Collaborate on Typical Invoice Format for Technical Support with Ease Using airSlate SignNow

Typical invoice format for Technical Support

hi many of you watching this video might already  know how AI tools are transforming the way we   extract and process information but today we  are diving into something really exited that   is structured data extraction so we are going to  use new Med new extract model which is available   in Huggingface it is an open source model it is  an open source large language model specifically   designed to make structured extraction from unst  structure text with more in a more efficient way   so whether you are working on a complex documents  or academic papers or any other type of text heavy   content this model can pull out the data you  need in organized or structured format so now   in this video I will walk you through how new  extract works and how you can implement it for   your one use cases so this model was basically  trained upon Microsoft PHI the specific use   case so now you know like when when working with  traditional tools the data extraction can often be   a pain it is either to rigid or not accurate  enough or it just takes too much of manual   intervention I experimented with the different  approaches but the solutions feel either too   limited or not flexible enough when dealing with  complex documents nuextract is open source it's   highly flexible and specifically optimized for  structured extraction tasks and since it uses a   power of llm's it doesn't just stop at keyboard  matching it understands the context which makes   a huge difference when you are dealing with a  Nuance or messy text in fact I recently came   across a real world use case where developers are  using nuextract to streamline data extraction for   invoices and HR documents and after experimenting  it with myself I realized how powerful and easy to   use it so in this video I will share how we can  deploy a new extract in AWS Sage maker and get   inference from it and also test it with uh using  a postman so let's let's quickly talk about new   extra version 1.5 by neumed it's a powerful open  source a model designed for structure information   extraction so here is a model card so um the key  things you need to know about this model is it's   a multi- language support it works not just in  English but also in French Spanish German and   Portuguese as well as Italian it hand long uh  documents so by using uh new extract version   1.5 you can easily extract long documents even  with up to 20,000 s to but you should you   should deploy it in your own endpoint so then  only it will allow model what they have given   for for a demo purpose it it doesn't take that  much s so um if it's 20,000 s you can   input large reports or research papers this is a  pure extraction so the model doesn't make things   up it only pulls information directly from the  input text which makes it super reliable um and   also one more thing is that uh it is a we can  customize the Json templates so you can tell   the model what to extract by using a simple Json  template for example if you're working on invoice   your template can be specify the fields like  name date amount this is optimized for accuracy   for the best performance set the temperature to  zero so that it will extract only the required   data so um this is mainly for focus on precise  extraction there is also new new extra tiny   version for lightweight task so but today we are  going to uh use the version 1.5 which is which   has more power and flexibility you can even try  the model for yourself using a playground demo in   huggingface you can experiment with it so just any  of these and then yeah so they given uh different   different um templates with with Json template  and the text let us test uh the first one okay   so this is the text so uh this is the model  template how we are going specify so it will   extract uh based upon the text what the parameters  needed like based upon the model template so um   let us click submit is going to give this same  kind of output let us see yeah so this is uh just   a demo which is which is given in the huggingface  itself but we are going to uh deploy this model in   Amazon sagemaker and then get inference from  made let us see how we can do that so let's   let's quickly talk about nuextract version 1.5 by  numed it's a powerful open source a model designed   for structure information extraction so here is  the model card so um the key things you need to   know about this model is it's a multi- language  support it works not just in English but also in   French Spanish German and Portuguese as well as  Italian it hand long uh documents so by using uh   new extract version 1.5 you can easily extract  long documents even with up to 20,000 s but   you should you should deploy it in your own um FR  endpoint so then only it will allow uh the model   what they have given for for a demo purpose it it  doesn't take that much s so um if it's 20,000   s you can input large reports or research  papers this is a pure extraction so the model   doesn't make things up it's only pulls information  directly from the input text which makes it super   reliable and also one more thing is that uh it  is a we can customize the Json templates so you   can tell the model what to extract by using  a simple Json template for example if you're   working on invoice your template can be specify  the fields like name date amount this is optimized   for accuracy for the best performance set the  temperature to zero so that it will extract only   the required data so um this is mainly for focus  on precise extraction there is also new new extra   tiny version for lightweight task so but today we  are going to uh use uh version 1.5 which is which   has more power and flexibility you can even  try the model for yourself using a playground   demo ning pH uh you can experiment with it so um  just select any of these and then yeah so they   given uh different different um templates with  with the Json template and the text let us test   uh the first one okay so this is the text so  uh this is the model template how we are going   specify so it will extract uh based upon the text  