtrable-hnsw/
html > body > div:nth-of-type(1) > div:nth-of-type(1) > div > section > article > figure:nth-of-type(2) > figcaption > p
[ 0.02286532334983349, -0.08817227929830551, 0.020560260862112045, 0.030585914850234985, -0.03129752725362778, 0.09054728597402573, 0.07536410540342331, -0.039445649832487106, 0.0948556512594223, -0.0541982501745224, 0.0625942125916481, 0.056879542768001556, 0.04445298761129379, 0.0303706321...
p
Dependency of connectivity to the number of point (no dependency).
[ "articles", "articles/filtrable-hnsw" ]
7,046
[ "Filtrable HNSW - Qdrant", "How can we modify it?" ]
/articles/filtrable-hnsw/
html > body > div:nth-of-type(1) > div:nth-of-type(1) > div > section > article > p:nth-of-type(11)
[ 0.020527055487036705, -0.06596172600984573, -0.011419967748224735, 0.04531334340572357, 0.05064483359456062, 0.005321791861206293, 0.0026849035639315844, -0.003812452545389533, 0.04830535501241684, -0.02706175670027733, 0.03866271674633026, 0.022808095440268517, 0.04872315376996994, -0.022...
p
There is a clear threshold when the search begins to fail.
[ "articles", "articles/filtrable-hnsw" ]
7,047
[ "Filtrable HNSW - Qdrant", "How can we modify it?" ]
/articles/filtrable-hnsw/
html > body > div:nth-of-type(1) > div:nth-of-type(1) > div > section > article > p:nth-of-type(11)
[ 0.07046439498662949, -0.030070310458540916, 0.037855636328458786, 0.014534304849803448, 0.04858735203742981, -0.0758068859577179, -0.02737274207174778, 0.026522811502218246, 0.051131267100572586, 0.062242936342954636, 0.0460616834461689, -0.010321379639208317, 0.014815696515142918, 0.03759...
p
This threshold is due to the decomposition of the graph into small connected components.
[ "articles", "articles/filtrable-hnsw" ]
7,048
[ "Filtrable HNSW - Qdrant", "How can we modify it?" ]
/articles/filtrable-hnsw/
html > body > div:nth-of-type(1) > div:nth-of-type(1) > div > section > article > p:nth-of-type(11)
[ 0.01685345731675625, -0.007973006926476955, -0.014025656506419182, 0.010182834230363369, 0.04983082041144371, -0.06535246968269348, -0.03625040873885155, -0.047200798988342285, -0.056012459099292755, 0.02334235981106758, -0.008230303414165974, -0.0046020615845918655, 0.07280397415161133, -...
p
The graphs also show that this threshold can be shifted by increasing the $m$ parameter of the algorithm, which is responsible for the degree of nodes.
[ "articles", "articles/filtrable-hnsw" ]
7,049
[ "Filtrable HNSW - Qdrant", "How can we modify it?" ]
/articles/filtrable-hnsw/
html > body > div:nth-of-type(1) > div:nth-of-type(1) > div > section > article > p:nth-of-type(12)
[ -0.07250309735536575, 0.001965514849871397, 0.004299175459891558, -0.011643731035292149, 0.0760824903845787, 0.04530646279454231, -0.009130417369306087, -0.09503579884767532, 0.03022714890539646, -0.039283763617277145, 0.02166396751999855, -0.013130243867635727, 0.04620172828435898, 0.0285...
p
Let’s consider some other filtering conditions we might want to apply in the search:
[ "articles", "articles/filtrable-hnsw" ]
7,050
[ "Filtrable HNSW - Qdrant", "How can we modify it?" ]
/articles/filtrable-hnsw/
html > body > div:nth-of-type(1) > div:nth-of-type(1) > div > section > article > ul:nth-of-type(2) > li:nth-of-type(1)
[ 0.04228685423731804, 0.0363638810813427, 0.03365592285990715, -0.009689723141491413, 0.04031877964735031, 0.05927532911300659, 0.13150554895401, -0.11285239458084106, 0.03139110654592514, -0.09210763871669769, -0.04706212505698204, -0.10548529028892517, -0.017390793189406395, -0.0242806654...
li
Categorical filteringSelect only points in a specific categorySelect points which belong to a specific subset of categoriesSelect points with a specific set of labels
[ "articles", "articles/filtrable-hnsw" ]
7,051
[ "Filtrable HNSW - Qdrant", "How can we modify it?" ]
/articles/filtrable-hnsw/
html > body > div:nth-of-type(1) > div:nth-of-type(1) > div > section > article > ul:nth-of-type(2) > li:nth-of-type(1) > ul > li:nth-of-type(1)
[ 0.07948527485132217, 0.078904889523983, 0.023638978600502014, 0.05251903086900711, 0.035301558673381805, 0.04459598660469055, 0.10987468808889389, -0.09340855479240417, 0.08668725192546844, -0.06278034299612045, -0.02914738655090332, -0.09118054807186127, -0.05655914545059204, -0.025223495...
