Beyond RAG: A Relational Context Engine That Cuts Token Burn — Unblocked

Are we on air? Good. Can you hear me like this Can you hear me like this ? Is everyone okay? ? Is everyone okay? ? Is everyone okay? Good. We're going to do a Good. We're going to do a Good. We're going to do a quick review because quick review because quick review because we only have 20 minutes, but we only have 20 minutes, but we only have 20 minutes, but I'm Peter from Unblocked. At I'm Peter from Unblocked. At I'm Peter from Unblocked. At Unblocked, we create a Unblocked, we create a Unblocked, we create a contextual engine. contextual engine. contextual engine. So I'll show you So I'll show you So I'll show you a few things. The speech a few things. The speech a few things. The speech will be divided into will be divided into will be divided into two parts. I'll show you two parts. I'll show you two parts. I'll show you what a context what a context what a context engine is and how it can engine is and how it can engine is and how it can be useful for be useful for be useful for your organization, and your organization, and your organization, and then we'll build a then we'll build a then we'll build a component of such an component of such an component of such an engine, and I'll show you engine, and I'll show you engine, and I'll show you some some some open source projects open source projects open source projects that you can that you can that you can work with after work with after work with after that. Here is our plan. that. Here is our plan. that. Here is our plan. What is a context What is a context What is a context engine? How does it create engine? How does it create engine? How does it create value? And then we'll value? And then we'll value? And then we'll start building. start building. Let's go back a Let's go back a few years few years , to the time before the advent of AI , to the time before the advent of AI agents. In those days, agents. In those days, agents. In those days, you were the you were the you were the context level. context level.

Imagine you are a new Imagine you are a new employee in an employee in an employee in an organization and you organization and you organization and you are trying are trying are trying to understand what is to understand what is to understand what is happening. happening. Usually you Usually you have to have to have to look through a bunch of look through a bunch of look through a bunch of code, flip through code, flip through code, flip through documentation, just documentation, just documentation, just trying to find trying to find trying to find the information the information the information you need to get the job done. you need to get the job done. Not to mention all the Not to mention all the history that history that history that has accumulated over has accumulated over has accumulated over the years in the organization, the the years in the organization, the the years in the organization, the incidents and the incidents and the incidents and the experience gained. The experience gained. The context engine context engine brings all this brings all this brings all this corporate knowledge corporate knowledge corporate knowledge together. You get a together. You get a together. You get a disjointed context disjointed context that goes into that goes into that goes into the engine, and at the output you the engine, and at the output you the engine, and at the output you have an ordered have an ordered have an ordered context where all the data context where all the data context where all the data is interconnected. He is interconnected. He is interconnected. He takes intentions into account. It is takes intentions into account. It is takes intentions into account. It is personalized. And personalized. And personalized. And it takes into account it takes into account it takes into account access rights. Now I'll access rights. Now I'll access rights. Now I'll show you exactly what I mean show you exactly what I mean show you exactly what I mean . Let's imagine . Let's imagine I switch to Claude.

