1 00:00:03,000 --> 00:00:06,800 Welcome to this video on building Reflexion Agents. 2 00:00:06,800 --> 00:00:11,960 In this video, you'll build a Reflexion Agent by applying prompt engineering and schema design. 3 00:00:11,960 --> 00:00:19,959 You'll analyze AI responses and tool call outputs, and create an iterative feedback loop using Responder and Revisor nodes. 4 00:00:19,959 --> 00:00:25,639 Finally, you'll evaluate how citation-backed revisions improve the depth of responses. 5 00:00:25,639 --> 00:00:30,920 To build Reflexion Agents, begin with the essential imports to set up your environment. 6 00:00:30,920 --> 00:00:35,400 The next step is to configure the Tavily Search tool for use in your agent workflow. 7 00:00:35,400 --> 00:00:38,759 Set up the API key for the Tavily Search service. 8 00:00:38,759 --> 00:00:43,080 Create a Tavily Search tool instance that returns up to 5 results per query. 9 00:00:43,080 --> 00:00:46,720 Test the search tool with a sample query about breakfast recipes. 10 00:00:46,720 --> 00:00:50,680 The invoke method executes the search and returns structured results. 11 00:00:50,680 --> 00:00:55,060 The search results return as a list of JSON dictionaries, each with a dictionary containing 12 00:00:55,060 --> 00:00:58,580 keys like title, URL, and content. 13 00:00:58,580 --> 00:01:04,260 Next, create an instance of the ChatOpenAI model using GPT. 14 00:01:04,260 --> 00:01:08,019 Send the question to the LLM and extract the text response, such as, 15 00:01:08,019 --> 00:01:11,019 Any ideas for a healthy breakfast? 16 00:01:11,019 --> 00:01:15,940 The response is generic advice, like eating oatmeal, Greek yogurt, or avocado toast. 17 00:01:15,940 --> 00:01:20,360 The purpose is to create a Reflexion Agent that generates the most cutting-edge, but 18 00:01:20,360 --> 00:01:23,459 not yet proven, health ideas. 19 00:01:23,459 --> 00:01:28,440 The agent uses Reflexion to improve and evolve its responses over time. 20 00:01:28,440 --> 00:01:33,620 Create a system message to tell the LLM to behave like Dr. Paul Saladino, specializing 21 00:01:33,620 --> 00:01:37,779 in controversial approaches like carnivore diets and extended fasting. 22 00:01:37,779 --> 00:01:44,540 The first_instruction is a variable to alter the prompt. 23 00:01:44,540 --> 00:01:50,620 Populate the first instruction variable with a specific prompt that requests a 250-word response. 24 00:01:50,620 --> 00:01:53,019 Connect the LLM to the system message. 25 00:01:53,019 --> 00:01:58,699 The model will now behave according to the defined persona, as illustrated in the diagram. 26 00:01:58,699 --> 00:02:01,419 This setup is for demonstration purposes only. 27 00:02:01,419 --> 00:02:02,779 Pass the same question. 28 00:02:02,779 --> 00:02:07,980 The LLM responds with a suggestion for a carnivore diet, a diet that may have some benefits but 29 00:02:07,980 --> 00:02:11,639 is not commonly recommended by mainstream health authorities. 30 00:02:11,639 --> 00:02:18,860 Now alter the LLM output to match the requirements. 31 00:02:18,860 --> 00:02:21,259 Create the reflection schema or class. 32 00:02:21,500 --> 00:02:23,539 The fields describe what each field does. 33 00:02:23,539 --> 00:02:27,820 The LLM outputs the missing field for what information is needed and the superfluous 34 00:02:27,820 --> 00:02:30,139 field for what's unnecessary. 35 00:02:30,139 --> 00:02:35,320 The AnswerQuestion class includes the answer, reflections, and search queries. 36 00:02:35,320 --> 00:02:39,800 The Reflection class is instantiated in the AnswerQuestion class. 37 00:02:39,800 --> 00:02:44,979 The LLM eventually outputs a JSON object with the attributes filled out. 38 00:02:44,979 --> 00:02:48,220 Like before, chain the LLM to the system message. 