text
stringclasses 1
value | content
dict |
|---|---|
Transcription
Welcome to the second last meeting. How are you guys feeling so far?
Teddy : Gina, are you doing fine.
No, I'm okay. I'm getting there. I think after this week I should be alright.
Teddy :No, I didn't expect you to finish last week.
Honestly, the parsing bit is the hardest part. I'm so stuck on that bit. I don't know... So I kind of followed what you gave me, right?
Teddy : Yep.
Gina : But I also, like, edited some bits myself. Like, moving to the notions of this line, but the problem is, like, parsing, going through and trying to parse the, like, for example, like, the markdown like that, right?
Teddy : Yep.
Gina : And, like, I don't know if you want it, like, exactly, like, you want things to be, like, separated, or it doesn't matter, like, as long as you can parse it and put it into the... the file? Separate it as in,
Teddy : what do you mean?
Gina : Because you know how you have, so for example, there's a date object in Notion, and then, or you can just parse it as a string. I was thinking, do I need to make it into a date time object? Because Notion has their own date thing, right? And then... For attendees, you can also have it as actual people that are in the Notion workspace.
Teddy : No, but for now, just parse it as string. That's the easiest.
Gina : Well, it's fine now. It's parsing it as a string and putting it into the page. So that's fine. But I just wasn't sure if you needed to separate it into separate... specific block type?
Teddy : Nah, too difficult. We will leave it to the next generation.
Gina : Okay. Well, the next generation is not you.
Teddy : It's not you either. You don't need to be in the team. You can go on.
Gina : I feel like I should... I don't like doing things halfway. I like to see things through.
Teddy : Okay. If you really want to do the block thing, then do it. But for now... string is fine if it's too much then yeah yeah but it goes in like this is like I parse like that file and it just goes in like that but then yeah so it I can make the database as a database yeah it's just a random like That looks nice. Thanks, that's one of my favourite. K-pop?
Gina : No, it's Chinese. Chinese! So this is just like a random page on my thing, right? And there's the database which I made through Python. And then the objects that I just saved in there. These are just empty because I was just trying to test out. But this last one I was able to put through. yeah just as like the blocks myself it's really good it's not like specific but I think I can like still clean it and stuff a little bit but yeah it works it works
Teddy : yeah that's nice um but the only thing is uh I didn't because you know how you gave me
Shreya: all these yeah I feel like what you can do is you already have the categories out like meaning information agenda you can create that as one big list and then create sub list into it where you can add your categories like you're talking about yeah that's fine yeah I think that would be able like I would be able to do that if we needed to but like the um extracting of the info like the markdown into the notion is working basically um but i also i didn't really use the because you know how you gave me like all this like uh fast api stuff yeah i ended up i didn't end up using it oh it's cool it's just like using yeah don't use that just the um notion yeah the notion client yeah yeah Yeah. That's cool. Like, I spent so much time because I thought... Wait, how long did you spend? Oh, not, no, not that much. Please don't be more than 10 hours. No, no, it wasn't more than 10 hours, don't worry. But I spent a lot of time, like, thinking about, like, I forgot that you told me that you just wanted, like, just a page is fine. Because I was thinking, like, oh, do I have to, like, make a date? And then, like, people, and then, like, all the information, like, as, like, a table. I was like, oh, like, you said it was just a page, so it's fine. Because I started doing it right here. I was doing it here and I was like, oh, wait. Oh, shit. That looks pretty neat, so I think that's fine. If you should prove the next... Yeah, maybe for the future, yes. This part, yeah. That's really good. I didn't get to end up doing the grok stuff though yet. Like, implement the grok stuff. Yeah. OK.
Shreya : All right, should I show up my stuff as well? You want me to connect mine as well? You can.
teddy : If this is too much, I can help. That's all right. I should be all right. We can work on it in the next two weeks. It should be fine.
Shreya : How does this thing work? I think first I have to change the display settings. Oh, the screen mirroring. Wait. Do you mind if I? Yeah. So yeah, just drag it like that. Oh, OK. Wait. Why is it not opening? Well, it's not opening. Wait, wait. OK. Oh, yeah.
