Let's now focus on unstructured data. You'll design two specialist agents to decide on the graph model that can be extracted from markdown files. Let's get to it. You have built a complete workflow for defining a construction plan that can build a knowledge graph from CSV files. Now you can move on to the unstructured data workflow. This starts in a similar way with user intent and file suggestion agents, but now for markdown files. Those agents are needed for the complete solution, but we'll simulate their output in this lesson. Instead, you will focus on the new concept, the Entity and Fact Type Proposal Agent. The Entity and Fact Type Proposal Agent is itself composed of two dedicated sub agents, the named entity recognition, or NER schema agent, and the fact type extraction agent. Notice that the output of these agents is a plan for how to perform knowledge extraction, not to do the extraction itself. The NER schema agent will read the markdown files looking for named entities. These are people, places, or things that are prominent in the text. In our example, you'd expect things related to furniture and reviews. The fact type extraction agent is a second pass over the text that looks for the kind of statements that appear about those things. So for example, some furniture has issues. Together, those agents will come up with a plan for extracting information that supports the approved user goal. To begin of course, we're going to have to go through some setup and import the usual libraries. And then we'll get OpenAI set up and check that it's working perfectly. That looks good. And we'll also check that Neo4J is available. It's ready too. Good. Now, the first agent that you're going to define is a Named Entity Recognition agent. Named Entity Recognition is a very common natural language processing operation that you can find in many different frameworks. It turns out, of course, that LLMs are very good with language, so they're very good with this task as well. Now, Named Entity Recognition, as it says, is the idea of like having entities that are recognized and giving them names really. That's that's what's kind of happening. An entity can be anything from a person, a place or a thing. And here we're going to ask the LLM to look for things that are relevant to the user's goal that it can find within the text that's available. You'll start by defining the instructions for the agent and as before, we're going to break this up into a couple different parts that we'll then compose together. First, define what the agent's role and goal are. Here, you're going to describe the agent as being a top-tier algorithm designed for doing natural language processing really. And its goal is to find named entities within the text, but it's not actually going to extract those entities. The goal here is for it to identify what kinds of entities are available. You can then give some more hints to the agent to let it know exactly what we're looking for here. And again, we're going to describe what an entity is, and then also we're going to break it down to two different categories of entities. We're going to describe what we're calling here well-known entities, and these are well-known because they exist in the structured data that we've already described in the previous lesson. So if that structured data, if some of those entities defined there occur in the text, we want those parts of the text to also be extracted. In addition to that, we have what we're calling discovered entities. And these are entities that may not exist in the graph data that we've described, but they might be appropriate to the user's goal. And if they appear consistently, then maybe there's something useful. We then give some more details about, you know, what the design rules are for identifying what a well-known entity is, and also what a discovered entity is. And as before, usually give a couple of examples to help the agent really understand what the purpose is. The final part of the instructions that you want to put together are the chain of thought directions and here we're going to describe exactly how to prepare for doing the task at hand, which is identifying these entities. So here's the tools that it needs to use to actually find out what is expected of it, give it the full context. And then here is the proposed series of steps that it should go through. It knows some files that are available. It has a sampling file tool. So go ahead and use that to take a look at some of the files. and see what it can discover and consider both the well-known entities and then also discovering new ones that are frequently mentioned. And then through that, put together an entire list of entities that it thinks are appropriate. And then use the get_proposed_entities tool to find that list. And as before, we'll turn back to the user to say, hey, does that look correct? And then there'll be a separate tool that is called for actually making the approval. Here it's called the approve_proposed_entities tool. You can then compose all those together into a single string and this will become our agent instructions for the named entity recognition agent. With the instructions defined, you can then provide the definitions of the tools themselves, and these are going to be pretty straightforward. There's not a lot of cleverness happening here. The tools we're going to define follow the same pattern that we've had. in previous lessons, we're going to ask the agent to first propose some things and then have some other tools that will take those proposals and turn them into approved versions of those. So here, the things we're proposing are a list of entities that are first proposed and then approved. And then you can also of course get those either proposals or the approvals. The next tool you need is a little bit different. It's going to be another getter tool. So this is going to be getting the well-known types from the schema that was proposed in the previous lesson from the structured data. So the well-known types will end up being the labels that were