5. Interacting Directly with the Model#
So far, all of our interactions with the LLM have been executed via the generative prompting mechanism. But a rich textual model lies behind the generative side of LLMs, and it is often beneficial, as well as interesting, to interact directly with the model itself, completely, or at least partially, bypassing the generative aspects of the LLM. In this final chapter we’ll setup a simple Retrieval-Augmented Generation (RAG) workflow as a means of demonstrating how interacting directly with the LLM model can be used to solve common research problems.
5.1. The Vector Space Concept#
LLMs are built primarily using a Transformer model design, which is an iteration of the word embedings models first introduced in 2013 by Mikolov, et al in Efficient Estimation of Word Representations in Vector Space.
Vector models represent the semantic space of every word in a known vocabulary as large vector of n calculated parameters, where the parametric values assigned to a given word contain weights that are related to the values of the same parameter for every other word. The details of the various methods of calculating these weights is beyond the scope of this workshop. (See the paper referenced in the above paragraph.). The important thing to know in this context is that if you introduce a new word or document into the model, the weights of every parameter for every work will change. A non-mathematical way to say this is the parametric definition of each word in the model contains within it the knowledge of every other word in the model.
The below output shows the vector space for a single word from the Ollama model:
[0.017877653, 0.06340178, 0.002512074, 0.02502087, 0.014879474, 0.033396315, -0.018965052, 0.021701865, 0.022903813, 0.005726911, 0.011753164, 0.024635881, -0.036029853, -0.0122078825, -0.038973905, 0.028633859, 0.042755026, 0.0041230093, -0.007491054, -0.004649939, -0.04057417, 0.06582168, -0.040506583, 0.047047015, 0.07758565, 0.04061356, 0.08639454, 0.008880966, 0.058461666, 0.020822758, -0.002838804, 0.029212574, 0.09078681, -0.057084266, -0.024549095, -0.016134499, 0.012695589, -0.002101303, -0.009117725, -0.08302931, 0.030918889, 0.004267565, -0.0040433845, 0.003311327, -0.019085197, 0.0013708055, -0.0049502417, -0.0026343798, -0.0055629555, -0.06046351, 0.00037648194, -0.00054742245, -0.028439399, -0.00605915, -0.005068259, -0.051835686, 0.041476496, -0.030044692, -0.009939256, 0.0122636845, 0.036116987, -0.0011576054, 0.017289555, -0.02137621, -0.009717823, 0.018864619, -0.041405194, -0.031779822, 0.016003272, 0.02662322, 0.042489186, 0.05851829, -0.028756421, -0.008845935, -0.022120439, -0.038079057, 0.015947865, -0.010375119, 0.007400938, 0.040579136, 0.042395342, 0.02481092, 0.0069094235, 0.0445682, -0.03433794, -0.006127028, 0.006102693, 0.04437743, -0.005359529, -0.009452275, -0.070931956, 0.024026003, 0.021285562, 0.043804023, -0.0062165977, -0.0031585386, 0.025973827, 0.0022389963, -0.020393115, -0.047750466, -0.016769083, -0.032374572, 0.009353231, 0.05064914, -0.05532496, 0.0037234968, -0.03282322, -0.019148953, 0.020609362, -0.035215434, 0.010370555, -0.01572176, -0.0131997755, -0.018954288, 0.001769391, 0.02663567, -0.052827142, 0.042304583, -0.07665151, -0.0022814118, -0.01620359, 0.004879272, -0.004069177, 0.029104479, -0.04665595, -0.056296546, -0.058358468, 0.06820146, -0.027006276, 0.048776966, -0.0001133284, -0.024395423, 0.0070581045, 0.006074796, -0.016159529, -4.531893e-05, -0.00082628976, 