WEBVTT 00:00.570 --> 00:02.790 Instructor: Just one minor change we need to make 00:02.790 --> 00:04.470 to our get text function. 00:04.470 --> 00:08.340 So sometimes the LLM may pass as our text 00:08.340 --> 00:12.030 some unnecessary characters which are non-alphabetical. 00:12.030 --> 00:16.350 So for example, backslash N or some single quote. 00:16.350 --> 00:19.980 So here I'm simply removing those and cleaning it up. 00:19.980 --> 00:22.980 Let's now review the React algorithm again. 00:22.980 --> 00:27.210 And after the tool execution, we got back the observation. 00:27.210 --> 00:30.200 It's the result of the tool after executing it. 00:30.200 --> 00:31.770 So now we have a choice, 00:31.770 --> 00:33.810 whether we want to give back an answer 00:33.810 --> 00:35.389 because we have enough information 00:35.389 --> 00:38.010 or we want to take the observation, 00:38.010 --> 00:40.080 the tool execution result, 00:40.080 --> 00:43.110 and we want to run another iteration of the React loop. 00:43.110 --> 00:45.060 But this time with all the history 00:45.060 --> 00:46.650 of what's been done so far 00:46.650 --> 00:49.960 so the agent won't be making redundant steps. 00:49.960 --> 00:53.340 So that's what we'll be implementing in this video. 00:53.340 --> 00:54.995 So let's go back to our code 00:54.995 --> 00:57.390 and let's head up to the React prompt. 00:57.390 --> 00:59.880 Now, remember that we removed the agent scratchpad 00:59.880 --> 01:01.876 from the original prompt. 01:01.876 --> 01:04.440 So now we're going to bring it back. 01:04.440 --> 01:06.060 And now we have better understanding 01:06.060 --> 01:07.140 what it's supposed to do. 01:07.140 --> 01:09.480 This is going to contain all the history 01:09.480 --> 01:11.459 and all the information that we had so far 01:11.459 --> 01:13.283 in the React execution. 01:13.283 --> 01:15.930 So we need to also add it to the dictionary 01:15.930 --> 01:17.310 that will be sending our prompt. 01:17.310 --> 01:21.060 So we'll have here another key of agent scratchpad. 01:21.060 --> 01:23.160 And just like before we're going to fill it up 01:23.160 --> 01:26.550 with a Lambda function that is going to access the key 01:26.550 --> 01:27.923 of agent scratchpad. 01:27.923 --> 01:31.620 So in order to keep track of the history of our agent 01:31.620 --> 01:32.850 and what happened so far, 01:32.850 --> 01:34.830 we're going to create a new variable 01:34.830 --> 01:36.420 called intermediate steps. 01:36.420 --> 01:38.220 And this is going to be a list 01:38.220 --> 01:40.710 which is going to start as an empty list. 01:40.710 --> 01:43.800 So let's plug it in into our agent invocation. 01:43.800 --> 01:45.731 So we're going to add here another he, 01:45.731 --> 01:48.420 and it's going to be the agent scratch pad. 01:48.420 --> 01:52.080 And this is going to hold the value of the empty list, 01:52.080 --> 01:54.480 which is the intermediate steps. 01:54.480 --> 01:56.670 Now, every time we run an iteration, 01:56.670 --> 01:59.066 we want to update this list and append to it the history 01:59.066 --> 02:01.410 and what we have performed. 02:01.410 --> 02:05.520 So after we perform the tool selection and tool execution 02:05.520 --> 02:07.784 and get back the observation, I'm going to append 02:07.784 --> 02:10.648 to our intermediate steps which hold the history 02:10.648 --> 02:12.630 our agent step. 02:12.630 --> 02:14.310 So this is what we got back 02:14.310 --> 02:16.445 after we parsed the LLM's answer, 02:16.445 --> 02:19.200 and I'm going to append the observation, 02:19.200 --> 02:21.990 what we got back from the tool after running it. 02:21.990 --> 02:24.510 So that way our agent will have both 02:24.510 --> 02:28.500 its reasoning engine history and what it has chosen already 02:28.500 --> 02:31.350 and also the result of what was the result 02:31.350 --> 02:33.390 of the tool execution. 