defined for nodes in the previous proposed schema. Here they're going to be used as the approved labels. We're going to pull them out of the construction plan and then just return that as a list. You'll go ahead and import the predefined tools that we've defined in previous lessons, and also the additional ones here. As before, you can combine them into a single list, and that's what we'll use as part of the agent definition. Now, to get an idea of what the agent is working with, you can use that sample_file function, just go ahead and call it directly on one of the available files and take a look at the content. So looking at this markdown file, you can see that this is a bunch of reviews in markdown format. It's got the headings, it's got some values that are embedded here like ratings. You can see there's usernames, locations, and of course the text of the review itself. So there's only a handful of reviews in here, but it's enough to kind of demonstrate how this all works. Okay, you've got your instructions defined, you've got your tools defined. Now we can construct the agent itself. Now to run this agent, because it's part of a longer workflow, it has some assumptions about what's really been accumulated in state. So we've got to create an initial state to try this agent out. And the initial state requires a couple of things. It needs to have the approved user goal, the approved files. Here you can see the markdown files it's going to take a look at. And then also the construction plan. And it's the construction plan that's used for those well-known entity types. Notice we're leaving out the relationship construction just because we don't need that in this step. So this is just kind of simulating what you would need. Now you're ready to go ahead and run the agent. We're going to use that make_agent_caller from the helper module and we're going to send in a simple request here. We're going to just basically tell it, you know, hey agent, can you do your job? Add some product reviews to the knowledge graph to take product complaints back through the manufacturing process. If you run into any kind of trouble here, of course, you can run the alternative version of this, which is a very short sentence, but it has the debug, comma True here that'll show you all the output. Once the agent is done kind of getting an idea of what it wants to propose and it makes a proposal, we'll take a look at the session state to make sure that it actually has made a proposal and hopefully it should not have automatically approved. It should be waiting for us to say, yeah, that looks good. So this will take a couple minutes to run because the agent is going to go off, take a look at a couple of these different files and try to come up with what it thinks is the best set of labels to use. for the entities. Okay, pretty good job. It's identified that there's products here, there's issues that are reported, complaints about customer experience, assembly instructions and maybe some customer feedback. Maybe a bit of redundancy, but this this looks good enough to me. and this is also important. Not only did it have a good response, but it also updated the session state, the memory correctly. So we've got proposed entities, but we do not yet have approval. So I've gotten a good result here. If you have not, feel free to run that cell again and see if you get a better set of proposed entities. Again, this is all set up assuming that we're going to be in an interactive environment. But if it works, you can go ahead and send a new message to the same agent saying I approve those proposed entities. Once this is done, we should find that it has actually gone and transferred the proposed entities into approved entities, and we should see that in the session state. Perfect. There we go. The approved entities matching what they had proposed before. You can then move on to the second agent. and this is going to be our Fact Type Extraction Sub-agent. And as the name implies, of course, and similar to what the previous agent did, we're going to be looking for the types of things that can be extracted. We're not going to do the extraction itself. So you can start with the instructions, because that's the most important part of both of these agents. And for this agent, the role and goal, again, we're going to say that this agent is a top-tier algorithm. It's going to be doing some text analysis. But here the goal is a little bit different. It's going to be looking for the type of facts that could be extracted. Don't actually extract those facts, just find what types of facts are possible. To help the agent out to understand what we're really talking about here. We're going to provide some hints, and the important part here that we're going to say and it's similar to what we had in the previous instructions for the the previous agents. Do not propose specific individual facts, but instead propose the general type of fact that would be relevant for the user's goal. and giving examples is always a great idea. Here my example is do not propose that ABK likes coffee, but the general type of fact that, you know, person likes beverage. So these form very simple sentences and we call these sentences triplets and they're of the form subject, predicate, object. And you can notice that here we've put it in parentheses, this is a classic thing to do. We'll see if the agent actually picks up on that format and provides that back to us in the response. We give it some additional design rules for how to think about this and what to look for, and then also how to use the tools. A little bit different in the previous agent. The previous agent had an entire list of proposed entity types. Here for each individual fact, we're going to just add that as an individual fact one at a time rather than here's a collection of facts that we're going to propose. subtle difference here. The importance really ends up being like both how the agent behaves and also what the cost of tokens really ends up being. The more round trips you have, it's going to be a bit