hey everybody this is Nic Chaillan I'm the former Chief software officer for the Air Force and Space Force I'm also the founder of Ask Sage bringing GPT to government teams and integrating it with custom datasets live queries to be able to bring these capabilities to government teams and contractors to accelerate your capabilities and save you time. At the time of this recording we have about 3 400 government teams and 750 companies both financial institutions and defense industrial base companies and government contractors that signed up for Ask Sage we're so thankful in three months to see such massive volume of engagements and we know you have questions so we wanted to share lots of insights today and pitfalls to avoid and things to pay attention to so let's get started! All right so the first step for you is to go to chat.asksage.ai to go and create an account so as you can see here very simple you go to the website and you register make sure you don't lie on the fields here and if you do have a CAC or PIV make sure it's inside of your device when you do the registration here but effectively you're gonna put your name first name company email phone and select your country make sure you read the terms and condition and then you're gonna subscribe you're gonna get a code in your email to validate if you don't get the code let us know we can force your account so just send us an email at support@asksage.ai will be able to help you to validate your account if you have any issue. all right so now that you have created that account you can log in either with a CAC here and simply click here after a registering first but if not you can enter your email and password and we'll show you later as well how to activate multi-factor authentication if you don't have a CAC or PIV as well. All right so the first thing you see when you log in is this pop-up here we
always tell you to read it because there's a lot of great insights also when you scroll down you're gonna see all the new features we're adding and a lot of great comments so check that out also watch those videos if you want to learn more about Ask Sage and if you have issues always you can find our Discord Community here and the support email and sales email as well. All right so first let's talk a little bit about Ask Sage architecture I know a lot of you have questions and that's understandable obviously first for us it was very important to be model agnostic so we do support open AI GPT 3.5, 4, DaVinci, GPT4 32k tokens but we also support cohere Dolly V2 Google Bard and any other large language model that we can use so we're agnostic and the beauty there is the training we do on top of those models is also agnostic you can train data and then ask questions to any of these models and it's going to have the same insights that we taught it so that's really enabling you not to get locked in to those large language models and try different things and see what sticks Ask Sage is hosted on Azure gov Impact Level 5 for CUI that's where we store all of our data, datasets that's a multi-tenant stack we also have an ability to host a dedicated tenant we'll talk about that later but the multi-tenant stack is on Azure gov and we have our dedicated Enclave with Azure Open AI with a dedicated endpoint that is on the FedRAMP high commercial regions where we brought all the controls to make it capable of doing impact level 5 work so while the Ask Sage stack is on Azure government the API for the Open AI models is on the FedRAMP High commercial regions and we have specific settings thanks to our partnership with Microsoft to do a fire forget API so effectively it's not logging any information, humans have no access to the prompts and responses but also so the data is never used to train the model so effectively when you ask a question to the bot using Open AI it's a fire and forget API so it does not remember what he just said and that's the beauty of what we built makes us capable of doing impact level 5 work and the way we make it remember what you ask is by passing again the history of the chat every time we have a conversation with the API as you can see Ask Sage has many datasets that we ingest and share across users but your custom datasets are only visible to you so if you train the bot on specific topics and create datasets they are only visible to you and to the people you decide to share those datasets with and that has to be done through us by the way for security reasons but know that when you train the bot on dataset this is not visible to other users and so effectively it's almost like having your own dedicated bot dataset experience on top of those large language models like Open AI. Obviously, I care very much about security and data centricity is essential we're built on top of zero trust we have this model of labeling data each dataset has a label that can be then shared to other users but by default they're only visible to you and that enables you to decide how you assign data and how you cut it into different datasets by creating your own data labels and we'll show you how to do that all right so why did we create Sage well you know like everybody we started playing with GPT back in November pretty much October started using it to write our video scripts and that was great I became kind of a pretty face reading the video scripts and doing all videos that was exciting but what really got us really excited is when we started to use it for Mission work and let me show you exactly the moment where I realized that this was way more than just some gimmick capability writing content for us in this case what really got me excited is when we had this data set with Chinese resumes of the CCP teams and it's unclassified and open source and you know they have 150 fields of different information but you know what was pretty mind-boggling is the volume of information and if you put this on Google