I switch to Claude. I'll show you some I'll show you some I'll show you some interesting sessions I've had interesting sessions I've had interesting sessions I've had . I asked . I asked . I asked Claude, without a Claude, without a Claude, without a context engine connected, context engine connected, context engine connected, to create a plan to create a plan to create a plan to optimize a to optimize a to optimize a component in our component in our component in our layer called Source layer called Source layer called Source Mark engine. I'm sure you Mark engine. I'm sure you Mark engine. I'm sure you know what Claude does know what Claude does know what Claude does when he starts a new when he starts a new when he starts a new task. He's like a task. He's like a task. He's like a new employee. He new employee. He new employee. He knows nothing about knows nothing about knows nothing about your code or your code or your code or organization. So the organization. So the organization. So the first thing it does is iterate through first thing it does is iterate through your your your codebase, codebase, codebase, trying trying trying to figure out what goes where. to figure out what goes where. to figure out what goes where. So, you can So, you can So, you can see it here. And see it here. And see it here. And you know, in the end he you know, in the end he you know, in the end he figures it out, or figures it out, or figures it out, or at least confidently at least confidently at least confidently demonstrates that he demonstrates that he demonstrates that he has figured it out. So he has figured it out. So he has figured it out. So he gives me something like a gives me something like a gives me something like a recommendation. recommendation. But now I'll show But now I'll show you what it looks like you what it looks like you what it looks like when you connect the when you connect the when you connect the context engine to the context engine to the context engine to the backend. Now it backend. Now it backend. Now it calls the calls the calls the Unblock context engine, and Unblock Unblock context engine, and Unblock Unblock context engine, and Unblock provides a very complete, provides a very complete, provides a very complete, but but but task-specific, set of task-specific, set of task-specific, set of context. And you context. And you context. And you can see that can see that can see that it references not it references not it references not only the source code, only the source code, only the source code, but also the Notion but also the Notion but also the Notion architecture docs, architecture docs, architecture docs, pull pull pull requests, and requests, and requests, and Slack discussions about Slack discussions about Slack discussions about optimizing this optimizing this optimizing this particular particular particular component. As component. As component. As a result, you a result, you a result, you get a very get a very get a very coherent set of coherent set of coherent set of context with all context with all context with all these links. So these links. So these links. So now, if, you know, now, if, you know, now, if, you know, Claude wants to Claude wants to Claude wants to come back and come back and come back and get more get more get more information. It has information. It has information. It has all these great points

all these great points all these great points to jump to. But to jump to. But to jump to. But the most interesting thing is this. the most interesting thing is this. Here I spent, Here I spent, let's say, $1.29, and let's say, $1.29, and let's say, $1.29, and it took about a it took about a it took about a minute and thirty minute and thirty minute and thirty seconds. This was with the seconds. This was with the seconds. This was with the contextual engine, and contextual engine, and contextual engine, and without it I spend without it I spend without it I spend up to $2.60 and three up to $2.60 and three up to $2.60 and three minutes. So you're minutes. So you're minutes. So you're seeing a seeing a seeing a 50% reduction in costs, which is 50% reduction in costs, which is 50% reduction in costs, which is just incredible. just incredible. just incredible. Now I want to show Now I want to show Now I want to show you another interesting you another interesting you another interesting demonstration. This is an demonstration. This is an demonstration. This is an interesting session that my interesting session that my interesting session that my colleague...oh, sorry. colleague...oh, sorry. colleague...oh, sorry. I'll bring it here. This is an I'll bring it here. This is an I'll bring it here. This is an interesting session my interesting session my interesting session my colleague did with Unblocked colleague did with Unblocked colleague did with Unblocked last week. We last week. We last week. We also have a also have a also have a code review product. And code review product. And code review product. And Richie is working on it. Richie is working on it. Richie is working on it. He found that He found that He found that the number of issues the the number of issues the code review product was detecting code review product was detecting began to decrease. began to decrease. began to decrease. So he just So he just So he just asked Unblocked, and Unblocked asked Unblocked, and Unblocked asked Unblocked, and Unblocked was able to get that was able to get that was able to get that information. And quickly information. And quickly information. And quickly narrow down the search narrow down the search narrow down the search to the likely cause, to the likely cause, to the likely cause, namely the transition from Opus namely the transition from Opus namely the transition from Opus 4.6 to Opus 4.8. And 4.6 to Opus 4.8. And 4.6 to Opus 4.8. And the characteristics the characteristics the characteristics of behavior behind of behavior behind of behavior behind the scenes the scenes the scenes caused a decrease in caused a decrease in caused a decrease in the number of these the number of these the number of these problems. So he problems. So he problems. So he just asked Unblocked just asked Unblocked just asked Unblocked to fix it. And now to fix it. And now to fix it. And now Unblocked has been able to create PR Unblocked has been able to create PR Unblocked has been able to create PR behind the scenes.