0.0637154, 0.0059138928, -0.03198953, 0.03083209, 0.04447194, -0.010676811, 0.059229825, -0.010191996, -0.0028025664, 0.015581719, 0.0017894701, -0.036908768, 0.032750253, -0.00061538146, -0.00023421139, 0.023999905, 0.05311479, -0.002025976, -0.035598464, 0.05171491, 0.016205553, 0.012384458, 0.015848337, -0.018880742, 0.009939101, -0.04564323, 0.04619895, 0.016052114, 0.029844409, 0.016591715, -0.03304172, 0.010738218, -0.034117375, 0.00042674044, 0.02169864, -0.029808298, 0.047498785, -0.0035721618, 0.0036177752, 0.0024102454, 0.015826527, -0.0043423865, 0.00017197774, -0.013263845, 0.0055757593, -0.021086643, 0.021501157, -0.008960701, -0.013033894, -0.035891533, -0.0023478405, -0.0105842315, -0.029424394, -0.040108405, 0.047428276, 0.032614276, 0.004700818, -0.029580154, 0.013410349, -0.04436798, -0.002524112, -0.043401685, -0.019115387, -0.0310016, 0.042518526, 0.0011768227, -0.009699604, 0.027930323, 0.004868278, 0.042531665, 0.023388566, -0.031542692, -0.022126734, 0.02991803, 0.020402728, 0.01740285, -0.03146585, 0.0006675659, -0.043296386, -0.024070092, 0.05440069, -0.091592595, -0.015312188, 0.021202749, 0.04029528, 0.033521898, 0.0662376, 0.041259386, 0.0019152533, 0.04165173, 0.015476845, -0.027217526, -0.013091816, -0.023338709, 0.046423323, 0.021806458, -0.028718244, 0.027715893, -0.008203636, 0.034483273, -0.013777358, 0.04319412, 0.01159729, 0.037991833, -0.0030139738, -0.012506061, -0.020538852, 0.07347958, 0.021088123, 0.015355857, -0.019658286, -0.0010304014, 0.0010367308, -0.0011895543, 0.03294703, -0.009129162, 0.052312456, 0.02493031, 0.025089836, 0.02234913, 0.010121085, 0.004974439, -0.017856015, -0.04785278, -0.06071542, 0.024878934, 0.022103539, 0.060682718, -0.029474825, -0.030727727, 0.0070011443, 0.012871844, -0.036305573, -0.04627666, -0.042255938, -0.035090264, -0.025596995, -0.0038353072, -0.023908876, 0.034437925, -0.011242904, -0.037104122, -0.017154701, 0.014371906, 0.0004429378, -0.070667475, -0.021502772, 0.003996584, 0.0016655694, -0.0037696788, 0.018496107, -0.0061283647, 0.029722132, -0.030419748, 0.008249602, -0.034229085, -0.034054644, 0.0022367463, 0.006322647, -0.0008654106, -0.013552419, 0.06015992, -0.04670435, -0.017072657, 0.011510181, -0.014638132, -0.01768795, -0.045838796, 0.0074388813, 0.014886965, -0.026288446, -0.017072698, 0.018787256, -0.0076605147, 0.043190476, 0.001253275, 0.040448133, -0.046431907, 0.06465034, -0.0042557577, 0.035780855, -0.007854722, -0.033858318, 0.013364367, 0.0051386314, 0.048734456, 0.028555216, 0.03291735, 0.051106986, -0.04020619, -0.039092552, 0.0004201808, 0.0558873, 0.0020150295, -0.060250822, -0.015357823, -0.014334893, 0.038407065, 0.062022515, -0.010649503, -0.020467315, -0.055101153, 0.031293012, 0.018142227, -0.029579649, 0.06103102, 0.03194345, -0.02866799, 0.016186291, 0.025313077, -0.01804975, 0.022064071, 0.057739165, -0.020422406, 0.015787605, 0.023415362, 0.035081822, 0.03416815, -0.0037626491, -0.010472144, 0.005786563, -0.029405741, 0.020895246, -0.0357163, 0.03586897, 0.01803324, 0.0036253817, 0.023799744, -0.02189161, 0.0013853569, -0.05124004, 0.006333707, -0.028821146, 0.018659985, 0.020827137, 0.022584653, -0.026804551, -0.050562136, 0.004859809, -0.00022457466, -0.060961388, 0.034191944, -0.0520906, 0.012428855, -0.0026420208, -0.013416474, 0.006723866, -0.00907954, -0.050680887, 0.020570999, 0.07329765, -0.021525482, 0.044000793, 0.022455772, -0.054601602, -0.001437439, 0.007880718, -0.027988305, -0.01858485, -0.012942602, 0.033937123, -0.028871242, -0.037330575, 