li
Select only points in a specific category
[ "articles", "articles/filtrable-hnsw" ]
7,052
[ "Filtrable HNSW - Qdrant", "How can we modify it?" ]
/articles/filtrable-hnsw/
html > body > div:nth-of-type(1) > div:nth-of-type(1) > div > section > article > ul:nth-of-type(2) > li:nth-of-type(1) > ul > li:nth-of-type(2)
[ 0.08973832428455353, 0.051340874284505844, 0.06233077496290207, 0.03245372325181961, 0.01064669992774725, 0.0147702069953084, 0.14743417501449585, -0.07683724910020828, 0.06016991659998894, -0.045450951904058456, -0.035049036145210266, -0.07996651530265808, -0.027011550962924957, -0.010645...
li
Select points which belong to a specific subset of categories
[ "articles", "articles/filtrable-hnsw" ]
7,053
[ "Filtrable HNSW - Qdrant", "How can we modify it?" ]
/articles/filtrable-hnsw/
html > body > div:nth-of-type(1) > div:nth-of-type(1) > div > section > article > ul:nth-of-type(2) > li:nth-of-type(1) > ul > li:nth-of-type(3)
[ 0.04414979740977287, 0.037190031260252, -0.022963279858231544, -0.004084533080458641, 0.02932164818048477, 0.12076899409294128, 0.21246562898159027, -0.05401492491364479, 0.04812745749950409, -0.09916901588439941, 0.03277970850467682, -0.07595442235469818, 0.04488322511315346, -0.005591466...
li
Select points with a specific set of labels
[ "articles", "articles/filtrable-hnsw" ]
7,054
[ "Filtrable HNSW - Qdrant", "How can we modify it?" ]
/articles/filtrable-hnsw/
html > body > div:nth-of-type(1) > div:nth-of-type(1) > div > section > article > ul:nth-of-type(2) > li:nth-of-type(2)
[ 0.027900369837880135, 0.05024560168385506, -0.028082456439733505, -0.05009869486093521, -0.09782102704048157, 0.025706637650728226, -0.04600507393479347, 0.0576239712536335, -0.08125247806310654, -0.08111903816461563, -0.014051109552383423, -0.056877367198467255, 0.06887424737215042, -0.03...
li
Numerical range
[ "articles", "articles/filtrable-hnsw" ]
7,055
[ "Filtrable HNSW - Qdrant", "How can we modify it?" ]
/articles/filtrable-hnsw/
html > body > div:nth-of-type(1) > div:nth-of-type(1) > div > section > article > ul:nth-of-type(2) > li:nth-of-type(3)
[ 0.13504450023174286, 0.01842666044831276, 0.0870441198348999, -0.004619195591658354, 0.06958991289138794, 0.010389708913862705, 0.022219698876142502, -0.0696859285235405, -0.03593172878026962, 0.029557297006249428, 0.018866293132305145, -0.07904442399740219, -0.02804902382194996, 0.0304003...
li
Selection within some geographical region
[ "articles", "articles/filtrable-hnsw" ]
7,056
[ "Filtrable HNSW - Qdrant", "How can we modify it?" ]
/articles/filtrable-hnsw/
html > body > div:nth-of-type(1) > div:nth-of-type(1) > div > section > article > p:nth-of-type(13)
[ -0.06265492737293243, 0.0010969561990350485, -0.017846155911684036, 0.018878726288676262, 0.018762407824397087, 0.03670673444867134, -0.0642135813832283, -0.012524082325398922, -0.015269826166331768, -0.025339897722005844, 0.04731335863471031, 0.0321098230779171, -0.022623978555202484, -0....
p
In the first case, we can guarantee that the HNSW graph will be connected simply by creating additional edges
[ "articles", "articles/filtrable-hnsw" ]
7,057
[ "Filtrable HNSW - Qdrant", "How can we modify it?" ]
/articles/filtrable-hnsw/
html > body > div:nth-of-type(1) > div:nth-of-type(1) > div > section > article > p:nth-of-type(13)
[ 0.03446009382605553, 0.03360319137573242, -0.042238809168338776, -0.0047805397771298885, -0.03555438667535782, -0.08561887592077255, -0.10296503454446793, -0.08348585665225983, -0.06639818102121353, 0.004064504057168961, -0.01353126484900713, -0.0593351274728775, 0.014161111786961555, 0.03...
p
inside each category separately, using the same graph construction algorithm, and then combining them into the original graph.
[ "articles", "articles/filtrable-hnsw" ]
7,058
[ "Filtrable HNSW - Qdrant", "How can we modify it?" ]
/articles/filtrable-hnsw/
html > body > div:nth-of-type(1) > div:nth-of-type(1) > div > section > article > p:nth-of-type(13)
[ 0.06386639177799225, -0.049862924963235855, -0.018748221918940544, -0.010974561795592308, -0.056695640087127686, -0.03182503208518028, -0.11641620099544525, -0.026269380003213882, 0.0281209759414196, 0.0054510789923369884, 0.00757246371358633, -0.015891101211309433, -0.04066770523786545, -...
p
In this case, the total number of edges will increase by no more than 2 times, regardless of the number of categories.
