Autonomous AGENTS by AWS BEDROCK vs GPT-4 Code Interpreter

Key Takeaways

a low Community yes today is about agents and we compare agents in gpt4 with code interpreter to the new AWS Amazon Bedrock agents what is the difference what is their performance so that in the last video I showed you here I have some Central Intelligence AI within gpt4 and I ask here that Central Intelligence to create three different agents agent is here data cleaner yeah yeah improve here to settings data cleaner functionality competence and the tools you can use and then we have here the second agent and the third agent functionality competences and everything you find here the summary in this video but now we're in the middle of August 2023 we have now Amazon Bedrock an Amazon Bedrock you will have llms like the one by E21 Titan BY Amazon clawed by Tropic stable diffusion by stability AI so you will have here with Amazon Bedrock a lot of large language all that you can choose from that is great however what is nice that Amazon provides Amazon Bedrock provides you here the ability just show us here some S3 and you can fine tune your foundation model for particular Downstream task in the easiest way an Amazon makes a separate copy of the base llm of the base Foundation model that is accessible only to you and trains this private copy of the llm so all of this is behind the proven Amazon firewalls so since some user asked make it clear when there is something new so you can fine tune your private llm copy with Amazon Bedrock by the way not sponsored great second we have agents that we can apply for Amazon Bedrock because imagine you want to make an API call to your company systems to your company database to your company inventory to your company document server so we can connect an Amazon provides here agents for you first you need to give the agent access to your corporate data sources and connected with existing apis second the llm needs to figure out what actions to take for a particular job that you want the llm and the agent to do so the llm has to configure the agent in a particular way so this includes building prompts that include here the definitions what to do how to perform the task an instruction for this task you notice so what they do the agents automate The Prompt engineering and the orchestration of the user requester task so the agent will automatically build the prompt and augmented with your company specific information that it got back from the API call to your business database that you have here in all of this within the secure environment of AWS you can break down the task in multiple steps a sequence of API calls you can go to multiple of your databases or whatever information sources you have and the nice thing a fully managed agent you don't have to worry about provisioning it or managing at all the cloud infrastructure this is done because you pay Amazon exactly for this and then I think it performs all the API calls using the proven AWS Lambda functions so for my viewers here the second interesting point the agents automate The Prompt engineering plus their API calls tour to your business database nice thing is we can help you our llms to reason to have some Advanced reasoning we wanted the llm reasons not on what it has been pre-trained or in what we fine-tuned it but we say hey for this particular task I want to show you how you have to tackle these problems step one two three four follow this we know this this is our old friend called react reasoning and acting you can structure prompts to show here our foundation model our llm how to reason here a particular Downstream task and decide on all the action that will help the llm to find the solution structure prompts include a sequence of questions or action observation example 0 point number three we can not only fine tune our llm but we can optimize our llm for a particular task that the agent should do that our llm will help Define and we can use the advanced reasoning with react Ben this is for you react what is react here we have the publication Google and Princeton here react synergizing reasoning and acting in language model and if you want to hear the example on the left side we have a chain of sort argumentation and on the right side we have react so we have here with chain of sort normally a sort an answer question sword answer question sword answer questions or answer you get it with react we have a different sequence we have a question great then we have a sort 1 and action one and observation one okay so two action 2 observation two and sort 3 are action 3 observation three you got it and then here we have a lot of examples this is one example to come to the right conclusion Advanced reasoning with react is it the best and optimal system no but it is great with llm so here we are since Amazon Bedrock agents are not operational and they are currently in an interim by invite only and I decided today during breakfast to do this video and in the last four videos in the last four hours Amazon did not come back to my request that I could access here to this preview version sorry Amazon you missed an opportunity so let's see here what they tell us what it should be able to do once it will be released to the public for example in the agent you have a shopping agent either it is predefined but Amazon or you can Define it whatever you have an agent as I've showed you in my video it is rather simple you prepared you deploy it and the description of the agent is enables human like chat interaction about shopping experience and you say wow and here you have here what Amazon gives us in the manual so system comes up how can I help you say hey I'm looking for some shoes amazing at Amazon you are looking for shoes systems as sure can you tell me what size or man or woman any style preferences so what you normally would have if you're at amazon.com you would have to type in now we have a conversation isn't that beautiful and then the system knows exactly what's in the inventory and says hey I suggest here this for this particular price they're on sale right now they're good for the purpose you are looking for so isn't this great and you say immediately hey I know you have this video here about the integration here of Salesforce their Einstein GPT system how they use it exactly for sales and services end and end yes of course they have the sales GPT the service GPT and and so this is not unique to Amazon this is what is coming to more or less every company it just depends on which platform you as a business owner you want to run this so what are the problems I showed you in my last video I was showing you how to build agents in gpd4 with code interpreter one problem was if I have a huge amount of data more than 5 000 real project with project description technical reports financial data time