0.0034082544, -0.012213574, 0.027358947, -0.012678767, -0.0012281493, -0.008216995, 0.0098119145, -0.013236902, 0.035422873, -0.05698667, 0.026304422, 0.025371647, -0.0687202, 0.047070544, 0.0033676485, 0.01868831, -0.042346913, -0.047602106, 0.0009473163, 0.0047479523, -0.025525795, -0.06124635, -0.0013746346, 0.012058469, 0.044922817, -0.031448822, 0.019288784, -0.012294686, -0.0029466779, 0.022633798, 0.046065114, -0.032607388, -0.05304123, -0.007924773, 0.03622815, 0.01062153, 0.019646611, 0.0034436306, -0.037126176, -0.018911213, 0.018784648, 0.024862075, 0.042500857, 0.006417808, 0.019235246, -0.013034909, 0.018130388, -0.0013267407, -0.00083176536, -0.00448969, 0.018655175, -0.057097565, -0.030697485, -0.013828505, -0.05019637, 0.0043303664, -0.014013816, -0.0069711655, -0.00027030072, -0.060510688, 0.0023435664, 0.015674897, 0.05176054, -0.014777224, 0.03503745, 0.02785944, 0.029323699, -0.009299999, 0.06398674, 0.046859678, 0.030094558, 0.01259041, 0.036267675, -0.052898552, -0.0006897955, -0.042908654, -0.0008491797, 0.0071274517, 0.04439769, 0.0022068666, -0.033317126, -0.041258458, -0.0012851228, 0.010489072, 0.025735304, -0.011091368, -0.045162655, 0.04294225, -0.007700677, -0.013193355, 0.004827113, 0.03687263, 0.027334342, 0.01824656, 0.018245494, 0.02662087, -0.034855142, -0.0003569712, 0.080406114, 0.0018468174, 0.013922922, -0.018535022, 0.03824286, 0.043014415, 0.03012092, 0.00803271, -0.03232023, -0.036566664, -0.04382392, 0.0040705246, 0.0150932595, -0.0033885865, 0.019958686, -0.034229252, 0.011928538, 0.02391127, 0.04755791, -0.029175205, -0.04682454, 0.02338032, 0.031248016, -0.06120794, -0.039181888, 0.042640693, -0.008935619, -5.115602e-05, 0.025923697, 0.03046568, -0.032538872, -0.07774579, -0.047744945, -0.032523815, 0.013560552, -0.019092428, 0.031365734, -0.0022067484, 0.032985773, 0.01676509, -0.038079567, -0.011245432, -0.018426247, -0.00963427, 0.036454313, -0.04823799, -0.007587352, 0.008115187, 0.048818875, 0.0153698195, 0.0077248164, -0.01555647, 0.04640162, 0.042655896, 0.02184045, -0.034443665, -0.05343133, -0.0039155954, -0.017624898, 0.055692866, 0.021329513, 0.058465578, -0.015564522, 0.02277778, -0.012915606, -0.007953365, -0.0055068564, -0.02201194, -0.025188457, -0.016094543, 0.013643234, 0.031373184, -0.01611006, 0.005265972, 0.040668696, 0.012431018, 0.047379185, -0.010484066, -0.080858625, -0.010537491, -0.02135213, 0.060060862, -0.028551204, 0.01085753, -0.0044772346, 0.011353918, 0.069825746, 0.0012897643, 0.045575302, 0.050233044, -0.04639454, -0.043274533, -0.0088698305, -0.009285012, 0.029509, -0.057046, 0.019981181, 0.030096136, -0.06739518, -0.04110572, -0.00971853, -0.0312545, -0.051547002, 0.048084266, 0.012502023, -0.010466205, -0.004660218, 0.011354275, -0.044684965, -0.027086467, 0.023491893, 0.042202327, -0.042279933, 0.06430072, 0.0076309554, -0.043298926, 0.025694486, -0.033425454, -0.022714725, -0.037531648, 0.05346618, -0.02026543, -0.048849754, -0.007577428, 0.026761733, 0.017290449, -0.057991516, -0.016920853, -0.027190786, -0.00909436, 0.032961052, 0.017874414, -0.06763719, 0.0006165022, 0.0087364875, 0.0142304795, 0.030081693, -0.0038981324, 0.011335561, -0.0009760232, -0.062857576, -0.0141597055, 0.051928427, -0.0019329971, -0.005905841, -0.027917175, 0.024023794, -0.029292746, -0.03817091, 0.0005423172, -0.011840818, 0.044430126, 0.016287055, 0.03094146, -0.026098186, 0.0069974707, -0.048900567, 0.026985843, 0.022422804, -0.019137315, -0.037108816, 0.038627103, 0.038475305, 