translate you get 150 fields and your brain is not able to really understand who this person is it's just too much information and so we'll show you what I did right so we simply did translate and summarize who this person is and this is a Json document here in Chinese so it's not in English other than the fields names and simply by asking it to translate and summarize and that's really where I decided to make this a company is seeing the immense value where now GPT is able to give us a clear rundown of who this person is in plain English and that was really mind-boggling you know the human brain would not be able to do this simply by translating things on Google Translate for example and so as you can see now we can pretty much picture who this person is that's why we decided to create Ask Sage we realized it was way more than just writing simple things and it was capable of doing pretty incredible things to really augment and empower people to get things done and bring tangible outcomes to the warfighters you know for me it's actually pretty interesting I don't Google anymore you know this is re-augmenting our time by you know 80% on average a day at least for developer we estimate you know a 10x ratio for teams responding to RFPs for example we've seen them be able to augment their volume by three to seven X on average which is pretty incredible if you look at GPT and the value it brings to the table it's really almost like a a company as a service capability what I mean by that is it's able to do pretty much anything if you look at Sage itself the logo of Sage was created by the bot the UI 90% of it created by about 90% of the backend and 100% of the SQL stuff 100% of the CI CD DevSecOps pipeline 100% of legal. you know if you look at all the capabilities we built really everything was orchestrated and empowered through the use of Ask Sage to improve itself almost like like an exponential capability so obviously a lot of people are worried about GPT a lot of issues a lot of challenges is it perfect no obviously not but can you drastically improve your velocity and your outcomes no doubt and we're just getting started you know things are improving every day and it's kind of mind-boggling to see all the research coming out from scientists and data scientists on the subject matter and it's really slowly but surely becoming more and more safe but also more accurate and factual all right so here let's look at the UI here a little bit as you can see we have different conversations and it's easy to rename them and we can easily do that and delete former chats ideally you want to create a new chat every time you change topic or you swap between personas and models we'll talk about that today you can edit you can delete and of course you know the conversation is kept but do keep in mind that because of the token limit of the models we only pass about five to ten of the previous messages to the query so it's not always tracking the entirety of your chat so try to get things done in as little prompts as you can ideally in one prompt or two but don't stop following up and keep asking questions it's always better to get to outcomes in the in the small number of questions all right so what is a token one token is about four English characters but that changes also with coding and Chinese and other languages so an easy way to see how many tokens you just entered if I write this is a test you'll see here that the bottom right here is going to show us this is 18 tokens of text right so this is a good way for you to try to estimate the volume of of tokens all right so let's start with a simple query here we're going to ask it who I am and see what kind of answer we get just to show you how this works of course we trained the bot on who I was so it has all that Insight so you know what's interesting obviously is we get references so we know where that came from and we have follow-up questions here so you could ask you know additional questions like what is the DoD Enterprise DevSecOps initiative that Nic Chaillan co led and this is going to give us additional insights here and give you additional references for that question as well so this is a very easy way to start asking questions like we talked about always start a new chat if you change topics you can just do /new with the command and create a new chat or you can also clear the chat by doing /clear as well all right so now let's let's look at a little maybe a more interesting example maybe writing some code so we're going to ask it to write the code for kubernetes nginx Ingress that leverages a mutual TLS authentication to authenticate users with a common access card and let's see what he's going to give us all right so he just gave us the code of that Kubernetes Ingress and it did add the verify depth and the verify client which is exactly how you would do this to activate and the TLS secret will be obviously passing the root CA with all the DoD certificate here which is exactly how we actually did build this CAC authentication feature in Sage for the DoD and gov so pretty cool as you can see it's giving us also some additional information here and you can copy the code here and you have a copy the entire block here but you also have a thumb down down button here to notify us when there is a bug don't use it when the answer is wrong we can't help that but notify us when you see a bug and there's an issue and you can just click this thumb down button here to help us fix it all right so first let's look a little bit at the different models like I said we are model agnostic so by the time you're watching this video we may have new models but at least for these models that we have today you'll want to be able to know which one to use they have different price points so you're gonna pay more if you use GPT4 or 32k than if you were using GPT3.5 for example when you buy an account with Ask Sage you get DaVinci token which is kind of the middle tier price point but these tokens can be exchanged just like a currency into any of these models so you do buy tokens 500 000 tokens for the 30 dollars a month account but keep in mind this gives you five million token with GPT3.5 which is way cheaper it's 10 times cheaper than DaVinci so
