Chatting with ArXiv Research Papers

00:00:00.000 | Okay, so today we have an update on the AI Assistant project that we've been working on for a little while.

00:00:07.680 | It's kind of been on the back burner for a while with everything else that's been going on,

00:00:13.200 | but today I think it's probably the first exciting part of the tool that you can actually use.

00:00:20.640 | And, well, let me just show you what it looks like.

00:00:23.840 | So come over to Colab here, I just do the PIP install, the most recent version, so 0.5.

00:00:32.240 | And then we set our environment variables, so that's just API keys that we need from OpenAI and Pinecone.

00:00:37.680 | And then I come down to here and I initialize, so this is a new thing, this archiver bot, all right?

00:00:44.800 | So initialize that, and this archiver is everything we need to begin interacting with the code library

00:00:52.080 | that we've already built in the previous two videos.

00:00:56.000 | So the downloading of papers and the search through those papers,

00:01:02.800 | that can now, or that will be fully controlled through the chatbot.

00:01:07.120 | At the moment, it's just a search that is implemented

00:01:10.240 | and kind of in the progress of building the upload part of that as well.

00:01:15.760 | So we can run this and this is a question that is kind of covered in one of the papers

00:01:21.920 | that I have added, and it's, what is the general scientific consensus

00:01:26.960 | on development tool usage in early hominids?

00:01:29.280 | So it's from this paper here on, it's not even a paper,

00:01:33.040 | I think it's a chapter of a book on the evolution of intelligence.

00:01:36.080 | Now, I noticed adding like a non, like AI, computer science article into the dataset,

00:01:44.400 | that there are a few things with the library that doesn't quite work yet,

00:01:47.200 | like references here are mostly not archive papers.

00:01:51.280 | And I don't, there might not even be any archive papers in there.

00:01:54.160 | So I need to figure out a better way of dealing with that

00:01:57.680 | when we're doing like the propagation approach to finding all of the relevant papers.

00:02:03.440 | And I'm not sure how well it's being processed,

00:02:06.960 | like the chunks, maybe the size of chunks needs to be increased or decreased.

00:02:11.040 | I'm not sure.

00:02:12.720 | So let's have a look at this.

00:02:14.720 | What did we get?

00:02:15.760 | There's a scientific consensus that early hominids developed tool usage with Homo habilis,

00:02:21.280 | making tools a long time ago,

00:02:24.800 | Homo erectus showing evidence of enhanced intelligence with task-specific stone hand axes.

00:02:30.960 | Okay. So, I mean, this is a pretty good answer, I think.

00:02:34.640 | And it also tells me the source, right?

00:02:36.800 | So I could ask chatGPT this, and it may or it may not give me a good answer.

00:02:41.840 | I think on this sort of topic, it probably would give a decent answer,

00:02:44.960 | but I don't know where that information is coming from.

00:02:47.280 | And I don't want to just blindly rely on that.

00:02:49.600 | Okay. Especially when you're doing actual research

00:02:52.000 | and you need to source what you're talking about, right?

00:02:57.600 | So I can click on here and it's going to take us to that paper that we saw before.

00:03:02.080 | Okay. And we see it's the evolution of intelligence, which is from a book, I think.

00:03:09.200 | Then, okay. Could you tell me about development tool usage in this context?

00:03:13.680 | So it's kind of interesting.

00:03:14.800 | They have like the, we have some basic simple stone tools,

00:03:18.640 | and then it's talking about stone hand axes.

00:03:21.600 | It's kind of cool.

00:03:22.320 | So, okay. Tell me more about that.

00:03:25.280 | And this does take a little while to run.

00:03:27.120 | Maybe that's something we can try and improve.

00:03:29.760 | But I think right now it's kind of a consequence of what is happening.

00:03:34.960 | We're like, we're making multiple calls, open AI.

00:03:38.640 | And it just, it takes a while to process.

00:03:41.040 | Even though we're using, in this case, we're using GPT-3.5 Turbo,

00:03:44.400 | which is one of the faster models.