[ "articles", "articles/filtrable-hnsw" ]
7,059
[ "Filtrable HNSW - Qdrant", "How can we modify it?" ]
/articles/filtrable-hnsw/
html > body > div:nth-of-type(1) > div:nth-of-type(1) > div > section > article > p:nth-of-type(14)
[ 0.03320576623082161, -0.0007627686718478799, 0.011732777580618858, -0.018935291096568108, 0.015148773789405823, -0.04372057318687439, -0.00236936891451478, 0.06184263154864311, 0.028101909905672073, -0.02020113170146942, 0.06079189479351044, 0.04772733524441719, 0.02582893706858158, 0.0329...
p
Second case is a little harder.
[ "articles", "articles/filtrable-hnsw" ]
7,060
[ "Filtrable HNSW - Qdrant", "How can we modify it?" ]
/articles/filtrable-hnsw/
html > body > div:nth-of-type(1) > div:nth-of-type(1) > div > section > article > p:nth-of-type(14)
[ 0.05824988707900047, -0.07109715789556503, -0.0590544156730175, 0.08610904961824417, -0.018879368901252747, -0.004652967676520348, -0.052254755049943924, -0.10884525626897812, 0.08006105571985245, -0.06771412491798401, 0.06252550333738327, -0.04242772236466408, -0.013721851631999016, -0.02...
p
A connection may be lost between two categories if they lie in different clusters.
[ "articles", "articles/filtrable-hnsw" ]
7,061
[ "Filtrable HNSW - Qdrant", "How can we modify it?" ]
/articles/filtrable-hnsw/
html > body > div:nth-of-type(1) > div:nth-of-type(1) > div > section > article > p:nth-of-type(16)
[ 0.028522901237010956, -0.01567145250737667, -0.04256248101592064, -0.02341410331428051, -0.014637894928455353, -0.06025610864162445, -0.1158967986702919, -0.0849970281124115, -0.03781524673104286, -0.009256727993488312, -0.018587268888950348, -0.012952124699950218, 0.05831953510642052, 0.0...
p
The idea here is to build same navigation graph but not between nodes, but between categories.
[ "articles", "articles/filtrable-hnsw" ]
7,062
[ "Filtrable HNSW - Qdrant", "How can we modify it?" ]
/articles/filtrable-hnsw/
html > body > div:nth-of-type(1) > div:nth-of-type(1) > div > section > article > p:nth-of-type(16)
[ 0.05876164138317108, -0.027225397527217865, -0.09154102951288223, -0.04488993436098099, -0.03641296923160553, 0.00891772098839283, -0.09416400641202927, 0.06671303510665894, 0.052602995187044144, -0.08668322116136551, 0.048389170318841934, -0.05194777250289917, 0.00013808715448249131, -0.0...
p
Distance between two categories might be defined as distance between category entry points (or, for precision, as the average distance between a random sample).
[ "articles", "articles/filtrable-hnsw" ]
7,063
[ "Filtrable HNSW - Qdrant", "How can we modify it?" ]
/articles/filtrable-hnsw/
html > body > div:nth-of-type(1) > div:nth-of-type(1) > div > section > article > p:nth-of-type(16)
[ 0.0599869079887867, -0.033331386744976044, -0.0016429077368229628, 0.07620910555124283, -0.004056375939399004, -0.014001643285155296, -0.037597160786390305, -0.05533100664615631, 0.02401539497077465, 0.0456874705851078, 0.03980293869972229, 0.0006420021527446806, 0.0376126803457737, -0.019...
p
Now we can estimate expected graph connectivity by number of excluded categories, not nodes.
[ "articles", "articles/filtrable-hnsw" ]
7,064
[ "Filtrable HNSW - Qdrant", "How can we modify it?" ]
/articles/filtrable-hnsw/
html > body > div:nth-of-type(1) > div:nth-of-type(1) > div > section > article > p:nth-of-type(16)
[ 0.010628901422023773, -0.08898404985666275, -0.02802846021950245, 0.05106719583272934, -0.0050024050287902355, 0.010929885320365429, -0.025684164837002754, -0.1312396377325058, 0.05320221558213234, -0.06044898182153702, -0.004310553893446922, 0.021109633147716522, -0.001697966014035046, 0....
p
It still does not guarantee that two random categories will be connected, but allows us to switch to multiple searches in each category if connectivity threshold passed.