02:33.390 --> 02:34.680 Let's run it in debug. 02:34.680 --> 02:37.747 And we want to examine the values of agent step 02:37.747 --> 02:39.480 and the observation 02:39.480 --> 02:41.910 and what we're saving in the intermediate steps, 02:41.910 --> 02:43.288 which is our history. 02:43.288 --> 02:46.680 So I'm simply going to select everything here 02:46.680 --> 02:48.205 and evaluate it. 02:48.205 --> 02:52.740 Now everything looks good so far except for the fact 02:52.740 --> 02:56.850 that the first element in our typo is an agent action, 02:56.850 --> 02:58.950 which is an LangChain object. 02:58.950 --> 03:01.890 And the LLM doesn't understand LangChain object. 03:01.890 --> 03:05.100 So we will need to translate this into text. 03:05.100 --> 03:07.143 So we'll need to format it nicely. 03:08.049 --> 03:10.950 So don't worry, we're not going to do it ourselves. 03:10.950 --> 03:12.570 LangChain is going to supply us 03:12.570 --> 03:15.026 with a utility function that performs it. 03:15.026 --> 03:19.873 So I'm going to import the function format log to string. 03:19.873 --> 03:23.850 And this function is going to take our intermediate steps, 03:23.850 --> 03:28.350 which is a list that contains the tool of the agent action, 03:28.350 --> 03:31.350 what the agent has decided, which tool to use, 03:31.350 --> 03:34.620 and the observation, what was the result of the tool. 03:34.620 --> 03:37.470 So it's then going to construct the scratchpad 03:37.470 --> 03:39.690 that is going to enable the agent 03:39.690 --> 03:41.910 to continue in its thought process 03:41.910 --> 03:45.329 and basically it's going to format all those strings nicely. 03:45.329 --> 03:49.110 So instead of passing the original agent scratchpad, 03:49.110 --> 03:51.665 we're going to first format it nicely 03:51.665 --> 03:55.680 by using the format log to STR function, 03:55.680 --> 03:59.040 and that way we'll pass to the LLM only strings 03:59.040 --> 04:01.500 and that's way the LLM would be able 04:01.500 --> 04:03.930 to take account of our history. 04:03.930 --> 04:06.810 So let's add another step of the agent. 04:06.810 --> 04:10.650 So I'm simply going to copy this agent step invocation 04:10.650 --> 04:12.180 and going to paste it. 04:12.180 --> 04:15.423 So let's print what we get in the second step of the agent. 04:16.650 --> 04:18.690 And now let's run everything in debug 04:18.690 --> 04:19.990 and let's see what we get. 04:21.660 --> 04:24.480 So when we run it in debug, we'll get an error 04:24.480 --> 04:29.480 that the output parser got as an input an input 04:29.790 --> 04:34.225 that also has a final answer and also a parsible action. 04:34.225 --> 04:36.360 And that confuses the LLM 04:36.360 --> 04:38.310 because it should either have an action 04:38.310 --> 04:39.930 and action input to perform 04:39.930 --> 04:41.460 or it should have the final answer. 04:41.460 --> 04:42.450 It can't have both. 04:42.450 --> 04:44.010 This is the React algorithm 04:44.010 --> 04:47.130 and this is the logic of the output parser. 04:47.130 --> 04:50.610 So from my experience, the usually fix for those kinds 04:50.610 --> 04:53.160 of issue is to work around 04:53.160 --> 04:56.100 and refine the prompt that is being sent to the agent. 04:56.100 --> 04:58.020 So you can see right now that in this prompt, 04:58.020 --> 05:00.720 we have a space at the end before the question mark, 05:00.720 --> 05:02.223 so we're going to remove it. 05:04.710 --> 05:07.707 And let's also not forget the previous invocation 05:07.707 --> 05:09.540 of the agent. 05:09.540 --> 05:11.409 So we also need to fix that. 05:11.409 --> 05:13.526 And let's now rerun everything 05:13.526 --> 05:16.209 and examine the output. 