behind the scenes. behind the scenes. He has all this He has all this He has all this context and he knows context and he knows context and he knows what to do. But what to do. But what to do. But the coolest thing is that you the coolest thing is that you the coolest thing is that you can see how can see how can see how he created this description, he created this description, he created this description, and it's just amazing. and it's just amazing. and it's just amazing. Like, he uh, understood the Like, he uh, understood the Like, he uh, understood the pull request (PR) pull request (PR) pull request (PR) that caused that caused that caused the regression from the beginning. the regression from the beginning. And then I found a And then I found a Slack discussion where Slack discussion where Slack discussion where it was discussed and it was discussed and it was discussed and used some of used some of used some of the data from there to the data from there to the data from there to help with the help with the help with the fix, which is fix, which is fix, which is just crazy. just crazy. So, I'll go back to... oops, I think I oops, I think I oops, I think I lost my lost my lost my Keynote presentation. Keynote presentation. Keynote presentation. Now let's see. Here Now let's see. Here Now let's see. Here she is. So you have the same she is. So you have the same she is. So you have the same problem with problem with problem with agents, right? Oh, agents, right? Oh, agents, right? Oh, sorry, this is the wrong sorry, this is the wrong sorry, this is the wrong presentation. This is not the presentation. This is not the presentation. This is not the same Keynote. Now that's same Keynote. Now that's same Keynote. Now that's good. So now, I good. So now, I good. So now, I think we'll just think we'll just think we'll just jump right into jump right into jump right into the creation, because we're the creation, because we're the creation, because we're very short on very short on very short on time today. So I'll just time today. So I'll just time today. So I'll just summarize it. You have summarize it. You have summarize it. You have context scattered context scattered context scattered everywhere. You have everywhere. You have everywhere. You have context in PR, context in PR, context in PR, reviews, comments, reviews, comments, reviews, comments, issues and tickets, issues and tickets, issues and tickets, documentation, documentation, documentation, conversations. And all of this is conversations. And all of this is conversations. And all of this is semantically semantically semantically queried, but queried, but queried, but at the same time at the same time at the same time structured. And RAG does a structured. And RAG does a structured. And RAG does a great job great job great job with the content part.

with the content part. It can handle content issues, find content issues, find the code, the PR that the code, the PR that the code, the PR that changed it, the Slack thread where changed it, the Slack thread where changed it, the Slack thread where it is discussed. And it is discussed. And it is discussed. And if you wrap that if you wrap that if you wrap that in an agency cycle, in an agency cycle, in an agency cycle, it can go very it can go very it can go very far. If you far. If you far. If you give the agent give the agent give the agent loop the tools loop the tools loop the tools to search for files to search for files to search for files afterwards, it will be afterwards, it will be afterwards, it will be very powerful. But it very powerful. But it very powerful. But it doesn't handle other types of doesn't handle other types of doesn't handle other types of structural questions structural questions , such as PRs, which I , such as PRs, which I , such as PRs, which I merged last merged last merged last week. Who week. Who week. Who checked the payments the most? checked the payments the most? checked the payments the most? Open PRs, etc. And these are Open PRs, etc. And these are Open PRs, etc. And these are requests. This is not requests. This is not requests. This is not semantic search. semantic search. So I wanted to So I wanted to show you again, show you again, show you again, going back to going back to going back to Richie's example, that he Richie's example, that he Richie's example, that he started the conversation with started the conversation with started the conversation with Unblock by asking Unblock by asking Unblock by asking a question based on an a question based on an a question based on an inquiry, because there inquiry, because there inquiry, because there 's a time element here. Can 's a time element here. Can 's a time element here. Can you show you show you show me a graph of me a graph of me a graph of the odds each week the odds each week the odds each week since January 1st? Okay, to since January 1st? Okay, to since January 1st? Okay, to do this, it can't do this, it can't do this, it can't just just just pull this pull this pull this information from the information from the information from the vector storage vector storage vector storage because that because that because that information isn't information isn't information isn't encoded there. So he will encoded there. So he will encoded there. So he will have to do have to do have to do it through some kind of it through some kind of it through some kind of query mechanism. And query mechanism. And query mechanism. And you know, search essentially has you know, search essentially has you know, search essentially has two halves.