Teddy : In the meantime. What is that Chinese actress? Is it an actress?
Gina : Yeah, actress.
Teddy : I don't know. But my sister would know. She might know. She loves watching Chinese dramas.
Gina : Oh, I like Chinese dramas as well.
Teddy : I'm surprised it's not K-pop.
Gina : I do like Koreans as well, but the last few years, I've been into Chinese dramas. Yeah, last few years.
Gina : So, how do you, why? Why? I feel like Chinese dramas are so diverse. it's like so interesting i think
Shreya : have you guys seen that korean web series uh when life gives you tangerine uh
Gina : i started watching it and then
Shreya : oh my god it's like i was too scared to continue because i knew i was gonna cry watching it
Shreya : i watched it during my mid-semester break last semester i was crying the whole time it was like so emotional so i'm telling you even if you watch it it's really emotional and i'm like it's it's based on a true story what was that called When life gives you 10 dreams. Oh, wait. It's a really popular one. Okay, I cannot open this. The guy dies 10 times, yeah. And the last time he went against the bad guy. No, no, no. That might be a different drama. This is like a family drama. Not even a drama. It's more of like their life story. I cannot open this file on my Mac anymore. Okay, what file? So it's like I have it open in VS Code. Yep. Where's my cursor? Oh, wait. No. See? Okay. Oh, yeah. Okay. Oh, this is how it works. Okay. So I can't open it on my computer, but I can open it on the monitor. Oh, okay. All right. So I just updated the RAG file a little bit more because I'm still facing a lot of errors. And...
Shreya : I've updated this. This would be like a code walkthrough when we have like the internal meeting with the whole DAPI team. Then instead of obviously like because it's really difficult to cover the code and all of this, then I can just show this. And it basically talks about what the rag file is doing, the components it's using, and we're doing data pre-processing and then embedding, then vector search, evaluation metrics. And I've created a separate file called metrics configuration. So we're adding the retrieval metrics like the K values, the MRR cutoff, And then this is a template where we can just keep on adding more questions if we want. And this is just like, you know, GitHub issues and all this. I'm not so sure about what exactly it's doing, so I still have to check on this. And yeah, there's just like requirements, just like the technical part. And then... I think it's gone to your main. Your main screen. Go to the left mouse. Yeah. Go past it. Okay, you do it. I've never done this before. Yeah. So it's connecting to both screens together. In here. Does that make sense? Okay. So if you go to the right mouse here, you will see the mouse over there. Yeah. And if you go to the left mouse, then you will see the mouse back here. Okay, just open this for me. Okay. Oh, okay, okay. Got it. And talk about this one. Okay. And so this is basically all of the evaluation metrics for RAC file. So the first part basically talks about the evaluation metrics. Remember when we talked about semantic search and the K values in the first meaning? So this is based on that. And then mean reciprocal rank. And it's also here in the file if you see it. MRR cutoff. So I've set it as 10. But ideally, a score of means one means the first result is always the correct one. So we should try to like lower the cutoff, I feel right. And then this is called hit rate, which is like a binary measure of whether at least one relevant document was found within the top K-retreat results. So there's something called, I'm gonna search for it instead. Yeah. Establish a gold standard data set. So this is basically where we would be giving a manually verified, first of all, transcript. So then it knows the format. It knows that Ted is always going to be the manager. Jenna and I are always the team members. A small example. So in the data set, we're basically going to be, it's like a baseline, a rule book on which it's going to work. Like, you know, a book on which RAC can always like refer to and know that this is what we're talking about. This is our relevant documents of like who's there in the team, who's the team lead, what projects we're working on. And if we have any other notes side by side, you know, anything like that, or we can try connecting it with Discord or something like that. And so then basically we can then retrieve all of those things from here. From the gold standard data set, it's going to retrieve the hit rate, the reciprocal rank. just to ensure that one relevant document was found when we got the meeting transcript. And then obviously the context should be relevant. And then this is to evaluate the generation it's doing, how accurate it is and how reliable it is. So there's something called faithfulness, which