you get 5 million token I know if you were to use GPT4 keep in mind that GPT4 is about five times more expensive than that DaVinci or 50 times more expensive than GPT3.5 so you're getting you know only a hundred thousand tokens when you have the paid subscription of GPT4 and gpt4 32k is nine times more expensive than DaVinci or 90 times more than GPT3.5 so you need to use the right model for the right job DaVinci is going to be less biased is going to be able to do more things so when you hear the model tell you oh I can't do X Y and Z, clear your chat do a new chat and try again with the same question with DaVinci and it should be able to do it now Cohere is another commercial model it's quite limited but it's it's a an option that you can use and it's the same price as DaVinci so when it comes to tokens keep in mind you're paying not just for the question you're asking but also the data we passed to the bot that we ingested that could be datasets or insights we have but also the reply that you're getting back right so you're paying tokens for all of the above and we can only estimate what you're typing we don't know how much we're going to pass in terms of data to that and we don't know also how much the bot is going to respond to that so it's impossible for us to predict how much one question is going to consume of tokens but at least here at the bottom right if I copy this text here and I copy and paste it into the the box here is going to tell us exactly here that it's 169 tokens all right so when you sign up for a free account you get 25 000 tokens and you don't get all of the plugins we have so if you want to pay for a paid account and get the 500 000 tokens so higher limit you can simply click on become a Sage customer today and you're going to be redirected to all stripe payment and you can use a credit card to pay and for government teams you can use the government purchase card for up to ten thousand dollars so you can actually pay for up to 27 users per team so that's very easy to to do directly there and reach out to us if you want to do bulk purchases you don't have to put the credit card for each account we can do that for you in the back end all right so let's look at one of the biggest impediment that is called hallucination it's when the bot is making stuff up it's just creating text and information that is just not accurate so let's look at an example and let's create an hallucination here I'm going to use a da Vinci model because it's more likely to hallucinate I'm going to say something as simple as so I'm going to ask who is Austin Bryan from Ask Sage and provide links with more information and it's wrong obviously because Austin Bryan works for defense unicorns and not Ask Sage so that's gonna make it hallucinate right we're misleading it to tell it he is working for Ask Sage and so here what's interesting right is that it gave us links but these links are not working and there are actually hallucinations so he's telling us hey this might be a sign that the answer might be a hallucination so that's something we build at Ask Sage to try to mitigate all these issues with hallucinations but let's try to ask the same question now on the GPT3.5 model and try to compare what happens here
so see here it's interesting because it's saying it's not sure but you can find more information about their website and it has a wrong page it's not that IO it is dot AI and so the link is wrong and so he knows it's hallucinating as well but at least as you can see with GPT3.5 is not making things up saying that Austin Bryan is a CTO of Ask Sage so to be clear it's very much less likely to hallucinate when he knows something so that's why the more we train it the more data we give it the less likely it is that hallucinations ought to happen you're also going to see things where Ask Sage is answering questions and making stuff up like if you ask it to send an email to somebody he's gonna say that it did but it doesn't right it's not able to send emails Ask Sage does not support that today so it's a lie right Ask Sage is the text based chat and we're working on a lot of different plugins so I'm not you know telling you that in the future we may not have an email plugin to email things to people but right now that's not a capability we have so as always trust but verify look at the answers you're getting and try to make sure this actually makes sense all right so we have three types of memory right short-term memory long-term memory and real time. short term is your chat with the bot that's just a current chat and the history of the chat the long-term memory is stuff you can store in datasets into our Vector database on Azure gov impact level 5 and that is you know all the way up to CUI right now and that is good for things that don't change a lot like documents policies and things like that that you know change maybe once a month or every six months or a year something like that and then you have real-time data that's where you want to tap into APIs data lakes data warehouses that's pulling real-time data at the time of the query to get the information you need something like the weather if you ask the question about the weather you don't want to get the weather of yesterday you want to get the weather of today so a good example of that is the plugin we made with the METAR with the FAA to get the weather from an airport so if I say what is the METAR of KIAD in plain English what's interesting is going to go use the FAA METAR API to pull that reference here that you see encoded but yet it's able to translate in plain English