what the parameters needed like based upon the   model template so um let us click submit is going  to give the same kind of output let us see yeah so   this is uh just a demo which is which is given  in the huging phas itself but we are going to   uh deploy This Modern huging phas sorry Amazon s  maker and then uh get inference from it let us see   how we can do that so before we jump into deploy  new extract model let us quickly talk about aw a   a maker is a fully managed service from Amazon web  services that allows you to build train and deploy   machine learning models at scale so um whether  you are a beginner or experienced developer Sage   maker make it easy to to take a machine learning  model like uh new extra and turn into working   solution um with as little effort as possible so  here we we have some things uh that Sage maker is   is very good at it that is managed infrastructure  you don't have to worry about setting up servers   or ensuring the scalability sage maker handles all  the lifting for you easy model deployment working   with aing FES tensor flow or pouch Sage maker  provides a buil-in support to deploy your model as   realtime endpoints or batch job services it's also  cost efficient see you only pay for the resource   you use which makes it great for anyone with  budget constraints but you should be very cautious   about that one so whether you are experimenting  or developing or running a production workloads   it's it's it's good at that point and also the  main thing is the integration with agging face   models so s maker has a direct integration with  agging face and makes deploying models from their   up directly like uh we will see it how we can  deploy uh while deploying a new extract model it   is simple and seamless so now let's move on um how  to navigate aw s maker from AWS console so now if   you open the console login since I already book  Market you can go here and just type Sage maker   so just click on that one it will open the stage  maker there yeah so before we jump into deploy new   extract model let us quickly talk about a stage  maker is a fully managed service from Amazon web   services that allows you to build train and deploy  Ma machine learning models at scale so um whether   you are a beginner or experienced developer Sage  maker make it easy to to take a machine learning   model like uh new extra and turn into working  solution um with as little effort as possible so   here we we have some things uh that Sage maker is  is very good at it that is managed infrastructure   you don't have to worry about setting up servers  or ensuring the scalability S maker handles all   the a Lifting for you easy model deployment  whether you're working with aing f tensor   flow or P Sage maker provides a built-in support  to deploy your model as real time endpoints or   batch job services it's also cost efficiency you  only pay for the resource you use which makes it   great for anyone with budget constraints but  you should be very cautious about that one so   whether you are experimenting or developing or  running a production workloads it's it's it it's   good at that point also the main thing is the  integration with a face models so s maker has   a direct integration with aing face and makes  deploying models from their up directly like we   will see it how we can deploy uh while deploying a  new extract model it is simple and seamless so now   let's move on um how to navigate aw s maker from  AWS console so now if you open the console log in   since I already uh bookmark it uh you can go here  and just type Sage maker so just click on that one   it will open the sage maker blade yeah so now we  are in AWS Sage maker so this left is sagemaker   blade where we can navigate uh into different um  applications like Studio canvas all this stuff you   can explore it but now in this uh video we're  going to aage Sage maker studio so first thing   you need to do is create a sage maker domain so um  this will create a a domain where it will create a   user and then it will give the permission to  the user to access other services as well in   order to uh do the development but um in in this  case if you want you can create a new domain but   I already created a one so I will go to e one so  I have the domain already created for me so this   what we call a default domain once you we create  the domain you can you can go to Open studio and   you can you can you can start developing but  if you want to do like instead of aw to create   domain for you to to Quick Start you can start  yourself like by by giving a fine gr necessary   permissions just create domain and you can start  creating it so I have the domain so I will just go   to The Domain okay so wait wait Studio okay open  Studio since we already created the domain we're   going to open the studio okay accept skip the tool  for now you can take the tool so now um here we   have um we are in stage maker studio so we have  applications where it will be used develop the   um the flow or you can develop using Jupiter  lab or Studio the code and then and then test   it um these are the the compute and what are the  things you are going to use uh for for preparing   uh machine learning model pipelines models all  those stuffs are there so you can explore one   by one but today we are going to see or leverage  Jupiter lab so I'll click Jupiter lab so already   I started um I tested previously so there are  two which is which is which is uh created and   then now it is in stoer state if you want to  create one just click on this uh create jbit   app space and then um you can see the options how  we can create so um in order to know more about   this I already created a video uh you can you can  you can check on that one so now we are going to   um this aing phase uh Jupiter lab it's stopped I'm  going to run it so it will take some time let us   give some time and I'll I'll come back later yeah  so still it is starting um if you go to the actual   specs uh so I'm using uh ml g5x Lodge which is  gpus so if you go to the spe here so g5x large   these are the pricing so please be aware of the  pricing as soon as you finish you can you can stop   the uh instance so I'm using g