WEBVTT 00:00.110 --> 00:00.710 Hey there. 00:00.740 --> 00:01.310 Eden here. 00:01.310 --> 00:06.500 And in this video I'm going to show what are the differences between the function calling agent and 00:06.500 --> 00:12.500 the react agent in the long chain implementation, when to use one over the other, and what are the 00:12.500 --> 00:15.350 advantages and disadvantages of each one? 00:16.460 --> 00:22.730 And like all things in software today, it all boils down where do we shift our responsibility to who 00:22.730 --> 00:27.770 is responsible for the agent selecting the correct tool with the correct input? 00:27.770 --> 00:34.910 And I think the analogy to serverless is appropriate here, because in serverless we shift all the responsibility 00:34.910 --> 00:42.530 for the availability, durability, scalability, all those ilities we shift it to the vendor, to the 00:42.530 --> 00:43.640 cloud provider. 00:43.940 --> 00:48.500 Now, when it comes to agents, we have two paradigms of tool selection. 00:48.500 --> 00:50.330 The one is function calling. 00:50.330 --> 00:57.350 And in function calling we supply a certain schema to the vendor, specifying all of our functions that 00:57.350 --> 00:59.450 we want to equip our agent with. 00:59.450 --> 01:03.900 And after that, our LM is sort of equipped with those tools. 01:03.930 --> 01:10.620 The LM may respond to us in its response that we need to invoke a tool with certain arguments according 01:10.650 --> 01:12.540 to the specification we send it. 01:12.780 --> 01:19.050 Now, just like in serverless here, we are not exposed for the logic that happens under the hood, 01:19.050 --> 01:23.370 for how the LM is determining how to select the tool. 01:23.520 --> 01:29.940 By the way, my guess here is that this is a fine tuned model, specifically fine tuned for tool selection 01:29.940 --> 01:32.040 similar to the gorilla paper. 01:32.820 --> 01:38.670 So in the function calling agent, the responsibility for selecting the tools is on the vendor, the 01:38.670 --> 01:39.510 LM vendor. 01:39.510 --> 01:45.750 And of course I have to note it's a shared responsibility model because we need to make sure that our 01:45.750 --> 01:49.920 tool descriptions are not ambiguous and that they are well defined. 01:49.950 --> 01:52.740 So the LM would have an easier job to do it. 01:53.190 --> 01:56.790 And today many vendors support function calling or tool calling. 01:57.000 --> 01:58.200 OpenAI. 01:58.230 --> 01:59.160 Google. 01:59.190 --> 02:00.030 Anthropic. 02:00.030 --> 02:00.690 Mistral. 02:00.720 --> 02:02.700 They all support tool calling. 02:02.700 --> 02:09.420 And luckily for us, Late Chain implemented an abstraction for this a well defined interface for using 02:09.450 --> 02:12.450 tool, calling for every vendor that supports. 02:12.480 --> 02:18.570 So at the moment, switching between models that support tool calling is super easy. 02:18.570 --> 02:22.260 And we can switch models with tool calling like we switch our socks. 02:22.770 --> 02:23.550 Alrighty. 02:23.580 --> 02:30.780 Now in a react agent, it's an agent that specifically is using the react prompting, which is based 02:30.780 --> 02:32.460 on the react paper. 02:32.520 --> 02:37.770 Here it all boils down to the react prompt that is sent to the LM. 02:37.770 --> 02:45.150 This is a well crafted prompt that the Link Chain team wrote and was inspired by the react paper, and 02:45.150 --> 02:50.250 incorporates a lot of prompt engineering techniques like chain of thought and few shot prompting. 02:50.250 --> 02:55.800 And I have to say, I think it's the most beautiful prompt today in prompt engineering, so well, job 02:55.800 --> 02:56.370 link chain. 02:56.370 --> 02:57.720 This is an awesome prompt. 02:57.720 --> 03:04.900 Anyways, the whole idea of this prompt is to make the LM come our reasoning engine to select which 03:04.900 --> 03:05.950 tools to use. 03:05.980 --> 03:11.680 And it turns out that this prompt is actually very useful for turning the LLM into a reasoning engine, 03:11.680 --> 03:18.580 and that in a lot of cases, it does return the correct tool to use with the correct inputs, and the 03:18.580 --> 03:26.110 LLM output is usually action and action input, which Linkchain knows how to parse and to deduce from 03:26.110 --> 03:26.290 it. 03:26.320 --> 03:27.970 Which tool do we need to use? 03:28.000 --> 03:30.520 They do it with some regular expressions. 03:30.550 --> 03:37.090 Now, after we run the tool and we get a result back from our tool execution, then Linkchain labels 03:37.090 --> 03:38.800 it as an observation. 03:38.800 --> 03:44.080 And then we start to reiterate and to run again a prompt with them. 03:44.080 --> 03:49.060 But this time it will contain the history of what the LLM decided so far. 03:49.060 --> 03:52.060 Which tool did we use and what was the result of that tool? 03:52.060 --> 03:53.410 What was the observation? 03:53.410 --> 03:57.970 And this was a very high level explanation of the react loop. 03:57.970 --> 04:05.530 But basically here it all boils down to the react prompt and the react prompt is what leverages the 04:05.530 --> 04:07.660 LLM to become our reasoning engine. 04:07.660 --> 04:10.660 So we have full control of this prompt. 04:10.660 --> 04:15.730 And if we want, we can tweak it around and we can customize it according to our needs. 04:15.730 --> 04:19.990 So we have much more flexibility here for the tool selection part. 04:20.410 --> 04:21.250 Alrighty. 04:21.250 --> 04:22.990 So which one is better? 04:22.990 --> 04:25.870 And to be honest, I don't have the exact answer for that. 04:25.870 --> 04:28.510 Like everything in software, it depends. 04:28.660 --> 04:34.180 If we want full control and full flexibility, we can use the react prompt. 04:34.180 --> 04:41.830 But the problem here is that all the responsibility is on us as developers while in function calling. 04:41.830 --> 04:48.460 Then the tool selection responsibility is on the vendor and this saves us a lot of work and a lot of 04:48.460 --> 04:53.200 thinking, because all the work for the tool selection was done by the vendor. 04:53.200 --> 04:54.730 We have here less control. 04:54.760 --> 05:01.000 However, we do have here much less headache and that was pretty much it. 05:01.000 --> 05:02.380 I hope you enjoyed the video. 05:02.380 --> 05:04.360 Please let me know in the comments what you think.