two halves. two halves. We have semantic We have semantic We have semantic search, and it can search, and it can search, and it can handle these handle these handle these types of semantic types of semantic types of semantic questions. But today questions. But today questions. But today we will work on we will work on we will work on that. Structural that. Structural that. Structural part. We part. We part. We will convert will convert will convert structured queries structured queries structured queries or natural or natural or natural language queries into structured language queries into structured language queries into structured queries. But creating queries. But creating queries. But creating queries is queries is queries is actually the easy actually the easy actually the easy part. The hard part. The hard part. The hard part is the framework part is the framework part is the framework that is that is that is built around it. So you built around it. So you built around it. So you need to do need to do need to do things like understand things like understand things like understand the schema so that you can the schema so that you can the schema so that you can pass it to the LLM and it will pass it to the LLM and it will pass it to the LLM and it will generate a well- generate a well- generate a well- constrained query. You constrained query. You constrained query. You also need to be able to also need to be able to also need to be able to match match match identifiers. If identifiers. If identifiers. If I ask a question, I ask a question, I ask a question, for example, "open for example, "open for example, "open PR from Rashin." He won't PR from Rashin." He won't PR from Rashin." He won't know who know who know who Rashin is, because his Rashin is, because his Rashin is, because his GitHub ID GitHub ID GitHub ID might be completely might be completely might be completely different. So, we have to different. So, we have to match it up somehow. And when you match it up somehow. And when you get a request from get a request from get a request from LLM, it is not reliable. LLM, it is not reliable. LLM, it is not reliable. It may contain It may contain It may contain dangerous operations. dangerous operations. dangerous operations. He may violate He may violate He may violate the tenant's boundaries. Therefore, the tenant's boundaries. Therefore, the tenant's boundaries. Therefore, you must have a you must have a you must have a way to verify that way to verify that way to verify that the request is valid and the request is valid and the request is valid and secure. And finally, secure. And finally, secure. And finally, even if you even if you even if you do such a do such a do such a check, check, check, errors may occur. And errors may occur. And errors may occur. And when you actually when you actually when you actually execute the query against the execute the query against the execute the query against the database, there database, there database, there can be errors there too.

can be errors there too. can be errors there too. So we can force So we can force So we can force LLM to try to LLM to try to LLM to try to fix itself if fix itself if fix itself if this this this situation arises. That's what situation arises. That's what situation arises. That's what we're aiming for here. We we're aiming for here. We we're aiming for here. We want to move on to want to move on to want to move on to this. We want to be this. We want to be this. We want to be able to say able to say able to say something like " something like " open PR from Rashin open PR from Rashin ." And then we want to ." And then we want to ." And then we want to receive such a request receive such a request receive such a request in response. Let's in response. Let's in response. Let's look at this for a look at this for a look at this for a minute. This is an minute. This is an open source repository on open source repository on which we which we which we will build the will build the will build the system. And I will immediately system. And I will immediately system. And I will immediately show you what this show you what this show you what this looks like in practice. looks like in practice. So, here is this So, here is this repository. It repository. It repository. It is called the is called the document query mechanism. I want document query mechanism. I want to show you what it to show you what it to show you what it looks like when you first looks like when you first looks like when you first launch it. So, I can launch it. So, I can launch it. So, I can say, for example, "PR say, for example, "PR say, for example, "PR from Peter." Oh, from Peter." Oh, from Peter." Oh, actually, we're actually, we're actually, we're not launching it right now. not launching it right now. not launching it right now. So I will go back So I will go back . Let's . Let's . Let's run it first. Now run it first. Now run it first. Now I can ask my I can ask my I can ask my question. So when question. So when question. So when you start, it's you start, it's you start, it's not set up at all.