0.017588839, -0.04594434, -0.022467958, 0.026242614, -0.003626525, -0.005459161, 0.005077498, -0.054963883, -0.038264934, -0.032436837, -0.042729624, 0.013707538, 0.020233767, 0.038336527, -0.015694082, 0.013093894, 0.0052572903, 0.023632064, -0.08146409, 0.050827812, -0.019896386, 0.010401246, -0.06778004, 0.021781597, 0.04689888, -0.04319867, 0.016965883, 0.022130005, 0.023766624, -0.013749355, -0.0060635735, 0.06975141, 0.018952014, 0.0069972123, 0.00504026, -0.023830919, -0.037890524, -0.016333593, -0.0120985545, 0.038249396, 0.02512755, -0.033442978, 0.025615495, 0.06467917, 0.018649945, -0.0045876806, -0.029679516, 0.007530205, -0.0055214935, 0.020197572, -0.024790697, 0.03483397, -0.024737723, -0.018295601, -0.042120326, 0.01887309, -0.009808555, -0.023777742, -0.061494462, 0.01789982, 0.02538138, -0.008253647, -0.035184555, 0.03329778, 0.01434242, -0.010552813, 0.03156485, -0.012651758, 0.015550292, -0.012159424, -0.008866914, 0.03016665, 0.00047264615, 0.002137789, 0.03800385, 0.013032143, 0.038531527, -0.023826802, -0.026870178, 0.008296305, 0.011327403, 0.011354583, 0.008343737, -0.03011617, -0.01695832, -0.030502837, -0.056637384, 0.007969275, 0.005983118, -0.014082555, 0.008388151, 0.0041729757, -0.002980083, -0.02793856, -0.0040570614, -0.03271183, -0.014064461, -0.0038818293, -0.021579802, -0.011951472, -0.009213087, -0.04639352, -0.019941697, -0.00862354, -0.040463027, 0.0027768584, 0.036274806, -0.006100972, -0.019409427, 0.019028286, 0.020864759, 0.023779923, -0.043357033, 0.014085626, -0.029877884, 0.0026885546, -0.020892423, -0.0138810035, 1.2074211e-05, -0.010356221, -0.047326554, 0.0014202517, 0.041650508, 0.008155009, -0.01450383, 0.014305613, 0.06366619, -0.02081839, -0.00822323, 0.048510242, -0.0037666806, 0.023664556, 0.0045068995, -0.032586563, 0.021237116, -0.02607813, 0.042608, -0.023525318, 0.021150945, -0.0029519917, -0.008412834, -0.0152705945, 0.022217998, 0.0040978184, -0.028645568, -0.011396287, -0.04190508, -0.0040457537, 0.0373357, -0.025256395, -0.066077806, 0.00964128, -0.04682747, -0.03272878, 0.027395088, 0.024279382, -0.005157404, -0.018477073, -0.07227116, 0.0078007714, 0.033792075, 0.0087622525, 0.024143495, -0.017744526, 0.06872509, 0.029899307, 0.015984805, -0.0032637268, -0.020403294, -0.01518212, -0.03185177, -0.03993953, -0.0066711167, 0.017834764, -0.013637137, -0.034408275, 0.00044049698, 0.03190554, -0.0017063806, -0.05234689, 0.010903661, 0.013306636, -0.057937864, -0.004879621, -0.022680474, -0.057177156, 0.012787167, -0.0035027317, -0.012204662, -0.043406643, 0.0021993837, 0.029110951, 0.071350016, -0.018334549, 0.005950288, -0.042299487, 0.011649722, -0.05013199, 0.012960693, -0.0038134519, -0.015828637, -0.0036978805, -0.014976756, -0.027449084, -0.051304407, 0.03896375, -0.026833659, -0.04779689, 0.008757001, 0.022397019, -0.022991782, -0.019497303, 0.0035975624, 0.017831575, -0.016879983, 0.005123394, 0.04163835, 0.0020756787, -0.009649309, 0.022750607, -0.036457054, -0.02670016, 0.042065192, 0.019326141, 0.06558085, -0.022085935, 0.0010581234, 0.019861665, -0.011214135, -0.028121853, 0.011319331, 0.0050522136, -0.03485415, 0.030448683, 0.0029888002, -0.01445645, -0.014735276, -0.01375081, -0.048828777, 0.03580206, 0.027761148, -0.05064619, -0.011247271, 0.04609971, -0.003069608, -0.05282368, -0.034661315, 0.033407543, -0.025460843, 0.008799487, 0.0091642765, -0.008024074, -0.0047479337, -0.016529204, 0.024267333, 0.035559356, 0.07985593, -0.05046684, -0.06790705, -0.0071336497, 0.011956307, 