LangGraph: Force-Calling a Tool

Key Takeaways

in this video we're going to make a really simple modification to the chat agent executive so that we always call a tool first so if you haven't already watched the chat agent executive video please do that um we will work off of that uh notebook that we have there and we'll only make a few modifications and in this video we're only going to cover those modifications so if you want more context on what's going on in the notebook please check out the the video and the notebook for the base chat execu all right so we're going to set this up where Force calling tool first most of the setup the same we're going to create our tools we're going to create this tool executor to column we're going to create our model we're going to bind the tools to the model we're going to Define an agent State this is all exactly the same and we're going to actually Define the same exact uh states that or nodes and and edges that we had before but we're also going to Define an additional node so this is we're going to call this the first model node and this is basically uh this is going to be the first node that gets called and we want this to return a message that says to call a specific tool so we have this tavil search results uh Json tool that's the name of the tool that we defined above if you wanted to call a different tool you have to you have to pass in the different name here and then we're going to pass in the inputs that we want it to call with so we're going to basically call it with query and we're going to get the most recent message and get the content from that message and pass that in as a query so we're going to create this function and now we're going to modify the graph so compared to the previous graph we now have this new entry point which is called first agent we have the same two other entry points as before agent and action and we're going to set the entry point to be first agent so this is a change from before as well so now we're setting that this first agent node is always called first we then have a conditional node from agent to action or end this is the same as before we now have this node from action to agent this is the same as before but but we have this new node from first agent to action which we're going to set and this is basically saying after we call the first agent we always want to call an action and we know that we always want to call an action because we are forcing it to call a AI message with this function call so we always want it to call this tool so we can create this graph um and now if we use it we can see that we get the first result back really really fast because not actually calling a language model it's just passing this message in as the argument um and then we can see the other things as well if we take a look at what's going on under the hood in lsmith we can see this even better so we have this result here we can see that the first agent you know we pass in this message and we get back this other message but there's actually no language model call going on under the hood um and the tool is really the first thing that's ever invoked um and then there is a language model call at the end but we can see that we skip this entirely because we are forcing the uh the the agent to basically always call the tool first