not set up at all. not set up at all. He won't give anything away He won't give anything away . So we're going to . So we're going to . So we're going to gradually build gradually build gradually build this system, okay? this system, okay? The first thing we'll The first thing we'll talk about is talk about is talk about is automatic automatic automatic schema detection. This is a schema detection. This is a schema detection. This is a completely dynamic completely dynamic completely dynamic process. [ __ ] does not have a process. [ __ ] does not have a process. [ __ ] does not have a fixed schema. Not fixed schema. Not fixed schema. Not all databases are like this, all databases are like this, all databases are like this, but the same principle but the same principle but the same principle can be applied to can be applied to can be applied to many. You can many. You can many. You can dynamically discover dynamically discover dynamically discover the schema at the schema at the schema at runtime. The technique runtime. The technique runtime. The technique consists of sampling consists of sampling consists of sampling many documents many documents many documents in a repository and then in a repository and then in a repository and then inferring a schema inferring a schema inferring a schema based on that. And I want to based on that. And I want to based on that. And I want to show you one show you one show you one tricky part of tricky part of tricky part of this process— this process— defining defining defining enums. The result enums. The result enums. The result is that you is that you is that you can can can use the LLM as a use the LLM as a use the LLM as a powerful tool, powerful tool, powerful tool, and it does and it does the job admirably. But the job admirably. But in reality, you can in reality, you can in reality, you can go very, very go very, very go very, very far just far just far just by using by using by using traditional traditional traditional procedural methods. procedural methods. procedural methods. So you can, you know, So you can, you know, So you can, you know, not always go not always go not always go for an LLM to do the for an LLM to do the for an LLM to do the kind of work that kind of work that kind of work that seems seems seems non-deterministic.

non-deterministic. non-deterministic. It's actually pretty It's actually pretty It's actually pretty deterministic, deterministic, deterministic, that kind of thing. that kind of thing. So, now that we've So, now that we've discovered the schema, I'll discovered the schema, I'll discovered the schema, I'll just show you what just show you what just show you what it looks like in our it looks like in our it looks like in our demo. Oh. demo. Oh. demo. Oh. First we First we First we need to move to the need to move to the need to move to the next level. And next level. And next level. And then launch. So then launch. So , now we have discovered , now we have discovered , now we have discovered the schema from our the schema from our the schema from our GitHub collection. And these are GitHub collection. And these are GitHub collection. And these are just just just pull requests. OK? There pull requests. OK? There pull requests. OK? There is not only a schema here, but is not only a schema here, but is not only a schema here, but also indexes that also indexes that also indexes that will help improve will help improve query performance. The next query performance. The next step is step is identity recognition. This is what identity recognition. This is what I was talking about, about I was talking about, about I was talking about, about matching matching matching users to users to users to real real real IDs. And IDs. And IDs. And again, this can be again, this can be again, this can be solved solved solved with simple with simple with simple procedural tricks procedural tricks . So that's actually more or . So that's actually more or less how we less how we less how we do it "under the hood" do it "under the hood" do it "under the hood" at Unblocks. When we take at Unblocks. When we take at Unblocks. When we take usernames usernames usernames or IDs from or IDs from or IDs from different systems and different systems and different systems and compare them, we can compare them, we can compare them, we can use use use fuzzy fuzzy fuzzy matching. And then, matching. And then, matching. And then, eventually, if eventually, if eventually, if necessary, you can necessary, you can necessary, you can submit the list of names submit the list of names submit the list of names to the LLM to to the LLM to to the LLM to act as an arbitrator. And act as an arbitrator. And act as an arbitrator. And finally, the last finally, the last finally, the last step before we step before we step before we can show the can show the can show the interesting stuff—this is it. This is the interesting stuff—this is it. This is the interesting stuff—this is it. This is the synthesis stage. And, roughly synthesis stage. And, roughly synthesis stage. And, roughly speaking, this is actually the