00:03:46.080 | I mean, if you did this with GPT-4, it would take a very long time.

00:03:49.040 | But anyway, so, okay.

00:03:51.920 | We've got this output they developed.

00:03:54.720 | Is this what I want?

00:03:57.920 | Sorry. It's a development of tool usage.

00:04:00.880 | And I'm kind of asking the same thing.

00:04:03.120 | Okay. Let me change that.

00:04:05.760 | So let's ask, okay.

00:04:07.840 | Tell me more about those hand axes.

00:04:09.920 | Tell me more about these hand axes.

00:04:15.200 | Okay. Let's try that.

00:04:18.320 | Okay. So it's being specific.

00:04:21.760 | Like it knows, I just said these hand axes,

00:04:24.240 | but because we have the chat history in there,

00:04:27.440 | it knows to actually search for the Homo erectus hand axes,

00:04:31.200 | because it can see the previous interactions that we had.

00:04:35.120 | Okay. And the answer is just that.

00:04:37.440 | Okay. It doesn't actually tell us much.

00:04:39.120 | Okay. And let's see if there's something else that we can have.

00:04:44.320 | So there's also evidence of great similarities

00:04:46.400 | between humans and great apes.

00:04:48.080 | So let's ask about that.

00:04:50.000 | So what are the similarities between humans and great apes?

00:04:58.800 | Okay. You can see again, it's creating the search term here.

00:05:06.000 | And our answer is similarities between humans and great apes

00:05:09.440 | include basic symbolism and creativity,

00:05:11.520 | which evolved as an adaption to forested environments

00:05:14.480 | of Eurasia during the Miocene.

00:05:16.160 | And then evidence of great similarities

00:05:20.560 | between humans and great apes and intelligent,

00:05:22.560 | in intelligence and traditionally believed has been found

00:05:25.920 | the apes ancestors from the mid-late Miocene

00:05:28.960 | had brains of comparable size.

00:05:30.880 | So these intellectual capabilities

00:05:33.280 | may have been potentiated as early as 12 to 14 million years ago.

00:05:41.520 | All right. Okay.

00:05:43.440 | So I think the point here is that we probably had like

00:05:47.120 | a common ancestor around that time ago.

00:05:49.520 | And at that point, we already, you know,

00:05:52.240 | the intelligence of humans was already well on its way

00:05:54.720 | to becoming what it was now.

00:05:56.640 | Right. Okay. Cool.

00:06:00.240 | So that is like the bot interface.

00:06:04.960 | Let me talk a little bit more about where this is going.

00:06:09.440 | Okay. Okay.

00:06:10.640 | So come over to, this is a new code that we have.

00:06:14.720 | So this is like the chat interface thing that we're doing.

00:06:18.160 | We're using line chain to create all of this.

00:06:20.400 | So we have a few things.

00:06:23.040 | So we have a couple of prompts that we're creating here.

00:06:26.080 | The initializing function of this class

00:06:29.520 | is basically just setting everything up.

00:06:32.000 | So initializing the large language model,

00:06:34.640 | the memory retrieval component,

00:06:37.040 | the chat bot, and also this splitter.

00:06:40.400 | So this, we haven't used this yet.

00:06:42.480 | This is when we start uploading documents via this function.

00:06:47.120 | So that's something I'll talk about a little bit later.

00:06:51.440 | So when we call this function, it just executes, right?

00:06:55.360 | So it executes on the agent that we create.

00:06:57.440 | So let's have a look at how we actually create those.

00:07:01.120 | So first we initialize a large language model.

00:07:03.760 | We can see that here.

00:07:05.440 | So that's just a chat model via line chain,

00:07:08.400 | gpt 3.5 turbo by default.

00:07:11.200 | And then we go on and we initialize our memory.

00:07:15.600 | So this is one of the more complicated parts.

00:07:17.760 | So the memory, we have this pine cone object,

00:07:22.160 | which is actually from here.

00:07:25.120 | Okay. So we have this pine cone class that we created.