is basically to see if it has generated any sort of false information or just random stuff on its own, like how ChatGP does it sometimes. Then answer correctness, answer relevancy. And then this is basically all the technical things that we have to keep in mind. Like the goals turned data set. Then automated scoring. And obviously there has to be a human in loop at all times. And then establish automated testing pipelines. So are we gonna do this in the hugging phase thing? Or is that gonna be different? we're going to do this with the hugging face and other stuff. Yeah, we can implement this in the queries that we were talking about, positive retrieval and negative retrieval. Okay. And then this is just like more on a theory base now on what practices we can do to make sure that we're following the steps for accurate metric, which is faithfulness, answer correctness, the context relevancy and tailor specific domains like how I was talking about in the first meeting that it could be possible that a different team even though they might be in the technical arena but they could still be talking about something completely different. So it has to be tailored to their specific domain also. And now these are the possible test cases that we have to deal with in our code, which is basically like whether, let's say the chunks of information it's retrieving, is it losing the context or is it too small, too large, or it's just too much of irrelevant information? And then weak re-ranker, which is basically like You know, the, like, RAG file, it might lose certain key points that we discussed during the meeting, so it might not include that. And what metrics we have to deal with if we want to improve it. Okay, and then the uniform uniformity of embedding which is basically like when we get the final transcript is a uniform is it consistent is it domain aware all of those things, and this is again regarding faithfulness and then ignoring context, so we have to deal with these test cases on our code itself. And I've like added links also from where exactly I got it. Just give it a minute. So see wrap pipelines and like exactly how we can make sure. So I think we have to like work on these things also. So yeah, and then this is just like a real world application and then all the sources I've used. So what I've done is from part time to time, like you can just open any of these tools and see how, like this is from Geeks for Geeks. So how we can maintain these things and it's got like a sample code also. So yeah, this is what I've worked on so far. All right, over to you. And it's almost done.
Teddy : And what Shreya is working on is uploading to Hugging Face. And she will do this later. And you follow it in this so that it improves the time. OK? Make sense? Yeah. OK. Right now, I'm working on getting feedback from Notion. Right, so we got chunks from all the queries here, which used to generate the meeting minutes. And what if the chunks are not correct? Like Shreya said, it's not accurate. So we need to make a data set to fine tune this. system right and what we need is but how do we know if this is incorrect right so we need to get feedback right yeah and what i've done so far is that see this is a simple Notion page. And we got a transcript here. And we got error in here. Student D doing API performance while the student C doing it. Right. Oh, this is a bad example, but it's an example. So you see the button here? You can click on it, and by linking We can get the retrieved issue. What is the query of that part? What's the wrong retrieval, which is the current retrieval? And what is the correct retrieval? Right. And then this will send to a hugging face with the pipeline to create a data set. Yeah. So what I need is when you are generating the Notion page, I want the ID of a specific block. So maybe we should add a section like chunk sending and buttons.
Gina : So how are you planning to get the feedback?
Teddy : Yeah, this part is too hard. No, no, no. That's the manual.
Manually, right? No, that is not. So when we're making the block, we don't have this part. So we just have maybe a section saying. Yeah, that section that chunks to send bit. But then you have the one that you want to send, and then you manually have to send it by pressing the button. Is that what you mean? Yes. So let's say you see, oh, this one is wrong, and you said, I want this chunk to be the correct one. Put the incorrect one in here, and then you send it by clicking the button. So that would be done manually, because you have to click the button. And what I need is the ID of this block. So when you are generating the page in Notion, you will have the block, and I need the ID. And also, what I need from Shreya is what kind of chunks I need to send it into the database.
Shreya : For the feedback, right?
Teddy : For the fine-tuning.
Shreya : For the fine-tuning.