because I ask it to do that and so it gave me all the information here in plain English that's an example of a plug-in we built that's pulling from real-time sources to get the information you need and then use the bot capabilities to read that back to the user in the plain English context so we talked about the token consumption like I said you know you're paying for both the question you're asking the data we pass to the model whether it's real-time API like here or it's simply a data set that we ingested in the vector database and you pay for the answer you're getting back keep in mind that if you don't need to pass data sets you can simply set data sets drop down menu here to none and that way by doing that you make sure you make sure that the bot is not using any data sets from our Vector database so you're not paying for tokens you don't need to pay for all right so prompt engineering is going to become the most important skill for most people to get particularly right now and so watch our curated YouTube playlists of prompt engineering videos to know how to ask a question to the bot you know the tone you can use different words like you know concise detailed summarize extract verbs matter how you phrase things matter try to get to your desired outcome in one prompt or two maximum don't chit chat with a bot that's not the most efficient way to get where you want to be and the bot might forget all that past context along the way so and if you can't achieve something most likely blame yourself so far I've yet to be able to find the limit of GPT so it's all about prompt engineering and try to reflect and for that you know we actually created a persona to help you with that and it's called the prompt engineer Persona and I'm going to use an example here we have a we have a prompt here I'm just going to copy and paste this and I'm gonna ask the prompt engineer Persona and I always clear the chat when I change topic so I'm going to say help me improve this prompt to gather medications and illnesses from a veteran medical record right and so this prompt was written to extract medication and dates and illnesses but let's see if the bot itself can help us improve it by itself and what kind of questions it is going to ask us to do that that's always an interesting subject so what's the purpose of the summaries is for research legal purpose can you provide more information about the medical reports all there are they from a specific time period this is to be able to file medical claims we need to collect to collect all medication dates and illnesses and the reports are for the lifetime of the veteran so then it's effectively continuously trying to improve the prompt and it just gave me an updated prompt that I could use to get better outcomes and use that prompt engineer Persona effectively to improve your questions so that you can get the right answer when you're struggling and trying to get things done all right so a lot of people ask us about use cases and there are so many right it's pretty much unlimited but we have people on the government side in acquisition writing RFPs with sage and RFIs we have people on the contractor side using Sage to respond to RFPs and we have people grading bids and categorizing labeling data it's very good at labeling content coding translation reviewing code commenting code using DevSecOps engineering principles to write YAML summarization and sentiment analysis and so on sky is really the limit let me show you how I wrote one of my bid answer by using Ask Sage and this was a bid coming from Tradewind which is a great DOD CDAO program and and so all I did which is pretty incredible I copy and paste the bid I'm gonna show you my prompt here I'm gonna do a new chat and I'm gonna put Contracting officer and I'm going to use a GPT 4 32k model here and the bot is already aware about Sage but it's always good to give it a little bit more context so I've responded as a Sage to this bid from the government and so I'm just giving a little bit of context I'm the CEO of Ask Sage blah blah blah context of what Ask Sage is I wrote a couple of sentences about Ask Sage here based on the RFP and things I thought you know the product and the company will bring to the table and then I wrote the government RFP information I just copy and pasted the PDF no formatting anything it's just the entire PDF here and then I said end of context and I said action make sure to follow all their requirements so that paper will be graded as acceptable by the JFAC and follow their proposed order answering all the required questions blah blah blah and then I say you know following their specific guidance right the two page 10 000 character detailed Discovery paper to show how Ask Sage meets other RFP requirements and he gave me this you know perfect thing I tweaked a little bit a few paragraphs and something that would have taken me probably a day or two took me 37 minutes all right so other field here is a temperature here and that field is important it's enabling you to customize the level of randomness in the generated text so if you want to stick to facts you probably want a temperature down to zero but if you want it to be more imaginative you can increase that to 0.5 maybe all the way to one but you know if you want to stick to facts stay to down to zero so the live query is also essential you know as you probably know the the large English models are trained at specific dates and times and they don't know what happened after that so if I were to ask who is the governor of Arkansas that's going to be wrong because it's going to be stuck in time so it thinks it's Asa Hutchinson and that's wrong but if I clear the chat and don't forget to clear the chat between the queries if I do the the live queries now it's going to pull from being in Google and it's going to give us now the right answer because it has the latest information here as you can