[ "articles", "articles/filtrable-hnsw" ]
7,065
[ "Filtrable HNSW - Qdrant", "How can we modify it?" ]
/articles/filtrable-hnsw/
html > body > div:nth-of-type(1) > div:nth-of-type(1) > div > section > article > p:nth-of-type(16)
[ -0.015710318461060524, -0.022172333672642708, -0.02510368451476097, -0.007304381113499403, 0.010548872873187065, -0.0358663871884346, -0.0847087949514389, -0.023124145343899727, 0.06909137219190598, -0.0359969325363636, 0.04728368669748306, 0.0471845343708992, 0.023928824812173843, -0.0236...
p
In some cases, multiple searches can be even faster if you take advantage of parallel processing.
[ "articles", "articles/filtrable-hnsw" ]
7,066
[ "Filtrable HNSW - Qdrant", "How can we modify it?" ]
/articles/filtrable-hnsw/
html > body > div:nth-of-type(1) > div:nth-of-type(1) > div > section > article > figure:nth-of-type(3) > figcaption > p
[ 0.014577987603843212, -0.050503458827733994, 0.025831693783402443, 0.06750885397195816, -0.0064574978314340115, 0.03323967382311821, 0.01619436778128147, -0.11194662749767303, 0.05922003835439682, -0.05238800123333931, 0.009409812279045582, 0.025379031896591187, 0.014117008075118065, -0.00...
p
Dependency of connectivity to the random categories included in search
[ "articles", "articles/filtrable-hnsw" ]
7,067
[ "Filtrable HNSW - Qdrant", "How can we modify it?" ]
/articles/filtrable-hnsw/
html > body > div:nth-of-type(1) > div:nth-of-type(1) > div > section > article > p:nth-of-type(17)
[ -0.048394348472356796, -0.056967269629240036, -0.03335699066519737, 0.012747739441692829, -0.060267675668001175, -0.05671557039022446, -0.021848857402801514, -0.0220168549567461, 0.025907248258590698, -0.004620696883648634, 0.099248506128788, 0.036616384983062744, 0.04856168478727341, 0.00...
p
Third case might be resolved in a same way it is resolved in classical databases.
[ "articles", "articles/filtrable-hnsw" ]
7,068
[ "Filtrable HNSW - Qdrant", "How can we modify it?" ]
/articles/filtrable-hnsw/
html > body > div:nth-of-type(1) > div:nth-of-type(1) > div > section > article > p:nth-of-type(17)
[ 0.039297476410865784, 0.02876005321741104, -0.10000845044851303, -0.013213296420872211, 0.06572721898555756, 0.03371096029877663, 0.045549262315034866, -0.033828750252723694, -0.05509177967905998, -0.06900277733802795, 0.01875522918999195, -0.06455982476472855, 0.020029988139867783, 0.0439...
p
Depending on labeled subsets size ration we can go for one of the following scenarios:
[ "articles", "articles/filtrable-hnsw" ]
7,069
[ "Filtrable HNSW - Qdrant", "How can we modify it?" ]
/articles/filtrable-hnsw/
html > body > div:nth-of-type(1) > div:nth-of-type(1) > div > section > article > ul:nth-of-type(3) > li:nth-of-type(1)
[ 0.07950256764888763, 0.09460633248090744, -0.00004541312227956951, -0.05332015082240105, 0.1051400750875473, 0.08991347998380661, 0.10431716591119766, -0.06240161135792732, -0.002974729984998703, -0.07326018810272217, -0.01659967191517353, -0.08399459719657898, 0.0411515086889267, -0.03100...
li
if at least one subset is small: perform search over the label containing smallest subset and then filter points consequently.
[ "articles", "articles/filtrable-hnsw" ]
7,070
[ "Filtrable HNSW - Qdrant", "How can we modify it?" ]
/articles/filtrable-hnsw/
html > body > div:nth-of-type(1) > div:nth-of-type(1) > div > section > article > ul:nth-of-type(3) > li:nth-of-type(2)
[ 0.09921976178884506, 0.011773063801229, 0.048135802149772644, -0.022676343098282814, 0.03749625384807587, -0.05491669476032257, 0.04538063704967499, -0.10416756570339203, -0.05743780732154846, -0.03456404432654381, -0.020633621141314507, -0.026490019634366035, 0.0010814942652359605, 0.0004...
li
if large subsets give large intersection: perform regular search with constraints expecting that intersection size fits connectivity threshold.
[ "articles", "articles/filtrable-hnsw" ]
7,071
[ "Filtrable HNSW - Qdrant", "How can we modify it?" ]
/articles/filtrable-hnsw/
html > body > div:nth-of-type(1) > div:nth-of-type(1) > div > section > article > ul:nth-of-type(3) > li:nth-of-type(3)
[ 0.09951461106538773, 0.0007448479882441461, 0.057393018156290054, -0.03146151080727577, 0.10355629771947861, -0.0380999930202961, 0.04477627947926521, -0.09339945018291473, -0.060991320759058, -0.041941169649362564, -0.002849257318302989, -0.027038443833589554, -0.021726572886109352, -0.00...
li
if large subsets give small intersection: perform linear search over intersection expecting that it is small enough to fit a time frame.