Teddy : Yeah, for the fine-tuning. And yeah, another thing. You don't need to look into the code. I just need the information. Oh. Yeah. Do you mean like the information in? No. The information is in the chunk. So this block, it will have an ID. And it will have an ID, yeah. And I want the ID. So you just need to extract that ID from there. But I still want to know like for feedback, what questions are you planning to ask the members to rate how accurate the Yeah, that is also a problem. Because as you can see here, the template is quite a bit messy. And I still haven't figured out how to add in that chunk to sentence and buttons inside that template so that people can give feedback. Adding that chunk at the bottom. At the bottom. But which one is referring to? I mean we can add more text in here so that people can better identify. What is that any function is like? Do you mean like for the user to differentiate? Yeah.
Gina : Do you mean the user or for the code to know?
Teddy : both because if you add in at the end then the code will never know which chunks is referring like which section is referring to and also for the user if you just accept chunks to send it's not descriptive enough yeah exactly so what i'm facing the issue is what how to design that section so that users can specify which sections they want to change.
Shreya : What we can do is we can add the link of a Google Form. We can make a Google Form where we can ask them questions. Because on Google Form, we can add drop down questions, we can add MCQ, stuff like that. Rate how accurate you found the transcript, what improvements we can do. So we give options for everything. Because if you just add everything on Notion, there is a possibility it could get a little bit cluttered. So I would say first get it on Google Form. So then we also have the IDs of each and every person, right? Because when you fill out a Google Form, you have to do it with your Google ID and stuff, right? And then when we have done it and we feel that, okay, this is what we're getting, then we can try to merge it with Notion. Because if you do it on Notion, it's like even if it messes up, you know, it's going to be like all over Notion and it might create a hurdle for like other things as well. In case if there's a problem in the code or if it doesn't fetch things correctly or if it messes up the ID or stuff like that. Right. Because in Notion, sometimes it's like it also gives you the option to go and make edits for somebody else. Like I can open your profile on Notion and change your profile picture for you. Right. So I mean, just to like avoid all of those things, we can first try it on like Google Form, where we know that the person who is taking the Google Form, like if Gina is giving me feedback on the transcript we got, then it is from Gina and, you know, we can say that it creates more like privacy and everything.
Teddy : I see. What do you think?
Gina : I'm kind of confused. Isn't the point of this one for like... So this is the meeting minutes that we designed. Or like, not designed, that's outputted. And then this part here is the... So you're reading it and then you notice that there's an incorrect output. Because it says student D but it should be student C. So you want to tell... You want to send that bit... to tell like whatever here is that it's incorrect and they need to like yeah because right now the chunks we're sending it's just like even if we add a few more buttons to it and even if we give like we can't add questions to it right like because we're just sending the chunk we only want the chunk that's incorrect right okay so let's say it's incorrect the manager can change it themselves it doesn't have to change through the code But the process is not robust enough, meaning, right?
Gina : You mean like currently, like in Notion itself at the moment?
Teddy : What do you mean induction?
Gina : Do you mean like this process at the moment?
Teddy : No, I mean the RAG system is not robust enough. Oh, okay, yeah. Yeah, because the RAG system provides the answer. Oh, yeah, because it's giving you the incorrect answer, yeah. Yes, so what we want to do is to make a data set to improve the rack system.
Gina : Mm-hmm.
Teddy : So that when you click in that button, it will send that into FastAPI. And we will get the query wrong chunks and correct chunks. Wrong chunks we know because it's the chunks that used to generate that. But correct chunks we don't know. So we also need the manager to give us that information.
Gina : Yeah, so that's why you're sending this, like sending the correct, what should have been outputted.
Teddy : And the problem is, how are we going to design that so that But the system knows it's that session that is wrong and also I think just for like if you want for now you can just add a few questions to it and like I said you know that's what I'm saying on Notion we are bounded We can't add a lot of diversity into the feedback. But I feel like if on Google Form, you know how we take polls and stuff as well on Google Form? So it generates bar graphs for you also to tell you what percentage of people said. We can get it. Let's say if we hand this out. like this DAPI know to other DAPI teams and they attempted and they're like yeah you know 40% people found the meeting transcript to be accurate 20% people said that it got the ID wrong you know first we can start asking like basic questions because we were at the initial stage right so we need to ask the basic questions like did it get the topic right did it get the student right did it get the notes right we have to ask specific questions also Right. I see what you... Yeah. Because until the time we don't try it out, like the hidden trial method. If you do it through Google Forms, then how are you going to send that? Like, for... Because you want to send this to... It doesn't have to be in FastAPI. That is just triggering a signal into the backend so that we can store the information. Yeah, I think she's trying to say how are we going to link Google Form with our code or Notion and stuff. That's the other thing. I think this is fine.