schedule and and end at a certain time if I wanted to perform some topological dimensionality reduction on the complexity of the project to thematic topics I ran into out of system Ram error so it is not that gpd4 fails it is not that the code interpreter fails it generates the code but the code execution there's only I don't know what limited amount of python environment code execution engine with a certain amount of RAM and if you really have thousands and thousands of projects and you run a complex mathematical operation you will get an out of system RAM and I didn't I was not able to cope with the system Ram because code interpreter is for me a black box with gpd4 I can't optimize the parameter of code interpreter this would be great open AI open AI if you're listening here this is something for you to optimize yes and I'm gonna pay for this too here on AWS Bedrock we have no problem because we have our Lambda functionality with AWS so I just tell them hey here is my second credit card and it says a great we switch on another Cloud machine for you isn't this fascinating but what's really interesting is you can build in addition your own specific agent so here important Point number four we have the agent Action Group here to your business databases or whatever information sources you have behind your firewalls so we can build our own agent isn't this great so overview this is just from the manual since everything is just an internal By Invitation Only preview so yes you know great so let's do this step one provide agent details the agent name is Insurance claim agent okay agent description user input yes I know a lot and I have here my security great and then I can choose here a model from the models that Bedrock provides to me let's say I go with and Tropic Cloud version 1 and Cloud version 2 or whatever I Define here the instructions for the agent UNH and design to help with processing insurance claims and managing pending paperwork up to 800 characters you can insert here with bedrock and you say go here now here this is the interesting part it is optional it is really optional because as I showed you in my video the llm will define an agent for you and gpd4 defined here exactly with code interpreter as a tool agent if you want or as an autonomous agent here exactly what I needed to do what mathematical operation for exploratory data analysis have a look at the video if you want to know more but here now I can Define here what I want I have a description and then you have two things you have your Lambda function that you have to Define that you have to code if you want and you have of course here provide here the open API structure in a Json file I would suppose so here you have now access to your internal databases if you're a business or whatever and we can Define here exactly the action that is going to happen and this is for example something I can not do out of the box with code interpreter with gbt4 maybe I would need some other additional llm that act as an agent for example gorilla but this is part already of another video I showed you so what do we do we say okay we created the agent I will show you in a second how we do this you select the model you have the action group and you have you find your Lambda function this is here and you have defined your Json file with the open API commands and then you say create agent that's it Now isn't this interesting simple example what we have we have here our Lambda function and Lambda Handler we got here specific function get the claims where Insurance Company and here we have now exactly the function and we have here the claim identification number blah blah blah and we have the identification of the policy holder and maybe you have the name and the address and whatever and there's a specific status to this process or to this document or whatever you have and this is it this is the easiest example I could think of if you want to make it a little bit more complicated well we have exactly the same we have your Lambda Handler you get do you define your function get claim documents get required documents check out if there are some missing documents that are not in the required document in a claimed argument compare it tell me what's missing blah blah blah and then simply you define here your functions document ID whatever you get the idea it is rather simple so with this third action group this is this is really interesting and I think this is really outstanding that you can build this in the connectivity to your secret private protected business databases so the Action Group the Json file specifies the action group that you want to add to your AWS agent each Action Group in the file has the properties of one two three four the Lambda function is responsible for processing the API call of course it is associated with this particular action group that you defined here the Lambda function itself must follow the properties ABC the Lambda function for example they get claims action group that has just showed you retrieves the claim from the database or any other data store document store a data Lake Delta Lake whatever you have and returns them in a Json response so you can augment now your query with this particular information for the llm and the associated agent to perform its task this was it I think very interesting what will come up here if Amazon will switch this operational for everybody especially the ability to build here let's call it llm induced or llm augmented autonomous agents laa and if you want to build your own agent to have a connection to your own business data you do not need to go with some long chains but you have everything here behind the closed door of cyber security provided by AWS for you for your data for your business data anything this is an interesting business implication so this is the first video I will make maybe another video where all those have here comparison then with hugging face hacking face agents and hacking face has some really nice feature coming up so I would like to show you how then the agents in gbd4 in AWS maybe in hugging phase Maybe also include some Google let's see what comes up but you see autonomous agent that are controlled and defined by the intelligence of our large language model or our visual language model our visual language action robotics models this is a topic that will I think in the future be of central importance for business application that you can run in any Cloud it does not really depend here on the llm that you choose and Tropic open UI Google or Amazon Titan I don't know what is Amazon Titan if somebody has some Benchmark on the logical reasoning capabilities of Amazon Bedrock Titan please leave a link in the description otherwise I hope it was a little bit informative and it would be great to see you in my next videos