x lar with 4 vcpus  having a 16 GB of memory so now uh uh this like   now since we are using uh jup for for uh for for  developing the U code but if you want to actually   create the inference endpoint for this particular  model we need to see the model size actually okay   so if you go to model code and then go to files  and versions here we have uh the size of the   model where there are two uh safe tenses one is  our five and other3 so almost 8 GB so if you use   G5 x large um it's comfortably sits into that  dpu where we have 16 GB of memory so that's one   thing you need to know and and let us see how we  whether it started still it starting let us give   some more time and I will come back yeah so now  you can see it is fully started so um I go to lab   and then open yeah so this this is 4 53 okay okay  so um now let us develop the code um we have to   create inference endpoint and and test it so this  is a okay so let us start from here um yeah okay yeah so uh now we have the um development  devopment ready so I just opened Jupiter   lab notebook and then I'm um developing the code  here so first I have configuration data where I   have um a region and AG pH um L language any  any authentication okay so now we are in AWS Sage   maker so this is left is Sage maker blade where  we can navigate uh into different um applications   like studio canas all the stuff you can explore it  but now in this uh video we're going to leverage   Sage maker studio so first thing you need to do is  create a sage maker domain so um this will create   a uh domain where it will create a user and then  it will give necessary permission to the user to   access other services as well in order to uh do  the development but but um in in this case if you   want you can create a new domain but I already  created a one so I will go to East one so I have   the domain already created for me so this what  we call a default domain once you we create the   domain you can you can go to Open studio and you  can you can you can start developing but if you   want to do like instead of AWS to create domain  for you to to Quick Start you can start yourself   like by by giving a fine gr necessary permissions  just create the domain and you can start creating   it so I already have the domain so I will just  go to The Domain okay so wait wait Studio okay   open Studio since we already created the domain we  going to open the studio okay accept skip the tool   for now you can take the tool so now um here we  have uh um we are in stage maker studio so we have   applications where it will be used to develop the  um the flow or you can develop using Jupiter lab R   Studio the code and then and then test it um these  are the the compute and what are the things you   are going to use uh for for preparing uh machine  learning model pipelines models all those stuffs   are there so you can explore one by one but today  we are going to see our leverage uh Jupiter lab   so I click Jupiter lab so already I started um I  tested previously so there are two which is which   is which is uh created and then now it is in stoer  state if you want to create one just click on this   uh create jupit La space and then um you can see  the options how we can create so um in order to   know more about this I already created a video uh  you can you can you can check on that one so now   we are going to um this agging phase uh Jupiter  lab it's stopped I'm going to run it so it will   take some time let us give some time and I'll I'll  come back later yeah still still it is starting   um if you go to the actual spec uh so I'm using  uh ml g5x Lodge which is gpus so if you go to the   spe here so GX large these are the pricing so  please be aware of the pricing as soon as you   finish you can you can stop the uh instance so I'm  using GX large with four vcpus having a 16 GB of   memory so now this like now since we are using jup  for for for developing the U code but if you want   to actually create the inference endpoint for  this particular model we need to see the model   size actually okay so if you go to model code and  then go to files and versions here we have uh the   size of the model where there are two uh safe  tenses one is our 5 and other 3 so almost 8 GB   so if you use G5 x large um it's comfortably sits  into that U uh GPU where we have 16 GB of memory   so that's one thing you need to know and and let  us see how whether it start still it starting let   us give some more time and I will come back yeah  so now you can see it is fully started so um I go   to lab and then open yeah so this is 4.3 exct okay  okay so um now let us develop the code um you know   to create inference endpoint and and test it so  this is a okay so let just start from here um yeah okay yeah so uh now we have the um development  inment ready so I just opened Jupiter lab notebook   and then I'm um developing the code here so first  I have configuration data where I have um AWS   region and agging pH so this aging pH   especially be need when any authentication okay   hey so now we have the development envirment ready  so um this is jupit notebook where I have uh a   configuration data which is in the form of a Json  where I have the envirment like en varable stored   here it contains ad default and a so this is  necessary in case if you need any authentication   from AG side so let us run one by one okay we  have default Ag and AG toen so I'm just paing   those things here and uh this is the actual code  what we require in order to the inference in point   okay so let us first run this so that it will  take some time I will expl in the meantime okay   so um it has already started creating end point  Sage maker so these are just Library inputs and   now we are um using Sage maker role as a role  what if you see like we already have a domain   here if you go to Sage maker Studio okay so this  domain consists of um all the required necessary   uh user profiles in order to access and then  uh um the app configurations everything was   done already if you use um uh the the the domain  which which we already created using uh uh create   domain button so now if you go here yeah so that  is from there it will it will get the IM role as   