Gina :But then if we add the Google Docs and stuff, you have to research more about how to link it all together. Or we can try to look for some other tool that is already doing that type of work. We can make a form here so that they can do it within Notion. That also works. Oh, they have forms. I don't know. I don't need to know. Same. That's why. We can just make a block that has all the questions and the response we want and make a button so that we can extract that in one go. That also works.
Teddy L It's just not aesthetic and elegant, I guess. But who cares? I think the notino is to... I'm fine with that. That's fine. It's just like the way Notion is developed, it just kind of makes it difficult.
Gina : Yeah, you're right. You're right.
Shreya : Okay, so it turns out we can connect Google Forms to Notion.
Teddy : How?
Shreya : Yep. Oh, so it's like... So it's like you take the data from Google Form and you connect it to Notion and it generates the same thing for you. And then you collect data inside Notion. But I want to, because now you are having Google Form to Notion, right?
Shreya: Yeah, it's saying you can just do it directly. It's called Google Form Add-on.
Teddy : But what I want is that Notion links to Google Form and Google Form back to the backend. No, we can do Google Form to Notion and then Notion to backend. Because we would anyway, like DAPI is going to be inbuilt in Notion, right? So Notion is our primary interface. Google Form is not the primary interface. I'm just suggesting Google Form because it has other features also, but Notion doesn't have that. Notion is mainly for making notes and structuring and organizing everything. So we get the data from Google Form to Notion, and then at Notion we can do whatever we want. We can get it in backend, we can keep updating DAPI, all of those things. How are you going to trigger that through Notion? Yeah, that is the main thing. I mean, this is a good situation. Either we connect it with our gold standard data set where we keep updating the documents and they have to be manually verified, like I researched in the metric thing. Because I still feel like I've not used Notion that much, but I still don't think Notion is that accurate when it comes to feedback, like the feedback that we want.
Teddy : I'll think about it, and also I will make...
Shreya : Yeah, you can also have a look at this, right, and see what we can do. Do you want to copy the link and just put it in the...
Teddy : Yeah, okay. I'll have a look at it. Just copy the link. All right. OK. Should be fine then. Is it too much work?
Shreya : No, it's OK. So Gina will finish that. And on Hugging Face, I have to do how the RAG file is going to get positive and negative retrievals. Yeah, and also what kind of positive and negative retrieval you're getting. Right, OK.
Shreya : So before you came, we were trying out Clue.ly, and turns out it is doing exactly what we're doing. It's giving you the transcripts. It can also give me individual transcripts, like what each one of us is talking about. It can also give you follow-up questions and everything. So I'm thinking we'll get today's meeting and the past meeting, because we only have two recordings so far, get their transcript and have it laid out. Because the RAC file that I've made, I've not even once tested out how it's generating the transcript, right? Because the RAC file that I have, like, if I give it, like, the main thing that we're trying to do, how am I going to test out my RAG file? Like, I've made it, but how are we going to test out the implementation it's doing? So for that, we need a transcript, right? So... How am I going to connect it with the data pipeline? That's what I'm wondering.
Teddy : Oh, you're not connecting that. I will do the connection. Okay. Yeah.
Shreya : Okay. So I'll just make the data pipeline thing. Yeah. Give you the rag file, the data, and everything, and you're going to connect everything. Yeah. Okay. And you're going to deal with the errors? That's cool. Deal with the errors. Because this file is still showing a lot of errors. I don't know why. So I still have to deal with it. And it's just showing, like, stupid, like, module error and, like, import errors for some reason. Oh. I don't know why.