39 00:02:48,220 --> 00:02:50,779 This will be a node in the LangGraph graph. 40 00:02:50,779 --> 00:02:56,820 You also need to bind the LLM to the AnswerQuestion schema using the bind_tools method. 41 00:02:56,820 --> 00:03:01,380 Essentially, the LLM will treat the AnswerQuestion class like a function. 42 00:03:01,380 --> 00:03:03,740 Pass the same message with the same question. 43 00:03:03,740 --> 00:03:07,440 The LLM responds with a suggestion for a carnivore diet. 44 00:03:07,440 --> 00:03:13,179 The result is not text but an AI message class that contains extensive information. 45 00:03:13,179 --> 00:03:17,899 But for reflection, the focus is on the tool_calls attribute. 46 00:03:17,899 --> 00:03:22,699 The tool_calls attribute contains the name of the class of the key-value pairs 47 00:03:22,699 --> 00:03:30,860 such as answer, reflection, missing, and superfluous with their corresponding text outputs. 48 00:03:30,860 --> 00:03:35,139 Response_list will be used to test the pipeline for one iteration. 49 00:03:35,139 --> 00:03:38,660 It will be analogous to the state variable in LangGraph. 50 00:03:38,660 --> 00:03:42,380 Append the user question and append the response from the responder. 51 00:03:42,380 --> 00:03:45,380 You can extract the search queries from the response. 52 00:03:45,380 --> 00:03:48,699 But let's create a LangGraph function for this purpose. 53 00:03:48,699 --> 00:03:54,100 Create a function for extracting the search queries from the responder message as a node for LangGraph. 54 00:03:54,100 --> 00:03:57,059 This node will also be used for the revisor. 55 00:03:57,059 --> 00:03:59,820 The state will be like the response list. 56 00:03:59,820 --> 00:04:01,779 Extract the AI message from the state. 57 00:04:01,779 --> 00:04:05,059 Next, extract the tool call parameters. 58 00:04:05,059 --> 00:04:07,539 Obtain the search queries from the LLM. 59 00:04:07,539 --> 00:04:11,059 Call the search tool and wrap the result in a tool message. 60 00:04:11,059 --> 00:04:15,639 Call the tool and append the tool response to response_list. 61 00:04:15,639 --> 00:04:18,420 There are many Reflexion prompt templates. 62 00:04:18,420 --> 00:04:23,920 Add a revisor system message guiding the LLM to act like Dr. Peter Accia, an expert in 63 00:04:23,920 --> 00:04:28,980 longevity and evidence-based health, and support responses with evidence. 64 00:04:28,980 --> 00:04:32,359 Update the prompt template to reflect this expert role. 65 00:04:32,359 --> 00:04:34,260 Create a schema for the revisor. 66 00:04:34,260 --> 00:04:40,339 The class is a subclass of answer question schema, so it will have the same fields and a citation list. 67 00:04:40,339 --> 00:04:44,480 Chain the system message and bind the LLM to the revise answer schema. 68 00:04:44,480 --> 00:04:47,100 This is another node in the LangGraph graph. 69 00:04:47,100 --> 00:04:51,600 The revisor passes the response list to revisor_chain. 70 00:04:51,600 --> 00:04:53,859 The result is an AI message. 71 00:04:53,859 --> 00:04:59,519 The args tool_call attribute has all the key-value pairs of the schema. 72 00:04:59,519 --> 00:05:01,940 Append the response to the response list. 73 00:05:01,940 --> 00:05:05,279 Now you have the second AI message from the revisor. 74 00:05:05,279 --> 00:05:08,579 You can feed the input back to the tool iteratively. 75 00:05:08,579 --> 00:05:12,420 Create a conditional node function that counts tool messages as a proxy for the number of 76 00:05:12,420 --> 00:05:15,579 iterations between the revisor and the search tools. 77 00:05:15,579 --> 00:05:18,179 It will count each time the tool is called. 