05:16.209 --> 05:19.740 We can see right now we got the first step of the agent, 05:19.740 --> 05:20.940 the first invocation. 05:20.940 --> 05:22.710 It was the first iteration. 05:22.710 --> 05:25.650 We saw that we got the answer from the LLM 05:25.650 --> 05:28.590 that decided that we should use the tool get text length. 05:28.590 --> 05:31.320 We then used the output parser to parse it, 05:31.320 --> 05:33.329 and we then chose the correct tool 05:33.329 --> 05:35.910 and we invoked its function. 05:35.910 --> 05:39.780 We did that using the agent action object of LangChain, 05:39.780 --> 05:42.960 which holds all the information about the tool to execute. 05:42.960 --> 05:44.340 We then executed the tool 05:44.340 --> 05:45.707 and we have now the result 05:45.707 --> 05:48.840 of that tool of get text length. 05:48.840 --> 05:51.585 So this result is considered as an observation, 05:51.585 --> 05:55.230 and then we take the agent action and the observation 05:55.230 --> 05:57.606 and we append it to the intermediate step list 05:57.606 --> 06:01.920 and we send it as the agent scratchpad. 06:01.920 --> 06:04.860 So this is the history of what has been done so far, 06:04.860 --> 06:08.040 what tools were chosen and what was the result. 06:08.040 --> 06:10.566 And this is being sent to the second iteration 06:10.566 --> 06:12.630 of the React loop. 06:12.630 --> 06:16.049 So now we're starting all over again, but with the history. 06:16.049 --> 06:20.460 However, now in the second invocation in the scratchpad, 06:20.460 --> 06:22.050 the agent has the observation, 06:22.050 --> 06:23.940 which is the result of the tool 06:23.940 --> 06:27.600 that counts the number of letters in the word doc. 06:27.600 --> 06:29.363 So we have the final answer. 06:29.363 --> 06:32.550 And actually the object that we're dealing with right now 06:32.550 --> 06:35.446 is called agent finish and not agent action. 06:35.446 --> 06:37.530 And I want to run it in debug 06:37.530 --> 06:39.787 in order for you to see it for yourself. 06:39.787 --> 06:44.040 So, remember that we told that the output parsing return 06:44.040 --> 06:46.860 as either agent action and agent finish. 06:46.860 --> 06:49.969 So in this case it returned the agent finish. 06:49.969 --> 06:53.820 So if we'll go to the agent step, we can see 06:53.820 --> 06:56.065 that we're dealing with an agent finish object, 06:56.065 --> 07:00.150 and this is what the output passer returned for us. 07:00.150 --> 07:02.010 And why did it return an agent finish 07:02.010 --> 07:03.540 and not an agent action? 07:03.540 --> 07:05.940 Because if you take a look at the log entry, 07:05.940 --> 07:08.785 the LLM responded to us, I know the final answer 07:08.785 --> 07:11.370 and then backslash then final answer, 07:11.370 --> 07:14.700 and the output parser parsed it and noticed it, 07:14.700 --> 07:17.730 and it decided that it now has the final answer, 07:17.730 --> 07:20.760 so it should return an agent finish object. 07:20.760 --> 07:23.280 Lastly, in the field of return values, 07:23.280 --> 07:25.804 we have a dictionary with the output key 07:25.804 --> 07:29.820 and the value is the answer of the agent for our question. 07:29.820 --> 07:33.450 So now I remove the print statement of the agent action, 07:33.450 --> 07:36.780 and we're checking if the agent step is an instance 07:36.780 --> 07:38.340 of the class agent finish. 07:38.340 --> 07:41.130 And if this is the case, then the agent step is going 07:41.130 --> 07:44.730 to have the field of return values and we can print it 07:44.730 --> 07:47.100 and finish our execution. 07:47.100 --> 07:48.843 So let's now rerun it. 07:51.840 --> 07:56.040 And we can indeed see that we had two React iterations 07:56.040 --> 07:58.410 and we got the final answer for our agent, 07:58.410 --> 07:59.493 which used our tool.