Gina : Yeah. So that's why. Well, I fixed one, so my imports are good. Oh, because she's not using my requirement.txt. Yeah, no, I edited it. Yeah. Yeah, because my PC is not working. I just removed some of the, what is it? I forgot what I removed. Tendency. Yeah, just a few of them. We can try out the requirement.txt if you want. That should be all of them. All of the libraries that we need? Yes, and you need to remove some of them though. Yeah, you told me there's like errors in the file. Oh, okay. Hugging face, got it. Yeah, that's it. Okay.
Teddy : Then, next week we actually need a proper demo. Right. So when are we planning to have the meeting? Meeting? You mean the... Yeah, the DAPI meeting. Is it decided yet? No. It's not decided yet. Okay. Yeah, should we have some... J, J, J. Yeah. Is he back from Miami yet? Or no? Oh, no, not J. It's J. Oh, J. Okay. J is always good. Did he say week nine? No, week 10. Oh, week 10. So that's after the break. That's right. Yeah, yeah, yeah. I believe it's week 10. Yep, it's after the break. Oh, that's when we have our showcase. Let me ask him. He will be here soon. If it's going to be week 10, there's nothing to showcase at that point. If we haven't gone through it all together. Then do you want to make the deadline next week? Yeah, I think we should set our internal demo without the DAPI team to have everything ready by next week. Yeah, can you guys do it? Yeah. No, you don't need to force yourself. I can help. Just ask for help, okay? Just ask for it. So stubborn. Yeah. I want to take the whole credit, you see. No, no, no. I'm just kidding. Yeah, I'll make... I'll figure it out next week. And don't work more than 10 hours. I don't, I don't. You don't? I don't. Okay, that's good. The dream is to make your work look like 10 hours, but you only spend like 5 hours.
Gina : That's like a corporate life.
Shreya : Yeah, that's how you want to do it. Like you just get your AI agent to do the whole thing for you, but they're like, oh, I'm still doing it, I'm still doing it. Okay, you're sure about that? Next week, deadline. Okay, so next week, let's keep two deadlines. So first deadline is just basically, like you said, you're going to do the whole integration, right? Yep. So you also need time. Because when you're going to integrate it, it's going to take you time.
Shreya : So first deadline is to just complete the whole code from my end, from Gina end, and we're going to hand over everything to you. And then you're going to sit down and do the integration. And then obviously we will have to sit down and fix a lot of other things. And in the meantime, we can work on the feedback thing for the notion and figure out what we're going to do. And then how we're going to connect with the backend, right? So then it works both ways. And then the second deadline of the week is the demo part, where we're done with the integration. And if we need to do it at the showcase, we would be able to do it. Let's keep it like that.
Teddy OK. Then next meeting will be on Friday. I will expect you guys to finish your features.
|
{
"action_items": [
{
"assigned to": "Gina",
"deadline": "Next week",
"status": "Not Started",
"task description": "Implement parsing of markdown into Notion"
},
{
"assigned to": "Teddy",
"deadline": "Next week",
"status": "Not Started",
"task description": "Create a data set to fine-tune the RAG system"
},
{
"assigned to": "Teddy",
"deadline": "Next week",
"status": "Not Started",
"task description": "Design a section for users to provide feedback on incorrect chunks"
},
{
"assigned to": "Shreya",
"deadline": "Next week",
"status": "Not Started",
"task description": "Research and implement a method to connect Google Forms to Notion"
},
{
"assigned to": "Shreya",
"deadline": "Next week",
"status": "In Progress",
"task description": "Complete the RAG file and data pipeline"
},
{
"assigned to": "Shreya",
"deadline": "Next week",
"status": "In Progress",
"task description": "Fix errors in the RAG file and implement the data pipeline"
},
{
"assigned to": "Teddy",
"deadline": "Next week",
"status": "Not Started",
"task description": "Integrate the code and complete the demo"
},
{
"assigned to": "All",
"deadline": "Next week",
"status": "Not Started",
"task description": "Prepare for the internal demo and showcase"
}
],
"agenda": [
{
"notes": "Gina discussed the challenges of parsing and integrating with Notion, including deciding how to handle date and time objects.",
"presenter": "Gina",
"time allocated": "",
"topic": "Parsing and Notion Integration"
},
{
"notes": "Shreya presented her work on the RAG file, including evaluation metrics such as faithfulness, answer correctness, and context relevancy.",
"presenter": "Shreya",
"time allocated": "",
"topic": "RAG File and Evaluation Metrics"
},
{
"notes": "Teddy discussed the need for a feedback mechanism to improve the RAG system and the challenges of integrating it with Notion, including designing a user-friendly interface for providing feedback.",
"presenter": "Teddy",
"time allocated": "",
"topic": "Feedback Mechanism and Notion Integration"
}
],
"discussion_summary": [
{
"decisions": [
"The team decided to parse the markdown as a string for now, rather than trying to separate it into specific block types.",
"The team decided to use Google Forms to collect feedback and connect it to Notion.",
"The team set a deadline for the next week to complete their features and have a demo ready."