0.009010545, -0.022069622, 0.04654688, -0.004984532, -0.05698449, 0.02789896, -0.006470848, 0.16390756, 0.008620992, 0.0386373, 0.004040578, -0.012715728, 0.0032905366, 0.03312697, -0.0031326956, 0.033616826, 0.0037641323, -0.011224089, -0.0056482824, 0.045344375, 0.034810774, 0.011302915, 0.031056117, -0.020407017, 0.021102114, -0.012683138, -0.03400104, -0.015533739, 0.03215273, 0.014327467, -0.029942159, -0.017634615, 0.044326916, 0.014144142, -0.0061987154, 0.021549068, 0.0029852581, 0.042642772, 0.012904342, 0.013874272, -0.023662219, -0.017450072, 0.024523398, -0.013644247, -0.049777966, 0.029591223, 0.0073114377, 0.007386899, 0.02843572, -0.017737953, -0.034420118, -0.057508748, 0.056002054, -0.049297955, -0.01018815, 0.014461512, -0.008235714, -0.022464707, -0.039012287, -0.009419257, -0.05011966, -0.06671197, 0.03013893, 0.0028313391, -0.0165345, -0.06355927, 0.022126166, -0.0014949412, 0.038758017, -0.02068345, 0.028776323, -0.041437782, 0.018729644, 0.037065763, -0.017174859, -0.010060867, -0.037954986, -0.015225971, -0.017563468, -0.04180092, -0.040190753, 0.009099932, 0.007841908, 0.010079515, -0.0028929978, 0.014422519, 0.041526373, -0.025540803, -0.04385223, -0.0336605, 0.016504893, 0.008888867, -0.025477355, -0.0071784807, 0.01078627, 0.026098343, 0.028917082, 0.055467572, -0.027621781, 0.015327966, 0.03297872, 0.027355123]
Given the relativistic nature of word vectors, if mapped to a Cartesian plane of n dimensions, the vectors of words that share similar meanings will be closer to each other than the vectors of words that are semantically dissimilar.
Given this mathematical reality, we can calculate how similar any two words are by calculating the cosign similarity of two word vectors. The more similar the two vectors are, the more semantically similar the two words are.
We can extend this vector space concept to deal with entire texts by averaging all of the word vectors for each text into text-vector representations of each text and then calculating the similarity of the text-vectors using same cosign similarity measure. The more similar the cosign similarities of the two text-vectors, the more semantically similar the two texts are.
This process could be extended yet again to classify and compare entire corpora.
5.2. Environment Setup#
To work directly with the embedding vectors you must download the Ollama vector model as a separate model. To do so, run the following command from your CLI.
ollama pull mxbai-embed-large
5.3. Retrieval-Augmented Generation (RAG)#
RAGs are a popular way of working directly with LLM models to leverage their power to direct people towards libraries of existing texts that relate to their specific prompts/queries. For example, Help Desk chatbots frequently use RAGS to deliver specific help documentation to users based on their prompts rather than simply generating a response directly from the LLM. Using a RAG framework, for example, one could create a chatbot that receives medical queries and uses RAG to find relevant, peer-reviewed journal papers and returns a relevant list of those articles to the user instead of an AI generated summary of everything ever said about the query.
Here we will build a simple RAG application based on Ollama text embedding vectors. For this example, we are providing the retrieval library and user prompt programmatically, but the code could be easily modified to use a training corpus that is stored on disk and to receive the prompt programmatically (as we did in our chatbot example earlier).