[ "articles", "articles/filtrable-hnsw" ]
7,072
[ "Filtrable HNSW - Qdrant", "How can we modify it?" ]
/articles/filtrable-hnsw/
html > body > div:nth-of-type(1) > div:nth-of-type(1) > div > section > article > p:nth-of-type(18)
[ -0.0033875275403261185, 0.06374076753854752, 0.08497770130634308, -0.06308944523334503, 0.0311481524258852, -0.008460708893835545, -0.066097192466259, 0.0170676838606596, 0.018457772210240364, -0.04299675673246384, -0.018819158896803856, -0.010183572769165039, 0.07837992906570435, -0.00846...
p
Numerical range case can be reduces to the previous one if we split numerical range into a buckets containing equal amount of points.
[ "articles", "articles/filtrable-hnsw" ]
7,073
[ "Filtrable HNSW - Qdrant", "How can we modify it?" ]
/articles/filtrable-hnsw/
html > body > div:nth-of-type(1) > div:nth-of-type(1) > div > section > article > p:nth-of-type(18)
[ 0.01451070886105299, -0.05655446648597717, -0.0028648071456700563, -0.017375657334923744, -0.031026624143123627, -0.03992459177970886, -0.07043959945440292, -0.008783126249909401, 0.043697696179151535, 0.04996773600578308, -0.014952012337744236, 0.023561948910355568, 0.030525989830493927, ...
p
Next we also connect neighboring buckets to achieve graph connectivity.
[ "articles", "articles/filtrable-hnsw" ]
7,074
[ "Filtrable HNSW - Qdrant", "How can we modify it?" ]
/articles/filtrable-hnsw/
html > body > div:nth-of-type(1) > div:nth-of-type(1) > div > section > article > p:nth-of-type(18)
[ 0.029710892587900162, 0.08252530544996262, 0.0111321322619915, -0.06653445959091187, 0.06850922107696533, -0.030824467539787292, -0.022894855588674545, -0.046600595116615295, 0.0002548817137721926, -0.009907548315823078, -0.09053995460271835, -0.04394509270787239, 0.007709538098424673, 0.0...
p
We still need to filter some results which presence in border buckets but do not fulfill actual constraints, but their amount might be regulated by the size of buckets.
[ "articles", "articles/filtrable-hnsw" ]
7,075
[ "Filtrable HNSW - Qdrant", "How can we modify it?" ]
/articles/filtrable-hnsw/
html > body > div:nth-of-type(1) > div:nth-of-type(1) > div > section > article > p:nth-of-type(19)
[ 0.08319376409053802, 0.009268890134990215, 0.03533262014389038, -0.06543142348527908, -0.08512721210718155, -0.10765205323696136, -0.07362909615039825, 0.007994337938725948, 0.01721223257482052, -0.02238335646688938, 0.041660748422145844, -0.01883356086909771, 0.02802438661456108, 0.067400...
p
Geographical case is a lot like a numerical one.
[ "articles", "articles/filtrable-hnsw" ]
7,076
[ "Filtrable HNSW - Qdrant", "How can we modify it?" ]
/articles/filtrable-hnsw/
html > body > div:nth-of-type(1) > div:nth-of-type(1) > div > section > article > p:nth-of-type(19)
[ 0.05286095663905144, 0.04217933490872383, 0.06431566923856735, -0.0739242285490036, -0.020858699455857277, -0.04021402820944786, -0.035005830228328705, -0.041398707777261734, -0.013237720355391502, -0.042161908000707626, 0.01715751737356186, 0.0013828481314703822, 0.027701949700713158, -0....
p
Usual geographical search involves geohash, which matches any geo-point to a fixes length identifier.
[ "articles", "articles/filtrable-hnsw" ]
7,077
[ "Filtrable HNSW - Qdrant", "How can we modify it?" ]
/articles/filtrable-hnsw/
html > body > div:nth-of-type(1) > div:nth-of-type(1) > div > section > article > p:nth-of-type(21)
[ -0.003084398340433836, 0.02231737971305847, 0.0012497494462877512, -0.040768906474113464, -0.06559201329946518, -0.017397519201040268, 0.015601756051182747, -0.09629154950380325, 0.026534901931881905, -0.029616106301546097, 0.023652110248804092, -0.023101693019270897, 0.02676357887685299, ...
p
We can use this identifiers as categories and additionally make connections between neighboring geohashes.
[ "articles", "articles/filtrable-hnsw" ]
7,078
[ "Filtrable HNSW - Qdrant", "How can we modify it?" ]
/articles/filtrable-hnsw/
html > body > div:nth-of-type(1) > div:nth-of-type(1) > div > section > article > p:nth-of-type(21)
[ -0.0018434131052345037, 0.017873303964734077, 0.004770074505358934, -0.010384521447122097, 0.045852869749069214, -0.0069743311032652855, -0.06020229309797287, -0.07632860541343689, -0.004048133734613657, -0.03279862552881241, 0.040116261690855026, 0.049417540431022644, -0.039161764085292816,...
p
It will ensure that any selected geographical region will also contain connected HNSW graph.