see so always uh interesting to to to look at live queries when you know it's time sensitive and you want to get latest information as well all right so now let's look at personalities we have a lot of options here going from an accountant to a Contracting officer to a decoration writer to an electrical engineer a DevSecOps engineer to a legal assistant of course read our terms and conditions you need to always validate legal and medical information with a professional but as you can see we we have a lot of customized personas all the way to a program manager an officer performance report writer so that really helps you customize the tone of the bot and the knowledge of the bot but also sometimes the formatting of the answers so if you want to have a specific format we can create customized personas for you to follow maybe a specific templating that you have to follow in your responses and so that already opens the door to a lot of possibilities by creating custom personas as well which can be created on a per user basis as well alright so don't forget to clear the chats between personas and also between models otherwise the bot might get confused now let's look at data ingestion so first of all we have a lot of parsers PDF HTML ingesting YouTube subtitle from videos structured unstructured data, you name it We have some parsers that are not accessible to you directly so you may have to reach out to us if you have questions about structured and unstructured data ingestion we can tap into APIs we can do a lot of different things we're bringing a lot of different plugins to life for our paid users including PDF word point parsers and plugins will allow you to visualize the content but also train it in in specific datasets and even summarize it if it's too long and so a couple of plugins as example here we have this import chain for text plain content so in this case since the size of the content could be hundreds of pages and we are limited in number of tokens we cut it into chunks then we train the chunks into the data sets and then what we also do is we propose to you to summarize the content into summaries and train the summaries into the datasets as well to have different versions of that content for querying it to find better results so if you were to use a book as an example if you try to adjust the whole book it's not going to work it's too many tokens so we would cut it into chapters then do a summary of the chapters and then do a summary of the summaries to get a summary of the whole book so then we can ingest the whole book and then if you have questions precise about one chapter of the book it will tap into the precise this chapter but if you have broad questions it will tap into the summarized versions so this plugin here helps you do that now another example that's interesting is let's say you ingest a database of resumes right you have 4 000 resumes the way we pass data into the model we only pass maybe four to five results from the vector database based on the query that you you're typing so it's not gonna go through all four thousand uh results so if you're asking you know who is Mr X that's going to work pretty well because it's going to look at the closest results to Mr X and he's going to return that information and that's gonna very much likely to be correct if it's in in the database obviously but if you were to ask you know how many of these users know how to code in Python for example it would only give the top four results so it will not be accurate because it might be more developers that can write in Python so that would be wrong the way we would have to ingest this instead it would be to ingest with categorization by programming languages so we could also have a separate entry for these categories of results but you could also connect to a database or an elastic stack or a postgres database for example and then the bot could convert that query into a SQL query and get the results back as well so many people think somehow that they need a lot of examples to train the bot that's actually not true you only want one or two you only want to train the right examples the right information for the stuff you're trying to achieve giving it the information he needs to give more context more information so don't give a lot just give what you need and you don't want it to get confused don't forget we only pass the top four results so if you give too much information it's not gonna work we also truncate the top four results to 500 tokens per result so you don't want to train results that are more than 500 tokens you want to cut it down to 500 tokens maybe summarize it or maybe cut into multiple pieces that's how we automate that for you inside of the plugins and the summarize plugins and the data ingest plugins now what's also pretty exciting with these plugins is we have this plug-in here with this example which is an admin plugin that you won't be able to see but this is connected to our Sage database and what's pretty amazing is I can ask questions in plain English and get the results in real time and just to show you what that looks like we have a couple of results here that are old but how many users have signed up in the month of April 2023 and it gave me the answer right here so it's converting this question into from plain English to a SQL query and then what's interesting is sometimes the query is wrong and so we get an error message and so we tell the bot to self-reflect and and improve its own SQL query by giving it the error and fixing it and then it fixes it behind the scene you don't you don't see any of this and it's you know giving us the right answer and so this is really a game-changing capability but as you can see here if we ask show me a table with the top five users who consume the most total tokens in April 2023 it knows it needs to give us a table so he's going to give us a table but when I ask it to add first