78 00:05:18,179 --> 00:05:21,220 Build the graph by importing the necessary components. 79 00:05:21,220 --> 00:05:26,220 Then initialize the message graph and set the maximum iteration limit to 4. 80 00:05:26,220 --> 00:05:32,519 To create the graph, first add the respond node to the graph that uses the initial_chain. 81 00:05:32,519 --> 00:05:38,500 Then, add the execute_tools node that runs the search queries generated by the draft node. 82 00:05:38,500 --> 00:05:44,179 Finally, add a "revisor" node that uses the revisor_chain to improve the draft 83 00:05:44,179 --> 00:05:47,500 response using the search results and original critique. 84 00:05:47,500 --> 00:05:52,299 Next, connect the respond node to execute tools via an edge. 85 00:05:52,299 --> 00:05:55,299 And connect the tool to the revisor via an edge. 86 00:05:55,299 --> 00:05:58,140 Next, add the start and exit points. 87 00:05:58,140 --> 00:06:03,220 First, add a conditional edge from the revisor via the event_loop function to 88 00:06:03,220 --> 00:06:06,279 decide whether to continue iterating or end. 89 00:06:06,279 --> 00:06:09,519 The end node is hidden in event_loop. 90 00:06:09,519 --> 00:06:12,640 This outputs the response after a set number of iterations. 91 00:06:12,640 --> 00:06:16,640 Finally, set the entry point to the draft responder node. 92 00:06:16,640 --> 00:06:19,399 This is where the query is first processed. 93 00:06:19,399 --> 00:06:20,760 Compile the graph. 94 00:06:20,760 --> 00:06:26,160 Input the query, I'm pre-diabetic and need to lower my blood sugar, and I have heart issues. 95 00:06:26,160 --> 00:06:31,119 The result is a human message followed by a list of alternating AI and tool messages. 96 00:06:31,119 --> 00:06:33,839 You can extract the answer from the AI response. 97 00:06:33,839 --> 00:06:39,200 The initial responder provided a general advocacy for animal-based nutrition, recommending 98 00:06:39,200 --> 00:06:46,359 eggs, fatty meats, and organ foods, while avoiding grains and plant foods based on broad nutritional philosophy. 99 00:06:46,359 --> 00:06:48,600 This is not medical advice. 100 00:06:48,600 --> 00:06:52,779 Similarly, you can get the answer from the final revisor iteration. 101 00:06:52,779 --> 00:06:59,079 The revisor improved it by adding five scientific citations, measurable outcomes like postprandial 102 00:06:59,079 --> 00:07:05,040 glucose levels, and more precise guidance distinguishing processed from unprocessed foods. 103 00:07:05,040 --> 00:07:09,559 Otherwise, neither response adequately addresses the heart health concerns from the original 104 00:07:09,559 --> 00:07:15,880 query, focusing instead on blood sugar management while largely overlooking cardiovascular aspects. 105 00:07:15,880 --> 00:07:17,679 In this video, you learned that 106 00:07:17,679 --> 00:07:24,000 A search tool like Tavily can be configured and invoked to enhance AI responses with external data. 107 00:07:24,000 --> 00:07:28,040 Prompt engineering and schema design guide the LLM to produce structured reflections 108 00:07:28,040 --> 00:07:29,739 and focused answers. 109 00:07:29,739 --> 00:07:35,040 The answer question and reflection schema capture answers, flag missing or irrelevant 110 00:07:35,040 --> 00:07:37,679 details, and generate queries. 111 00:07:37,679 --> 00:07:44,640 Tool outputs like tool_calls and schema fields help extract structured insights from AI messages. 112 00:07:44,640 --> 00:07:49,640 LangGraph chains responder and revisor nodes into an iterative feedback loop using prompt 113 00:07:49,640 --> 00:07:52,160 updates and evidence-based revisions. 114 00:07:52,160 --> 00:07:58,359 A message graph orchestrates the Reflexion agent, managing node routing, iteration limits, and control flow.