],
"issues": [
"The team is still having trouble with the parsing of markdown and the integration of Notion and Hugging Face.",
"The team needs to create a gold standard data set to fine-tune the RAG system and improve its accuracy.",
"The team is having trouble connecting Google Forms to Notion and getting feedback from users."
],
"key_points": [
"The team discussed the current status of their project, including the parsing of markdown and the integration of Notion and Hugging Face.",
"Gina mentioned that she had trouble with the parsing bit and was stuck on it, but Teddy suggested just parsing it as a string for now.",
"Shreya talked about her work on the RAG file and the evaluation metrics, including the mean reciprocal rank and hit rate.",
"The team discussed the need for a gold standard data set to fine-tune the RAG system and improve its accuracy.",
"Teddy mentioned that he is working on getting feedback from Notion and creating a data set to improve the RAG system.",
"The team discussed the possibility of using Google Forms to collect feedback and connect it to Notion.",
"Shreya mentioned that she found a way to connect Google Forms to Notion using an add-on.",
"The team set a deadline for the next week to complete their features and have a demo ready."
]
}
]
}
|
Transcription
Welcome to the second last meeting. How are you guys feeling so far?
Teddy : Gina, are you doing fine.
No, I'm okay. I'm getting there. I think after this week I should be alright.
Teddy :No, I didn't expect you to finish last week.
Honestly, the parsing bit is the hardest part. I'm so stuck on that bit. I don't know... So I kind of followed what you gave me, right?
Teddy : Yep.
Gina : But I also, like, edited some bits myself. Like, moving to the notions of this line, but the problem is, like, parsing, going through and trying to parse the, like, for example, like, the markdown like that, right?
Teddy : Yep.
Gina : And, like, I don't know if you want it, like, exactly, like, you want things to be, like, separated, or it doesn't matter, like, as long as you can parse it and put it into the... the file? Separate it as in,
Teddy : what do you mean?
Gina : Because you know how you have, so for example, there's a date object in Notion, and then, or you can just parse it as a string. I was thinking, do I need to make it into a date time object? Because Notion has their own date thing, right? And then... For attendees, you can also have it as actual people that are in the Notion workspace.
Teddy : No, but for now, just parse it as string. That's the easiest.
Gina : Well, it's fine now. It's parsing it as a string and putting it into the page. So that's fine. But I just wasn't sure if you needed to separate it into separate... specific block type?
Teddy : Nah, too difficult. We will leave it to the next generation.
Gina : Okay. Well, the next generation is not you.
Teddy : It's not you either. You don't need to be in the team. You can go on.
Gina : I feel like I should... I don't like doing things halfway. I like to see things through.
Teddy : Okay. If you really want to do the block thing, then do it. But for now... string is fine if it's too much then yeah yeah but it goes in like this is like I parse like that file and it just goes in like that but then yeah so it I can make the database as a database yeah it's just a random like That looks nice. Thanks, that's one of my favourite. K-pop?