speaking, this is actually the speaking, this is actually the easiest part. easiest part. easiest part. We simply take We simply take We simply take the schema that we the schema that we the schema that we dynamically defined dynamically defined dynamically defined earlier. We take the earlier. We take the earlier. We take the current date and time. current date and time. current date and time. Maybe a little Maybe a little Maybe a little information about information about information about the user the user the user asking the question asking the question asking the question so the system can so the system can so the system can easily distinguish between easily distinguish between easily distinguish between queries like "my queries like "my queries like "my PR" and not just " PR" and not just " PR" and not just " Peter's PR." And then we Peter's PR." And then we Peter's PR." And then we pass this to LLM and pass this to LLM and pass this to LLM and ask her ask her ask her to synthesize the query to synthesize the query to synthesize the query for us. The result for us. The result for us. The result will look will look will look something like this. something like this. Let's move on to this. So now I can So now I can go back to my go back to my go back to my engine and try to engine and try to engine and try to run this query run this query run this query again. And now he will again. And now he will again. And now he will pull up the documents pull up the documents pull up the documents for me and generate for me and generate for me and generate this request. Now, the this request. Now, the this request. Now, the next step next step next step I was talking about is I was talking about is I was talking about is validation. In fact, validation. In fact, validation. In fact, this is the most important thing. this is the most important thing. Your query may Your query may contain really contain really contain really dangerous operators dangerous operators , such as a , such as a , such as a function operator that executes function operator that executes function operator that executes pure JavaScript on the pure JavaScript on the pure JavaScript on the database server, database server, database server, which is definitely not desirable.

which is definitely not desirable. But you may But you may want to do want to do want to do other things with it. other things with it. other things with it. For example, you For example, you For example, you may have may have may have hidden fields that hidden fields that hidden fields that you do not want you do not want you do not want LLM to show or for LLM to show or for LLM to show or for LLM to retrieve LLM to retrieve LLM to retrieve and execute. If and execute. If and execute. If you have metadata fields, you have metadata fields, you have metadata fields, for example, containing for example, containing tenant IDs, and you tenant IDs, and you want to isolate want to isolate want to isolate each request to a each request to a each request to a specific tenant specific tenant , it is better not , it is better not , it is better not to allow LLM to allow LLM to allow LLM to interact with these to interact with these to interact with these fields at all, but to restrict their fields at all, but to restrict their fields at all, but to restrict their operation to the operation to the operation to the system level. You can system level. You can system level. You can insert these fields insert these fields insert these fields hidden. This is what it hidden. This is what it hidden. This is what it roughly looks like. roughly looks like. So, again, we'll So, again, we'll quickly show quickly show quickly show the result and what it the result and what it the result and what it looks like. Let's move on looks like. Let's move on looks like. Let's move on to our to our to our verification phase. Oh. Let's verification phase. Oh. Let's run it again. So run it again. So if I now ask a if I now ask a if I now ask a question like question like question like this: PR with this: PR with this: PR with authentication in the authentication in the authentication in the header. This will header. This will header. This will probably trigger probably trigger probably trigger some kind of regular some kind of regular some kind of regular expression. And here it is.

expression. And here it is. expression. And here it is. OK. So our OK. So our OK. So our validation engine validation engine validation engine said, “Sorry, said, “Sorry, said, “Sorry, regular expressions regular expressions regular expressions are not allowed. "You are not allowed. "You are not allowed. "You have been rejected." And now have been rejected." And now have been rejected." And now we are not we are not we are not allowed to do that. So, allowed to do that. So, allowed to do that. So, validation is cool. validation is cool. But what should we do about it But what should we do about it ? We received ? We received ? We received an error. But we an error. But we an error. But we can just take can just take can just take this error, feed this error, feed this error, feed it back into the LLM, and it back into the LLM, and it back into the LLM, and let the model let the model let the model correct correct correct itself. That's all, actually itself. That's all, actually itself. That's all, actually . From a technical . From a technical . From a technical point of view, this is point of view, this is point of view, this is exactly what it looks like. It's exactly what it looks like. It's exactly what it looks like. It's just a loop with a certain just a loop with a certain just a loop with a certain maximum maximum maximum number of attempts. And number of attempts. And number of attempts. And so if I try so if I try so if I try again, I'll increase again, I'll increase again, I'll increase the number of attempts and the number of attempts and the number of attempts and run this. Now, if run this. Now, if run this. Now, if I try this query I try this query I try this query again, oops, it's not the same one. again, oops, it's not the same one. Let's try this one. Now Let's try this one. Now you see that it you see that it you see that it took a few took a few took a few tries. But in the end tries. But in the end we achieved we achieved we achieved the result. However, we have the result. However, we have the result. However, we have limited it to limited it to limited it to using match.