00:07:28.560 | So we initialize that.

00:07:30.720 | We initialize a encoder via a line chain.

00:07:34.960 | We initialize a vector DB via line chain.

00:07:37.200 | So there's a bit of repetition here.

00:07:39.920 | We have this vector DB,

00:07:41.040 | and then we also have another version of it here.

00:07:43.120 | Something needs to be cleaned up.

00:07:45.360 | This is like the first version that I wrote pretty quickly.

00:07:48.720 | And then we have the retriever.

00:07:51.360 | So this is a retrieval Q&A with sources chain from line chain.

00:07:55.200 | So basically we're going to retrieve relevant documents,

00:07:58.160 | and we're also going to include the sources in there.

00:08:00.320 | That's super important.

00:08:02.000 | I tried in these prompts to make sure that it's going to include

00:08:05.360 | those sources of information all the time.

00:08:07.760 | Okay. So that's important.

00:08:10.800 | Okay.

00:08:11.940 | And then we come down here and we initialize a search tool, right?

00:08:17.200 | So this is just one of the tools.

00:08:18.880 | There will be multiple tools that the agent can use.

00:08:21.360 | This is just one of them.

00:08:22.400 | And it's like the core functionality,

00:08:24.640 | which is where we're chatting with the agent,

00:08:26.960 | and we want it to refer to these archive papers.

00:08:30.240 | Okay.

00:08:30.740 | This is how it does it, right?

00:08:33.040 | So it uses the search function, which is here.

00:08:38.320 | Okay.

00:08:38.820 | This is basically just a custom version of the retrieval tool

00:08:45.680 | that is already within line chain.

00:08:48.240 | The only difference is that we add in the sources

00:08:50.800 | to the answer response or the answer value here,

00:08:54.240 | because before it didn't seem to be using this,

00:08:58.160 | because you basically get a answer value and a sources value.

00:09:02.480 | It didn't seem to be considering the sources value in the final answer.

00:09:06.640 | So I just forced them into the actual answer value here.

00:09:10.000 | That's the only difference.

00:09:11.520 | I'll probably make some other changes to that in the future,

00:09:14.240 | but for now, that's fine.

00:09:17.120 | And then, okay.

00:09:19.120 | So that's the search function that our tool uses.

00:09:24.160 | We have a description, which is the search description here.

00:09:28.640 | So we need to use this tool when searching

00:09:30.560 | for scientific research information

00:09:32.400 | from our prebuilt archive papers database.

00:09:35.360 | This should be the first option when looking for information.

00:09:37.680 | When receiving information from this tool,

00:09:40.320 | you must always include all sources of information, right?

00:09:43.520 | So again, sources of information, super important.

00:09:46.560 | And then we append that tool to our tools list, right?

00:09:49.520 | So then that's kind of ready to be used by our agent, okay?

00:09:55.280 | And then we initialize our agent.

00:09:57.120 | That's the next step here.

00:09:58.960 | So the chatbot agent.

00:10:00.480 | So that is here, right?

00:10:04.000 | So we use the conversational memory.

00:10:06.640 | I haven't really messed around so much with this at the moment,

00:10:09.120 | but we use conversational buffer window memory.

00:10:11.440 | So we keep basically a track of the last five interactions.

00:10:15.120 | And then we forget it, okay?

00:10:17.440 | Then we initialize the agent.

00:10:20.400 | So we're using the chat conversational react description agent.

00:10:23.840 | What that means is that it's for a trap model,

00:10:27.040 | which is the GT 2.5 turbo model that we've initialized.

00:10:31.760 | It's conversational, meaning there is this conversational memory

00:10:35.600 | considered by the agent.

00:10:37.280 | So it's not just looking at a single interaction.

00:10:41.760 | It's looking at a history of interactions.

00:10:44.640 | The react is like a framework where you are basically saying,

00:10:49.920 | I want you to reason about what action to take and then take the action, okay?

00:10:56.320 | So the RE is reason and the act is action, okay?