[ "articles", "articles/filtrable-hnsw" ]
7,079
[ "Filtrable HNSW - Qdrant" ]
/articles/filtrable-hnsw/
html > body > div:nth-of-type(1) > div:nth-of-type(1) > div > section > article > h2
[ 0.0345105342566967, 0.13655169308185577, 0.0018493582028895617, 0.05189747363328934, -0.009813056327402592, 0.02261107973754406, 0.09343250840902328, 0.06036166846752167, 0.017523951828479767, 0.03887659311294556, 0.010159940458834171, 0.01514778658747673, 0.008142069913446903, 0.072624303...
h2
Conclusion
[ "articles", "articles/filtrable-hnsw" ]
7,080
[ "Filtrable HNSW - Qdrant", "Conclusion" ]
/articles/filtrable-hnsw/
html > body > div:nth-of-type(1) > div:nth-of-type(1) > div > section > article > p:nth-of-type(22)
[ -0.10003414005041122, 0.03256025165319443, 0.011857720091938972, 0.002083408646285534, 0.03098490461707115, -0.010773024521768093, 0.004222946707159281, -0.08695582300424576, -0.06045082211494446, -0.10185858607292175, -0.03425636887550354, 0.029466278851032257, 0.029921546578407288, -0.02...
p
It is possible to enchant HNSW algorithm so that it will support filtering points in a first search phase.
[ "articles", "articles/filtrable-hnsw" ]
7,081
[ "Filtrable HNSW - Qdrant", "Conclusion" ]
/articles/filtrable-hnsw/
html > body > div:nth-of-type(1) > div:nth-of-type(1) > div > section > article > p:nth-of-type(22)
[ 0.024640094488859177, 0.02957439236342907, -0.01631920225918293, -0.004570194985717535, 0.04779129475355148, 0.02435091696679592, 0.0618758499622345, -0.10660162568092346, -0.03377082943916321, -0.09352745115756989, -0.020919663831591606, -0.08866863697767258, 0.015138888731598854, 0.02465...
p
Filtering can be carried out on the basis of belonging to categories,
[ "articles", "articles/filtrable-hnsw" ]
7,082
[ "Filtrable HNSW - Qdrant", "Conclusion" ]
/articles/filtrable-hnsw/
html > body > div:nth-of-type(1) > div:nth-of-type(1) > div > section > article > p:nth-of-type(22)
[ 0.031889524310827255, 0.003307081526145339, 0.0072892108000814915, -0.03491171821951866, -0.09258370101451874, -0.03329218178987503, -0.08524132519960403, 0.017307519912719727, 0.010421845130622387, -0.04483785852789879, 0.03222711756825447, 0.022623298689723015, 0.03267061710357666, 0.063...
p
which in turn is generalized to such popular cases as numerical ranges and geo.
[ "articles", "articles/filtrable-hnsw" ]
7,083
[ "Filtrable HNSW - Qdrant", "Conclusion" ]
/articles/filtrable-hnsw/
html > body > div:nth-of-type(1) > div:nth-of-type(1) > div > section > article > p:nth-of-type(23)
[ -0.049439288675785065, 0.07258656620979309, -0.024823306128382683, -0.024335801601409912, -0.03152364864945412, -0.128509059548378, -0.04331214725971222, -0.030067645013332367, -0.1080763041973114, 0.00448864558711648, -0.00334400637075305, 0.07886502146720886, 0.08892512321472168, -0.0243...
p
Experiments were carried by modification python implementation of the algorithm,
[ "articles", "articles/filtrable-hnsw" ]
7,084
[ "Filtrable HNSW - Qdrant", "Conclusion" ]
/articles/filtrable-hnsw/
html > body > div:nth-of-type(1) > div:nth-of-type(1) > div > section > article > p:nth-of-type(23)
[ -0.023408692330121994, -0.03688640147447586, -0.073342464864254, -0.06706681847572327, 0.04999301955103874, -0.10465151816606522, -0.18890340626239777, 0.010882986709475517, 0.033434342592954636, -0.042818378657102585, 0.0378677137196064, 0.03520132973790169, 0.004654520191252232, -0.03813...
p
but real production systems require much faster version, like NMSLib.