name and email it can automatically add those results so that's a good example of a plug-in that's tapping into a live database to pull the results and format it in the right way to back to the user to be as efficient as possible all right so let's take a look at the data sets here and like we talked about by default it selects all the data sets but you can also put none if you're not using any of the trained data sets we have and you just want the the default large language Model results you can just set it to none so you're not paying for those tokens but then we have ingested a lot of different information from acquisition.gov and and DOD and Air Force and some of my content as well then you can create your own datasets and that really opens the door to your own customized content which really enables you to ingest your own data as well you can decide who you share those datasets with and you have to reach out to us to share it with another user in your team for example so they would be able to see different datasets and you decide how you cut your data and ingest it into which datasets just like labels and so it is very simple to to create a dataset we're going to show you how easy it is now we're gonna do /add-dataset Nic-Video and this is creating this this dataset so it's done and I'm going to do a simple training just to show you how easy it is and keep in mind we have a full API for a paid user and you can reach out to us to get full access to that API but if I if I now want to train uh that Nic Chaillan has a dog and uh the dog's name is Monk and he and and uh uh he is a French Bulldog um now I simply trained this and it took me 21 tokens for that training and if I ask does Nic Chaillan has a dog and yes it says a French bulldog named monk so now he knows and it's as simple as this train command of course you can ingest much larger you know documents and things keep in mind each chunks have to be under 500 token so we cut it for you using the data ingest plugins so you can do that with a plugin you can do that with API or just with the command line here and you see how simple it is now if I want to look at the data we ingested and then I want to find it into my data set I can simply scroll and I see here uh the the training I just did into the dataset and if I want to delete it all I got to do is take the ID and delete it here this item has been successfully deleted so now if I clear my chat and I ask you know does Nic has a dog and he says it's not sure so so like we talked about when you create an account you don't have access to all the paid features but if you pay 30 bucks a month per user you get access to 500 000 tokens for querying of DaVinci model you also get 100 000 token of training but then you also have a 50 dollars plan for one million DaVinci token and a 90 dollar plan for two million tokens per user per month so that's more capacity if you're doing bids and larger documents so reach out to us to upgrade for these accounts all right so let's take a look at some of the account settings that we have here I can see my first name last name company I can set up my phone number but more important I can see how many tokens I have for querying or how many I consume this month how many training tokens I have I can see which data sets I have access to but here is how I can click to activate my MFA option it's gonna use a QR code with Microsoft authenticator to register that but you can also see here the button to register your CAC with your account for your PIV or your CAC to work with your account if you do not register with a CAC in the device if it says CAC not found that means your local proxy is blocking your CAC and you need to create a local it ticket to your support team to ask for a white listing of your CAC pass through for the domain *.asksage.ai to be whitelisted to allow
the CAC passthrough to flow to our website otherwise it's being blocked by your proxy all right so another feature we built is the prompt template and here you're going to find pre- defined templates for prompts for different topics for acquisition and all these different personas and we can add more more templates obviously but here's some example of how to do different things and what's even more exciting is you have your private prompts so you can store your prompts and create new prompts to reuse them so you have to remember how you ask a question so it's as simple as writing a title description in the prompt and picking the Persona you can also share it with others if you want people to benefit from your research you can just click public and anyone can see those those prompts so that helps you really reuse prompts across use cases now another amazing set of features are the plugins we've been building and and we're releasing more and more plugins so by the time you watch this video we're gonna have more plugins but here you see our METAR plug-in we release this amazing git repo plugin which effectively scans a git repo that you give it and it's going to look at every file and see look for improvements both in performance quality and security and create a merge request with a new Branch with all the proposed changes for you to review so it's kind of a free audi can do it pretty much in any programming language and it's a game changer not every change is going to be good but it's giving you a lot of great Insight so try that out we have the commenting code emotions to pull emotion from content evaluating code importing chain which enables you to import text so you copy and paste a PDF or whatever document and it's gonna import it summarize it and train it into a data set of your choice we have this amazing PowerPoint generator which effectively you asked about to write a PowerPoint of X slides maybe five or whatever about X Y and Z and it's going to create the python code to create the the PowerPoint we have a python sandbox to run the code