WEBVTT 00:00.080 --> 00:03.000 Please take time to understand this process very thoroughly. 00:03.280 --> 00:08.880 React prompting is used to allow agents to deduce which tools it needs to use logically. 00:09.320 --> 00:15.720 We will use react, which prompts the agents iteratively and requires to generate three main steps. 00:16.160 --> 00:19.400 Let's say we want a description of how a bird chirps. 00:19.760 --> 00:24.080 Thought is basically a logical explanation of what the model should perform. 00:24.960 --> 00:30.600 For example, I'll use Doc Retriever tool to find description of how the bird chirps. 00:31.160 --> 00:36.800 Then the next step is an action, a precise operation that must be executed. 00:37.080 --> 00:44.320 So in this case we have the tool doc retrievers, which input says description of how the bird chirps, 00:44.480 --> 00:48.040 and then the observation is the output produced by action. 00:48.360 --> 00:51.240 That is a tool call after it has been performed. 00:52.040 --> 00:58.840 For example, here birds chirps is a light, melodic sound that often feels chirps and rhythmic with 00:58.840 --> 01:01.240 a sequence of short, high pitched notes. 01:01.680 --> 01:06.360 So now we'll create the agent prompt to execute the desired react prompting. 01:06.720 --> 01:08.760 So here is the agent prompt. 01:08.920 --> 01:15.680 You are designed to help with a variety of multi-modal task and can perform multi hops to answer questions. 01:16.560 --> 01:18.920 This is the thought that we have provided. 01:19.440 --> 01:22.320 Now the next logical step is action. 01:22.520 --> 01:26.280 So we specify here you have access to a wide variety of tools. 01:26.480 --> 01:31.840 You are responsible for using tools in any sequence you deem appropriate to complete the task. 01:32.200 --> 01:37.000 Break the task into subtasks and iterate to complete each subtask. 01:37.920 --> 01:40.200 And here is where we provide the tools. 01:40.560 --> 01:44.600 And then the last step here is the observation or the output format. 01:44.880 --> 01:49.600 If you need to make a tool call, your response should follow the following structure. 01:49.960 --> 01:51.280 Provide the thought. 01:51.640 --> 01:55.360 Your reasoning process the tool that you invoked. 01:55.760 --> 01:58.800 The tool input in the JSON format which is input. 01:58.840 --> 01:59.720 Hello world! 01:59.960 --> 02:05.120 We have provided how the input should look like and the observation as tool response. 02:06.000 --> 02:11.840 And then this is what is the stop token or stop words by text criteria. 02:12.200 --> 02:18.410 And then this is where the agent will stop creating the or generating the response based on the tool 02:18.410 --> 02:19.170 response. 02:19.290 --> 02:22.330 You decide whether you need another more information. 02:22.650 --> 02:26.650 If so, make another tool call with the same structure. 02:27.530 --> 02:33.490 And then if you have enough information without using any more tools, you must give your answer to 02:33.490 --> 02:38.330 the question with final answer and respond in the following format. 02:38.890 --> 02:41.810 Here is what we have given the format thought. 02:42.170 --> 02:43.090 The reasoning. 02:43.290 --> 02:45.930 Final answer to humans. 02:45.930 --> 02:47.170 User question. 02:47.530 --> 02:52.730 This is the react prompt and this is the prompt template with the system as role. 02:52.970 --> 02:58.610 The agent prompt, which is the prompt that we provided along with the chat as user and the question 02:58.610 --> 02:59.770 that goes to it. 03:00.690 --> 03:03.690 Please take time to understand this very thoroughly. 03:04.090 --> 03:09.890 It's very important to understand the entire Agentic prompt here that will guide the model or large 03:09.890 --> 03:12.890 language model to follow certain processes. 03:13.250 --> 03:17.690 Invoke tools and do necessary actions depending on this prompt. 03:18.250 --> 03:19.050 Thank you. 03:19.090 --> 03:20.930 I'll see you in the next video.