well as uh this is the main role which we need  for um Sage maker execution yeah so we importing   um libraries and then um we need a rle um these  Ro which is used to execute on Sage maker from   this Jupiter lab notebook or the dment and the  role is Sage maker execution role this is the   role necessary with as necessary permission uh  to execute um so if you go to the domains we   already shown like it is created let me see yeah  default execution to sageer execution to okay and   then um this is the like uh we are defining which  aging face model we are going to use so if we go   to the model card this is the model actually just  copy it and you can get the model card this is   new new X verion 1.5 and number of gpus so we are  going to use only one GPU and then um um so again   P model so we are getting the image from ning face  and then en from this here and then Ro we already   declared Ro here okay so now predictor theing pH  model deploy we are going to having instance count   of one instance type is MLG FX large so this is  the one we need since we have 16 GB this is this   is enough to to have 8 GB large large language  model and uh this is a timeout check for 300 Mill   seconds okay so now um let's go to Sage maker and  see the inference in points it should be starting   yeah so it is creating the inference end point it  will take some time to create inference end point   so let us wait until that time and then then I  will come back he so now we have the development   envirment ready so um this is Jupiter L notebook  where I have uh a configuration data which is in   the form of a Json where I have the envirment  like en varable stored here uh it contains ad   default and aish so this is necessary  in case if you need any authentication from   aing side so let us run one by one okay we have  default so I'm just Wasing those things here   and uh this is the actual code what we require  in order to the inference in point okay so let   us first run this so it will take some time I  will explain in the meantime okay so um it has   already started creating Ence in point Sage maker  so these are just Library inputs and now we are um   using Sage maker role as a role what if you see  like we already have a domain here if you go to   maker Studio okay so this domain consists of um  all the required and necessary uh user profiles in   order to access and then um the app configurations  everything was done already if you use um uh the   the the the the domain which which we already  created using uh uh create domain button so now   if you go here yeah so that is from there it  will it will get the IM r as well as uh this   is the main role which we need for um Sage maker  execution yeah so we importing um libraries and   then um we need a role um this is the role which  is used to execute on Sage maker from this Jupiter   lab notebook or the dment and the role is Hemer  execution role this is the role necessary with as   necessary permission uh to execute um so if you go  to the domains we already shown like it is created   let me see see yeah default execution to Sage  maker execution to okay and then um this is the   like uh we are defining which aging face model we  are going to use so if you go to the model called   This is the model actually just copy it and you  can get the model code this is new new version 1.5   and number of gpus so we are going to use only one  GPU and then um um so model so we are getting the   image from face and then en from this here and  then R we already declared Ro here okay so now   predictor theing pH model deploy we are going to  having instance count of one instance type is MLG   FX large so this is the one we need since we 16 GB  this is this is enough to to have 8 GB large large   language model and uh this is a timeout check  for 300 M seconds okay so now um let's go to   Sage maker and see the inference end points it  should be starting yeah so it is creating the   inference end point it will take some time to  create inference endpoint so let us wait until   that time and then I will come back yeah so  uh now we have the inference endpoint ready   so it's already in service so it took some time  almost uh 5 to uh 7 minutes okay so this is the   inference endpoint let us copy this okay so here  we need the endpoint name so that is only for   HTP request we go to endpoint name this is the  endpoint name we want let's copy this have it   here so this is the code how we invoke entrance  end point by pausing a text and then giving the   template in the format how we need or how we need  to extract the information from the given text   um so this is small function where we send the  payload and then prompt plus uh the param M if   necessary and then we are invoking the end point  getting the inference from there okay so let us   run this one yeah so uh dror patient discussion it  h the end point and we get the inference okay so   initial observation symptoms temperature s thought  okay so I just took the same example what we have   already in the uh inference examples here if we  go to inference yeah so this one I took this one   so we got the output let us try with this and  change the model card sorry template card as   well okay yeah so uh the template uh we need to  change ing to the requirement let us go   here and then get this gra this template from here  okay hopefully this will work okay let us try this   okay so we are issue in the template okay this is  correct yeah I found the issue like we have some   like previously we given the template wrongly so  we couldn't be able to the inference correctly so   I I changed the uh template and now if you see  we get the output as requested so we gave the   test text the uh text is about v1.5 I just grabbed  from here this one and the template given what is   model name number of parameters number number of  uh so it currently get the inference like it   paed on and then um we extracted the text based  upon our template so now the same thing we will   try to test using Postman let us see how we can  do that yeah so uh now we are the inference end   point ready so it's already in service so it took  some time almost uh 5 to uh 7 minutes okay so this   is inference