[ "articles", "articles/filtrable-hnsw" ]
7,085
[ "AWS Marketplace - Qdrant" ]
/documentation/cloud/aws-marketplace/
html > body > div:nth-of-type(1) > section > div > div > div:nth-of-type(2) > article > h1
[ -0.06789988279342651, 0.014377804473042488, -0.0475853830575943, -0.013357804156839848, -0.025738654658198357, 0.05410824716091156, -0.055006083101034164, -0.008974313735961914, 0.04451215639710426, 0.0726894736289978, 0.03579834848642349, -0.03936470299959183, 0.055983301252126694, -0.000...
h1
Qdrant Cloud on AWS Marketplace
[ "documentation", "documentation/cloud", "documentation/cloud/aws-marketplace" ]
7,086
[ "AWS Marketplace - Qdrant" ]
/documentation/cloud/aws-marketplace/
html > body > div:nth-of-type(1) > section > div > div > div:nth-of-type(2) > article > h2:nth-of-type(1)
[ -0.04313991963863373, 0.04466705024242401, -0.001381761976517737, 0.03237174078822136, -0.0007944217650219798, 0.03537530452013016, 0.1064491719007492, 0.08225446194410324, -0.0832243412733078, 0.0094078304246068, -0.04622291028499603, 0.021912911906838417, 0.018983757123351097, -0.0449770...
h2
Overview
[ "documentation", "documentation/cloud", "documentation/cloud/aws-marketplace" ]
7,087
[ "AWS Marketplace - Qdrant", "Overview" ]
/documentation/cloud/aws-marketplace/
html > body > div:nth-of-type(1) > section > div > div > div:nth-of-type(2) > article > p:nth-of-type(1)
[ -0.07241352647542953, 0.007111965212970972, -0.08501242846250534, -0.02500370889902115, 0.01181589625775814, 0.02156519517302513, -0.0006346515147015452, 0.024238120764493942, 0.03072076477110386, 0.024319197982549667, -0.0032484885305166245, 0.02061302587389946, 0.032037150114774704, 0.02...
p
Our AWS Marketplace listing streamlines access to Qdrant for users who rely on Amazon Web Services for hosting and application development.
[ "documentation", "documentation/cloud", "documentation/cloud/aws-marketplace" ]
7,088
[ "AWS Marketplace - Qdrant", "Overview" ]
/documentation/cloud/aws-marketplace/
html > body > div:nth-of-type(1) > section > div > div > div:nth-of-type(2) > article > p:nth-of-type(1)
[ -0.06487268954515457, -0.04491674154996872, -0.02368333749473095, 0.0028387350030243397, 0.01504296064376831, 0.019490689039230347, -0.11653430759906769, -0.0005970335332676768, 0.03974038362503052, 0.048857588320970535, -0.01080469973385334, -0.02632634900510311, 0.04482880234718323, 0.03...
p
Please note that, while Qdrant’s clusters run on AWS, you will still use the Qdrant Cloud infrastructure.
[ "documentation", "documentation/cloud", "documentation/cloud/aws-marketplace" ]
7,089
[ "AWS Marketplace - Qdrant" ]
/documentation/cloud/aws-marketplace/
html > body > div:nth-of-type(1) > section > div > div > div:nth-of-type(2) > article > h2:nth-of-type(2)
[ -0.08086583763360977, 0.03642721474170685, -0.017634602263569832, -0.0480148121714592, -0.048209913074970245, -0.007742108311504126, 0.07038458436727524, 0.0033433239441365004, 0.06500361114740372, 0.004625571426004171, -0.05813894048333168, -0.020015662536025047, -0.036263249814510345, -0...
h2
Billing
[ "documentation", "documentation/cloud", "documentation/cloud/aws-marketplace" ]
7,090
[ "AWS Marketplace - Qdrant", "Billing" ]
/documentation/cloud/aws-marketplace/
html > body > div:nth-of-type(1) > section > div > div > div:nth-of-type(2) > article > p:nth-of-type(2)
[ -0.047809988260269165, -0.02217816561460495, 0.03983146697282791, -0.00025837341672740877, -0.018390515819191933, 0.06725877523422241, 0.007722153328359127, -0.013148508034646511, 0.10483010858297348, -0.03152777999639511, 0.005983221344649792, -0.037532705813646317, -0.008411132730543613, ...
p
You don’t need to use a credit card to sign up for Qdrant Cloud.
[ "documentation", "documentation/cloud", "documentation/cloud/aws-marketplace" ]
7,091
[ "AWS Marketplace - Qdrant", "Billing" ]
/documentation/cloud/aws-marketplace/
html > body > div:nth-of-type(1) > section > div > div > div:nth-of-type(2) > article > p:nth-of-type(2)
[ -0.08217734098434448, 0.017563167959451675, -0.05574150010943413, -0.024544619023799896, -0.017977284267544746, 0.028510477393865585, 0.0015596020966768265, 0.0001383090711897239, 0.16395770013332367, 0.05510488897562027, 0.007484142202883959, -0.02150365523993969, 0.03693167492747307, 0.0...
p
Instead, all billing is processed through the AWS Marketplace and the usage of Qdrant is added to your existing billing for AWS services.
[ "documentation", "documentation/cloud", "documentation/cloud/aws-marketplace" ]
7,092
[ "AWS Marketplace - Qdrant", "Billing" ]
/documentation/cloud/aws-marketplace/
html > body > div:nth-of-type(1) > section > div > div > div:nth-of-type(2) > article > p:nth-of-type(2)
[ -0.02314305491745472, 0.032660070806741714, -0.047233473509550095, -0.017420271411538124, 0.018729008734226227, -0.009025566279888153, -0.01549796387553215, 0.023573296144604683, 0.15073774755001068, 0.09545135498046875, -0.03552926704287529, -0.01373355183750391, 0.031216608360409737, 0.0...
p
It is common for AWS to abstract usage based pricing in the AWS marketplace, as there are too many factors to model when calculating billing from the AWS side.