more expensive to do, but if it's going to have better results, then maybe that's a good tradeoff. And finally, you can go ahead and add the chain of thought directions. as before, we're going to follow the same pattern that we always do for these directions. Here's how you prepare for the task and then here's our suggested step-by-step approach for how to actually execute on this task. use the tools in this way, sample some of the files, look for subjects and objects that are related to the text, and then add proposed facts about those subjects and objects. So what you end up with here should be a couple of small sentences, three-word sentences. You can then compose that all together in a single string and we're ready to move on to the tool definitions. Now, I'm going to spend a little bit more time looking at the tool definitions for the fact type extraction agent and that's because it does a little bit of sanity checking along the way. And that's part of the intention of actually separating these out into different agents. In the previous agent, the named entity recognition agent, We give it some guidance about how to look for entities, but we didn't really constrain it too much about what it found. Here we're going to be very particular that the facts that it's actually going to propose must match up with existing entity types that were defined by the prior agent. Splitting it up in this way gives us a little bit better guarantee that we're not going to have facts that actually don't line up with entity types that we want. So, when you look at the definition of add_proposed_fact here, we're getting what we're even saying by argument name is an existing approved subject label, proposed predicate label, and then an approved object label. So that again is the sort of the triple. And these triples though, both the subject and the object should already exist from the previous agent. So we're going to put a little bit of guardrails around that. We're going to check, here's the approved entities, we're going to get that from state and make sure that both the approved subject and the approved object label that are passed in as arguments are in that list. And if not, we're going to then inject an error from this tool call and tell it that label doesn't exist, you know, that it should try again. part of the reason why we're doing things one piece at a time here so that we can one piece at a time actually make a little bit of corrections to the agent's behavior. If that all looks fine, then we kind of rearrange this into just a triple, save that into the current predicate list and create a fact list out of that. The other two functions here are just following the usual pattern, getting that list of proposed facts and then approving the proposed facts into the approved list of facts. You can go ahead and put that list tools all together, ready to use by the fact agent, and now we can go ahead and construct it. Constructing the agent, since we've done all the hard work of already defining things, is pretty straightforward. And now you can go ahead and run it. Now, the initial state for this agent is the same as the initial state for the previous agent. So we're going to take advantage of that and just go ahead and make a copy of the end state of the previous agent for this agent. In the full multi-agent system, these would just be run in sequence and, you know, they would have the shared state in that way. so you can run them by themselves. We're going to make a copy and then go ahead and make a caller for the agent. And the message we're going to use for actually kicking things off is asking it to go ahead and make some proposals. There's a lot of output here, so I'm actually going to switch over to doing it this way without the comma True at the end. That will just let the agent do its work without seeing all of the output. If you run into a problem, use this one with the comma True at the end instead. Once the agent is done with all of its fact proposals, we're going to go ahead and take a look at the session state. We're going to assume that it has made a proposal and then complain if it hasn't. And also we're going to assume that it has not approved that proposal. Hopefully that works out. Okay, let's take a look at the proposals that it's made here. It's proposed facts like that a product has an issue, and that a product has received some customer feedback, product has an assembly time. That's an interesting one. A lot of different ones here. This looks like a pretty good list. And let's check the session state. Ah, so something did go wrong here. It actually sent us back this proposal, but it didn't save it to the session state. Now, why is that? Let's see if it has any complaints that it says. So that's unfortunate. So it's decided to actually make this proposal in just plain text, but it hasn't followed through with actually making the tool call that we thought that it would. So I'm going to go ahead and run this again. And of course in the full agent system, If it gets to the point where it makes this kind of proposal and at the end we realize that there are no proposed facts, we would go ahead and automatically just retry. So this rerun should work pretty cleanly because it's going to reinitialize the state and then go through the exact same steps. Okay. So we've got what look like pretty much the same fact types as before. It aligns with the user goals, asks us if we want to proceed. And let's see if it is actually It has correctly proposed those facts, but it's awaiting the approval. Okay. Great job, agent. You get an A+ this time. So again, if this doesn't work out for you, can just try to run it again, depending on the temperament of OpenAI on any given day. If it'll work first time or second time. It's usually not three times required. I'm going to go ahead and to send another message approving the proposal and that should just be copied over to the approved fact types. Looking at the session state, we see the approved fact types at the end. That's perfect. This has been a pretty good run.