end point let us copy this okay so  here we need the endpoint name so that is only for   HTP request we go to the endpoint name this is the  endpoint name we want let's copy this have it here   so this is the code how we invoke inance endpoint  by pausing a text and then giving the template in   the format how we need or how we need to extract  the information from the given text um so this   is small function where we send the payload and  then prompt plus uh the parameters if necessary   and then we are invoking uh the endp point getting  the inference from there okay so let us run this   one yeah so uh dror patient discussion it h the  end point and we get the inference okay so initial   observation symptoms temperature s thought okay so  I just took the same example what we have already   in the uh inference examples here if we go to  inference yeah so this this one I took this one so   we got the output let us try with this and change  the model card sorry template code as well okay so   uh the template uh we need to change ing to  the requirement let us go here and then get this   this template from here okay hopefully this will  work okay let us try this okay so we have issue   in the template okay this is yeah I found the  issue like we have some like previously we given   the template wrongly so we couldn't be able to the  inference correctly so I I changed the uh template   and now if you see we get the output as requested  so we gave the test text uh text is about v1.5 I   just grabbed from here this one and the template  given what is model name number of parameters   number number of uh so it correctly get  the inference like it paed on and then um we   extracted the text based upon our template so now  the same thing we will try to test using Postman   let us see how we can do that yeah so now we are  going to get inference um test inference endpoint   using postp so let us go to post method and this  is a text this is a this is the text I got from   Wikipedia put a ballot of a play so from from  this text we are going to extract name of the   ACT director location City when all this stuff let  us see how we can let us see how how will it it   extract the test um sorry extract the information  so let us go to P maker and grab the in point okay   grab the URL so this you can you can get it from  the inference end points the invocation URL and   let us send so before that um in order to access  inance endpoint um from your developer machine um   to the salemaker in order to send a post request  to the stemer inference endpoint you need you need   to have a proper authorization um so here I I  am using a signature as Au type if you select   here you can see but um you need to create a new  access key and secret key um no need to give any   service name if you want you can give a Sage maker  and then yeah so if we go to body and then let us   send the okay we are getting an error what is it  missing field input fail to G so there is some   issue in Json let me figure it out what's going on  there and then I will come back so previously we   messed up with the format but now um we need  to give in a correct format to the inference   and point especially how the new extract expects  so then only we can able to uh get the output or   the extracted uh data from the text so um these  are the input and in the inputs I I give in this   format input and then the template and then and  then finally the text okay so let us try with   this one actually I got the output correctly  let's try one more time yeah so the required   output is that um output the funding so inside the  funding how much the funding 27 million invested   investor and then KB investment uh these are the  investors and then company name identify activity   so these are this this um uh output is what we  expected this is the sample text and template I   grabbed it from the new text new new extract model  model page itself I will show you yeah so this is   the example I chose um yeah so this one yeah so  if you see um this is the template what we need   like uh what the information to be extracted so  funding new funding investor all the stuff so   this is the text given so um our INF end point  be the same way as expected it given all these   values so hope you find Value from this video so  you can further fine tune uh you can further use   this model we extract uh complex documents uh with  template um I think you need to play around with   this and then uh get the um the point where how  you want to PR the template and how much is the   text size and also if you go to the uh model card  let me show you if you go to the model card this   spaces okay if you go to the model card there is  one more option where if there the text size is   more or if you want to process more of text so  they give a option called sliding window promp   so um this one it will take uh you can you can  give the you can specify maximum input size   and maximum new s as well and then split the  documents they gave the sample uh code you can use   this one and try or you can find you can you can  tweak to fit ing to your uh use case yeah   and then one more important thing is that once you  done uh developing uh you need to um terminate the   end point point or delete the endpoint so just  run this code so before that I will show you it   is running now you see the endpoint is in service  you go inside the endpoint you can see all the   memory so we we fired some queries post queries so  that it gives the uh the percentage of utilization   here um yeah so all these things you can see here  okay so now if you go to the endpoint it's in   service so make sure you delete the endpoint  before um uh like once you tested everything   you developed the code you tested it and then  make sure you delete the endpoint otherwise you   will the cost you already saw the cost it it's  almost $1 per hour so make sure you delete the   endpoint okay so as soon as we did now if you go  and check if there are end points there are no   current resources so it's in it's deleting the end  point started deleting the end point now yeah so   this is what I want to share you with today hope  you like this video please subscribe thank you