using match. using match. Unfortunately, this is not an exact Unfortunately, this is not an exact Unfortunately, this is not an exact match of the title. We match of the title. We match of the title. We don't have a PR with that don't have a PR with that don't have a PR with that exact name. So exact name. So exact name. So we need something else, we need something else, namely full-text search full-text search . So, if we add . So, if we add . So, if we add full-text search full-text search full-text search to our engine, to our engine, to our engine, hopefully we'll hopefully we'll hopefully we'll get a result get a result get a result similar to this. similar to this. Let's try again. Perfectly. That's all. We Perfectly. That's all. We went all the way from went all the way from went all the way from start to finish. This is an start to finish. This is an start to finish. This is an open open open source project, so you can source project, so you can source project, so you can see how it's see how it's see how it's implemented and implemented and implemented and try it try it try it yourself. This is what we yourself. This is what we yourself. This is what we created. Something that created. Something that created. Something that effectively effectively effectively handles the other handles the other handles the other half of RAG. Everything half of RAG. Everything half of RAG. Everything related to related to related to structured data structured data . You can . You can . You can use use use filters, aggregations, and filters, aggregations, and filters, aggregations, and time constraints. And time constraints. And time constraints. And it is it is it is self-correcting.

self-correcting. self-correcting. But the coolest thing is that But the coolest thing is that But the coolest thing is that it also powers a it also powers a it also powers a chat agent. chat agent. So if I go So if I go here and say something here and say something here and say something like, "Um, what like, "Um, what authentication-related issues did we authentication-related issues did we work on last work on last work on last week?" Now you week?" Now you week?" Now you see everything that see everything that see everything that shows what exactly it shows what exactly it shows what exactly it does, all the queries does, all the queries does, all the queries it creates. And now it creates. And now it creates. And now it has the data, it just it has the data, it just it has the data, it just processes it in the processes it in the processes it in the background, and that's it background, and that's it background, and that's it . Very . Very . Very cool. So there you have it—a cool. So there you have it—a structured structured structured context available context available context available for querying by both for querying by both for querying by both agents and agents and agents and humans. There's humans. There's humans. There's just one last thing left just one last thing left . Um, in the description you . Um, in the description you . Um, in the description you probably saw that we probably saw that we probably saw that we promised a promised a promised a context engine simulator. context engine simulator. If you don't want If you don't want to connect Unblocked to connect Unblocked to connect Unblocked right away due to certain right away due to certain right away due to certain limitations in your limitations in your limitations in your environment, we have an environment, we have an environment, we have an open open open source project that will allow source project that will allow source project that will allow you to do it you to do it you to do it completely locally.

completely locally. The idea is that it The idea is that it creates a context creates a context creates a context package for each package for each package for each individual task and individual task and individual task and then uses then uses then uses that package to that package to that package to conduct A/B conduct A/B conduct A/B testing of the testing of the testing of the same task, same task, same task, demonstrating demonstrating demonstrating the benefits of having the benefits of having the benefits of having such context at such context at such context at hand for your hand for your hand for your agents. Here's the QR code for agents. Here's the QR code for agents. Here's the QR code for it if you it if you it if you need it. need it. OK. And lastly: OK. And lastly: if you want coconuts, if you want coconuts, if you want coconuts, come to our come to our come to our stand. Thank you.