[ "documentation", "documentation/cloud", "documentation/cloud/aws-marketplace" ]
7,093
[ "AWS Marketplace - Qdrant", "Billing" ]
/documentation/cloud/aws-marketplace/
html > body > div:nth-of-type(1) > section > div > div > div:nth-of-type(2) > article > p:nth-of-type(4)
[ 0.022806992754340172, 0.04369789734482765, -0.022188875824213028, -0.013505839742720127, 0.06094726547598839, 0.0187898688018322, 0.024898486211895943, -0.059993501752614975, 0.08313388377428055, 0.08076240122318268, 0.03876318782567978, -0.05117752403020859, 0.059456538408994675, -0.07435...
p
The payment is carried out via your AWS Account.
[ "documentation", "documentation/cloud", "documentation/cloud/aws-marketplace" ]
7,094
[ "AWS Marketplace - Qdrant", "Billing" ]
/documentation/cloud/aws-marketplace/
html > body > div:nth-of-type(1) > section > div > div > div:nth-of-type(2) > article > p:nth-of-type(4)
[ -0.009099346585571766, 0.004505238961428404, -0.006223360076546669, -0.039187055081129074, -0.037746645510196686, -0.008768337778747082, -0.05506357178092003, 0.048252519220113754, 0.10008213669061661, 0.05231517180800438, -0.042714498937129974, -0.03526292368769646, 0.005799760110676289, ...
p
To get a clearer idea for the pricing structure, please use our Billing Calculator.
[ "documentation", "documentation/cloud", "documentation/cloud/aws-marketplace" ]
7,095
[ "AWS Marketplace - Qdrant" ]
/documentation/cloud/aws-marketplace/
html > body > div:nth-of-type(1) > section > div > div > div:nth-of-type(2) > article > h2:nth-of-type(3)
[ -0.04103842377662659, -0.09806843101978302, -0.05204593390226364, -0.009730827994644642, 0.0008853606414049864, -0.02908790484070778, 0.0005397621425800025, 0.044279564172029495, 0.005584837403148413, 0.0028678940143436193, -0.07755840569734573, -0.01356692798435688, 0.024302341043949127, ...
h2
How to subscribe
[ "documentation", "documentation/cloud", "documentation/cloud/aws-marketplace" ]
7,096
[ "AWS Marketplace - Qdrant", "How to subscribe" ]
/documentation/cloud/aws-marketplace/
html > body > div:nth-of-type(1) > section > div > div > div:nth-of-type(2) > article > ol:nth-of-type(1) > li:nth-of-type(1)
[ -0.0003279636730439961, 0.0029478976503014565, -0.06027214601635933, 0.008880825713276863, 0.03347885236144066, 0.035090669989585876, -0.03894713148474693, 0.015410864725708961, 0.013734799809753895, 0.06763027608394623, 0.03683972358703613, -0.032704561948776245, 0.011101022362709045, -0....
li
Go to Qdrant’s AWS Marketplace listing.
[ "documentation", "documentation/cloud", "documentation/cloud/aws-marketplace" ]
7,097
[ "AWS Marketplace - Qdrant", "How to subscribe" ]
/documentation/cloud/aws-marketplace/
html > body > div:nth-of-type(1) > section > div > div > div:nth-of-type(2) > article > ol:nth-of-type(1) > li:nth-of-type(2)
[ 0.00918828509747982, -0.0043913619592785835, -0.009119597263634205, 0.029901787638664246, 0.08344636112451553, 0.06378857046365738, 0.0459897480905056, -0.01902049407362938, 0.04601418599486351, -0.02986358106136322, 0.04456588998436928, 0.05245513841509819, -0.1008821353316307, -0.0192806...
li
Click the bright orange button - View purchase options.
[ "documentation", "documentation/cloud", "documentation/cloud/aws-marketplace" ]
7,098
[ "AWS Marketplace - Qdrant", "How to subscribe" ]
/documentation/cloud/aws-marketplace/
html > body > div:nth-of-type(1) > section > div > div > div:nth-of-type(2) > article > ol:nth-of-type(1) > li:nth-of-type(3)
[ -0.04491692781448364, -0.08059558272361755, -0.04769459739327431, -0.033585745841264725, 0.047513242810964584, 0.03388064727187157, -0.004682694096118212, 0.0067694843746721745, 0.012006470933556557, 0.004009230062365532, 0.01836288347840309, 0.03785767778754234, -0.01608443446457386, 0.01...
li
On the next screen, under Purchase, click Subscribe.
[ "documentation", "documentation/cloud", "documentation/cloud/aws-marketplace" ]
7,099