get the PowerPoint back and give it back to the user seamlessly it's kind of a it's kind of mind-boggling we have the sam.gov search which helps you search for bids on sam.gov split text lets you cut text into chunks summarize lets you take a long piece of text could be hundreds of pages and cut it into summary chunks and then summarize the summary of the chunks we have the medical version of it which is kind of the same concept we have the summarize website which is also awesome which lets you put a URL and gets a summary of what's inside of that page and then we have the train content into data set which you know is also used into the import chain here but it's directly ingesting into a data set it's not cutting it or summarizing it for you first so you probably want to use import chain if you have a long piece of text only use the trained content into data set for text that's already short and to the point so the beauty with plugins is also the ability to do agents and complex chaining that's a good example here to take a look at we're going to take a piece of text and I have this article that I wrote on LinkedIn recently about about AI so we're going to use that to try to see how this capability can take a long piece of text then cut it into chunks and give you the ability to ingest this into your data set and so I'm going to copy and paste my article right here I'm gonna leave the default of the summarization and the cut and I'm going to want to train it into my Nic-Video data set and so it pre-populates my prompt here I just have to click send and so what this is going to do first this has um split the text into chunks so I have my different chunks already ready to go chunk one two three four it's asking me do I want to train each chunks into the video data sets I'm gonna say yes and now what's gonna happen is it's it's training and it's ingesting each of these chunks which is on average 340 350 tokens perfectly under the 500 limit now it's asking me hey do I also want to summarize this original content so I'm going to say yes so what's gonna happen here is going to take the the text the long text and summarize it so that I have a summarized version of it and it's gonna give me back that summary another little hint when you use a plug-in make sure that you pick the model that you want to use because it's going to use the model you specify here so use gpt35 if you if you don't need something else or gpt4 but be careful you know only use the right model when you run your plugins so it is going to adapt to the the token limit of these models so here I got my summary so he gave me one two three four pieces of summary of my article into much shorter chunks that are much easier to to summarize so now I can decide if I want to also train the summaries into my datasets and I'm going to say yes and now my data sets have ingested now 180 150 90 215 tokens here so now this summarized version is also in my training so now depending on the questions I'm going to ask it's gonna pull the best results based on this article that I just suggested in a second so as you can see with the chain first he cut it into chunks train the chunks summarize the article train the summarize chunks into the data set all seamlessly all with an Agent plug-in and so we can build these for you to effectively automate a day of work or a whole set of steps and stages and conditions and decision trees to really automate a piece of work another good example that we we have built is the go no go flying decision with the Air Force and it's effectively tapping into 20 APIs from the FAA and Air Force and Gathering all this icing weather data to help the pilot make a flying decision and so that shows you also the the beauty of the decision Tree by aggregating and putting data from different sources all into the single place in Sage to help make the right decisions and so that's also a great example of a plugging agent that could be a game changer all right so we talked about this but we have obviously this multi-tenant stack on chat.asksage.ai you know hundreds of and thousands of people on this but we also have the ability to host a dedicated Enclave for you on your dedicated Azure gov Enclave so reach out to us for these that's good if you if it's scaling up and you have maybe already a lot of users on the multi-tenant side and you want to get a dedicated Enclave maybe for security reasons or whatnot we don't think it's necessary honestly when it comes to security but some people feel better having a multi-tenant stack and we can do that and that way no one else has access to your data other than you inside of your enclave now a lot of people ask us what are we working on when it comes to air gapped classified work well open AI is not bringing their API model anytime soon on the high side but we do have Partnerships with companies like cohere and databricks and others to bring their large large language models to the classified side so we have an engagement with Microsoft and one with Amazon on the secret and top secret fabric to bring large language models and fine-tune them for different use cases but keep in mind those models are not as good as open Ai and so it's pretty tough not very good at coding either so you know very limited but that's the best we can we can do now on the high side and Sage is agnostic to the model so the more Models come the more we're going to be able to bring these options to the table all right I hope this was helpful to you if you want to come chat with us come join us on the Discord Community we have hundreds of people now on the chat always great to share insights and ideas and ask questions I hope this got you excited to come and try Ask Sage and see what we can do to augment your time and your capabilities and if we can help you reach out to us at sales at sage.ai and we'll be happy to answer any of your questions stay safe see you soon
2023-05-18