00:11:04.480 | And then description is basically saying,

00:11:06.000 | base your decision on whether to use this agent on the agent description,

00:11:12.640 | which we defined here, okay?

00:11:16.320 | So that is our initialized agent.

00:11:22.240 | And then we just update the prompt.

00:11:24.480 | So we have this custom system message up here, okay?

00:11:28.880 | Which is your expert summarizer and deliver technical information.

00:11:32.320 | You make complex information incredibly simple to understand,

00:11:36.160 | so on and so on, okay?

00:11:37.920 | I just wrote this quickly.

00:11:39.280 | It definitely needs some work.

00:11:41.680 | But one thing I did add that again is pretty important

00:11:44.720 | is just really telling the agent again and again

00:11:48.880 | to include a source of information at the end of responses, okay?

00:11:53.680 | So I just like hammer that in as often as I can

00:11:57.520 | just to make sure that it's actually doing that, okay?

00:12:00.640 | So after that, we're done, right?

00:12:05.280 | So that is the agent, it's created.

00:12:08.400 | And then the final bit that is in this initialization component

00:12:12.800 | is initialize the extraction tooling.

00:12:15.680 | That is for later that I'm not using that right now,

00:12:19.760 | but I will be at some point.

00:12:22.400 | And I kind of show you what that is going to look like.

00:12:24.800 | So if we come over to here,

00:12:26.880 | we have basically the same code we saw before.

00:12:29.600 | This is running locally.

00:12:31.280 | So I can like test the most recent version of the bot.

00:12:37.680 | So we just asked questions.

00:12:39.440 | It's going to do the exact same thing, right?

00:12:42.080 | I'm AI language model, right?

00:12:43.840 | It's not accessing the database at this point.

00:12:46.080 | If I ask you a technical question,

00:12:47.520 | it's going to access the database, right?

00:12:50.640 | Okay, cool.

00:12:52.160 | Then the next thing that I'm doing here

00:12:56.880 | is I want to be able to use this chat interface

00:13:00.080 | to add new articles to the database, right?

00:13:03.120 | At the moment, we have to use a different script

00:13:06.640 | where we manually, not manually,

00:13:08.880 | but we write code in order to add those papers

00:13:14.560 | to the database.

00:13:15.520 | I don't want to do that.

00:13:16.320 | I want to let users interact with the chatbot

00:13:18.800 | and say, I would like to talk about this particular paper

00:13:24.240 | and for that paper to not be within the database already.

00:13:28.000 | And the chatbot know that, okay,

00:13:31.680 | if they want to talk about that,

00:13:33.280 | we need to search the internet.

00:13:34.400 | We need to retrieve that paper.

00:13:36.640 | We need to embed it, add it to the database,

00:13:39.280 | and then we can start talking about it, right?

00:13:41.600 | I'm not sure because that takes a little bit of time.

00:13:43.760 | So at some point there'll need to be

00:13:46.560 | some like async code in here,

00:13:48.160 | which handles that and then says, okay,

00:13:49.840 | you need to give me a bit of time to think about it.

00:13:51.840 | Something like that.

00:13:53.600 | I'm not sure yet,

00:13:54.400 | but for now it's just creating that, right?

00:13:59.520 | So the way that that will be implemented,

00:14:03.040 | I'm not a hundred percent sure yet,

00:14:04.480 | but I've started doing it.

00:14:06.160 | So we have this add article to database function, right?

00:14:11.920 | That is just getting the archive ID, basically.

00:14:16.000 | It's using that to download.

00:14:18.880 | So this is from constructors, I think.

00:14:21.120 | Yes, it's an archive class.

00:14:25.120 | Okay, archive class here.

00:14:26.400 | This is another function from the library.

00:14:28.720 | So that downloads the archive object.

00:14:32.880 | Or we get information on it.

00:14:34.480 | Then we download it.

00:14:36.080 | We get the metadata for the paper,

00:14:38.560 | and then we shrink the paper in smaller parts.

00:14:40.640 | So into the like 300 token slices,

00:14:44.160 | and then we add that to the database.