Our first step is to load the libraries that we will need for our RAG into our Python environment.
import ollama
import numpy as np
from numpy.linalg import norm
Now that our environment is ready, let’s define the corpus of texts that are available for retrieval and delivery based on our prompt.
documents = [
"Llamas are members of the camelid family meaning they're pretty closely related to vicuñas and camels",
"Llamas were first domesticated and used as pack animals 4,000 to 5,000 years ago in the Peruvian highlands",
"Llamas can grow as much as 6 feet tall though the average llama between 5 feet 6 inches and 5 feet 9 inches tall",
"Llamas weigh between 280 and 450 pounds and can carry 25 to 30 percent of their body weight",
"Llamas are vegetarians and have very efficient digestive systems",
"Llamas live to be about 20 years old, though some only live for 15 years and others live to be 30 years old",
]
We’ll also here provide our hard-coded user prompt. As noted earlier, in a real-world application this would be dynamically received, as in our earlier Chatbot exercise.
prompt = "What animals are llamas related to?"
Next, we need to gather the embedding vectors for each of the training texts in our documents collection.
# create an empty list to hold the returned embeddings for each document
embed_list = []
# store each document embedded space in a list
for d in documents:
response = ollama.embed(model="mxbai-embed-large", input=d)
embeddings = response["embeddings"][0]
embed_list.append(embeddings)
The embed_list object now holds a collection of document vectors, one each for each of the texts in our training corpus.
At this point, we are ready to process a user query to our RAG. We’ve already defined our text query in a prompt variable. Our first action is to generate a vector that represents the embedded space our prompt.
# get the imbedded space for the prompt
response = ollama.embed(
model="mxbai-embed-large",
input=prompt
)
# get just the embeddings vector from the response
prompt_vector = response["embeddings"][0]
Our prompt_vector variable now contains the vector representation of our prompt. With this in hand, we can now compare the prompt_vector to each of the training texts vectors using a cosign similarity measure. The most similar vector represents the text from our document set that most closely semantically matches our prompt.
# set some base values
max_similarity = -1
most_similar = -1
# loop through our training data and calculate vector similarity to target
for n in range(len(embed_list)):
# calculate cosign somilarity between prompt vector and training vector
cosine_similarity = np.dot(embed_list[n], prompt_vector) / (norm(embed_list[n]) * norm(prompt_vector))
# check if this is more or less similar than last vector and update return values accordingly
if cosine_similarity > max_similarity:
max_similarity = cosine_similarity
most_similar = n
Now that we’ve identified the most similar text in our document library, we return that document to the user as the answer to their prompt.
if most_similar > -1:
print(documents[most_similar])
else:
print("Nothing similar")
5.4. Putting it all together - the complete RAG#
The code below presents the complete code for our simple RAG workflow. Inline comments explain each step of the RAG process.
import ollama
import numpy as np
from numpy.linalg import norm
# define a set of training documments. [These would normally be loaded from a collection
# of files on the computer]
documents = [
"Llamas are members of the camelid family meaning they're pretty closely related to vicuñas and camels",
"Llamas were first domesticated and used as pack animals 4,000 to 5,000 years ago in the Peruvian highlands",
"Llamas can grow as much as 6 feet tall though the average llama between 5 feet 6 inches and 5 feet 9 inches tall",
"Llamas weigh between 280 and 450 pounds and can carry 25 to 30 percent of their body weight",
"Llamas are vegetarians and have very efficient digestive systems",
"Llamas live to be about 20 years old, though some only live for 15 years and others live to be 30 years old",
]
# set a user prompt (normally we'd do this through a user input)
prompt = "What animals are llamas related to?"
# create an empty list to hold the returned embeddings for each document
embed_list = []
# store each document embedded space in a list
for d in documents:
response = ollama.embed(model="mxbai-embed-large", input=d)
embeddings = response["embeddings"][0]
embed_list.append(embeddings)
# embed_list now contains a list of document vectors for the augmenting documents
# get the imbedded space for the prompt
response = ollama.embed(
model="mxbai-embed-large",
input=prompt
)
# get just the embeddings vector from the response
prompt_vector = response["embeddings"][0]
# set some base values
max_similarity = -1
most_similar = -1
# loop through our training data and calculate vector similarity to target
for n in range(len(embed_list)):
# calculate cosign somilarity between prompt vector and training vector
cosine_similarity = np.dot(embed_list[n], prompt_vector) / (norm(embed_list[n]) * norm(prompt_vector))
# check if this is more or less similar than last vector and update return values accordingly
if cosine_similarity > max_similarity:
max_similarity = cosine_similarity
most_similar = n
# return the most similar document fromt the training set
if most_similar > -1:
print(documents[most_similar])
else:
print("Nothing similar")