00:14:47.040 | Okay, right?

00:14:48.080 | And then we return,

00:14:49.280 | or we're going to return something like this.

00:14:50.880 | So we're going to return that saying,

00:14:52.240 | I've added this to my memory.

00:14:53.760 | It's now accessible via the search archive database tool.

00:14:57.920 | So this is both kind of for the user to see,

00:15:01.360 | but also actually more for the chatbot to see.

00:15:04.400 | So once it's seen that,

00:15:05.520 | okay, we have this in our memory now,

00:15:08.000 | we can now access it through this other tool, right?

00:15:11.280 | So if it's going through those multiple steps,

00:15:13.200 | it can download it,

00:15:13.920 | and then it can ask a question to the tool.

00:15:16.240 | Okay?

00:15:17.860 | So that is a function.

00:15:20.640 | I've tested it here.

00:15:21.600 | So I added the attention is all you need paper.

00:15:24.880 | You know, that worked.

00:15:25.840 | And then I asked the question,

00:15:29.760 | and it retrieved the relevant information from that paper.

00:15:32.560 | That's kind of what I want to do,

00:15:33.920 | but I don't want to write code like this to do it.

00:15:37.680 | I would rather write something like,

00:15:39.280 | but can you add the like 1706.03762 paper to your memory?

00:15:54.800 | Okay?

00:15:55.360 | I want to say something like that,

00:15:56.880 | or rather than that,

00:15:58.480 | I would like to say, can you add the attention

00:16:01.920 | is all you need paper to your memory,

00:16:03.840 | or can you add some papers, right?

00:16:07.200 | So this is where we might search for multiple papers,

00:16:09.840 | some papers about NLP attention to your memory, right?

00:16:15.280 | I want to be able to ask these things to the bot,

00:16:17.440 | and they actually go through that process

00:16:18.960 | that I just described.

00:16:19.920 | So that's what we're working on now.

00:16:24.480 | But for now, like you can actually kind of use this

00:16:28.720 | to some degree without too much effort.

00:16:31.280 | So you can use some code that will look like this.

00:16:37.280 | Okay.

00:16:37.600 | So you'd come here,

00:16:38.720 | you would initialize your knowledge base.

00:16:41.920 | You'd come down to here, right?

00:16:46.880 | So you'd use this sort of code here

00:16:48.720 | in order to add new papers to the database.

00:16:52.560 | Okay.

00:16:52.800 | If we come up to here,

00:16:53.840 | this is where we initialize that, right?

00:16:56.320 | So we have the archive bot

00:16:57.920 | and we initialize it with this paper here.

00:17:00.480 | Okay.

00:17:00.720 | And the extractor is what you can see here.

00:17:05.440 | Okay.

00:17:06.320 | So with that, you can actually use this library.

00:17:11.040 | I'll add like a notebook or something

00:17:14.080 | that will show you how to do that.

00:17:15.680 | But this is still like not ready yet.

00:17:19.520 | Okay.

00:17:19.920 | Because we want to basically make every interaction

00:17:24.160 | with this library accessible via the chat bot.

00:17:27.520 | But for now, you know, that's where we are.

00:17:31.200 | So I'll leave this video.

00:17:33.120 | I hope this has been a useful update.

00:17:36.080 | Hopefully there'll be a bit more progress on this very soon.

00:17:38.880 | And I'll be able to show you like an actual chat bot

00:17:41.680 | that you can genuinely use just like out of the box

00:17:45.680 | as a essentially like research assistant for archive papers

00:17:51.280 | and maybe in the future, other things as well.

00:17:53.280 | But yeah, that's everything for now.

00:17:55.920 | So I hope this was interesting and useful.

00:17:59.440 | Thank you very much for watching

00:18:01.520 | and I will see you again in the next one.

00:18:03.600 | Bye.

00:18:04.640 | (soft music)

00:18:20.560 | (music fades)

00:18:22.560 | you

Chatting with ArXiv Research Papers — AI Assistant #3

Chapters