Gina : No, it's Chinese. Chinese! So this is just like a random page on my thing, right? And there's the database which I made through Python. And then the objects that I just saved in there. These are just empty because I was just trying to test out. But this last one I was able to put through. yeah just as like the blocks myself it's really good it's not like specific but I think I can like still clean it and stuff a little bit but yeah it works it works
Teddy : yeah that's nice um but the only thing is uh I didn't because you know how you gave me
Shreya: all these yeah I feel like what you can do is you already have the categories out like meaning information agenda you can create that as one big list and then create sub list into it where you can add your categories like you're talking about yeah that's fine yeah I think that would be able like I would be able to do that if we needed to but like the um extracting of the info like the markdown into the notion is working basically um but i also i didn't really use the because you know how you gave me like all this like uh fast api stuff yeah i ended up i didn't end up using it oh it's cool it's just like using yeah don't use that just the um notion yeah the notion client yeah yeah Yeah. That's cool. Like, I spent so much time because I thought... Wait, how long did you spend? Oh, not, no, not that much. Please don't be more than 10 hours. No, no, it wasn't more than 10 hours, don't worry. But I spent a lot of time, like, thinking about, like, I forgot that you told me that you just wanted, like, just a page is fine. Because I was thinking, like, oh, do I have to, like, make a date? And then, like, people, and then, like, all the information, like, as, like, a table. I was like, oh, like, you said it was just a page, so it's fine. Because I started doing it right here. I was doing it here and I was like, oh, wait. Oh, shit. That looks pretty neat, so I think that's fine. If you should prove the next... Yeah, maybe for the future, yes. This part, yeah. That's really good. I didn't get to end up doing the grok stuff though yet. Like, implement the grok stuff. Yeah. OK.
Shreya : All right, should I show up my stuff as well? You want me to connect mine as well? You can.
teddy : If this is too much, I can help. That's all right. I should be all right. We can work on it in the next two weeks. It should be fine.
Shreya : How does this thing work? I think first I have to change the display settings. Oh, the screen mirroring. Wait. Do you mind if I? Yeah. So yeah, just drag it like that. Oh, OK. Wait. Why is it not opening? Well, it's not opening. Wait, wait. OK. Oh, yeah.
Teddy : In the meantime. What is that Chinese actress? Is it an actress?
Gina : Yeah, actress.
Teddy : I don't know. But my sister would know. She might know. She loves watching Chinese dramas.
Gina : Oh, I like Chinese dramas as well.
Teddy : I'm surprised it's not K-pop.
Gina : I do like Koreans as well, but the last few years, I've been into Chinese dramas. Yeah, last few years.
Gina : So, how do you, why? Why? I feel like Chinese dramas are so diverse. it's like so interesting i think
Shreya : have you guys seen that korean web series uh when life gives you tangerine uh
Gina : i started watching it and then
Shreya : oh my god it's like i was too scared to continue because i knew i was gonna cry watching it
Shreya : i watched it during my mid-semester break last semester i was crying the whole time it was like so emotional so i'm telling you even if you watch it it's really emotional and i'm like it's it's based on a true story what was that called When life gives you 10 dreams. Oh, wait. It's a really popular one. Okay, I cannot open this. The guy dies 10 times, yeah. And the last time he went against the bad guy. No, no, no. That might be a different drama. This is like a family drama. Not even a drama. It's more of like their life story. I cannot open this file on my Mac anymore. Okay, what file? So it's like I have it open in VS Code. Yep. Where's my cursor? Oh, wait. No. See? Okay. Oh, yeah. Okay. Oh, this is how it works. Okay. So I can't open it on my computer, but I can open it on the monitor. Oh, okay. All right. So I just updated the RAG file a little bit more because I'm still facing a lot of errors. And...
Shreya : I've updated this. This would be like a code walkthrough when we have like the internal meeting with the whole DAPI team. Then instead of obviously like because it's really difficult to cover the code and all of this, then I can just show this. And it basically talks about what the rag file is doing, the components it's using, and we're doing data pre-processing and then embedding, then vector search, evaluation metrics. And I've created a separate file called metrics configuration. So we're adding the retrieval metrics like the K values, the MRR cutoff, And then this is a template where we can just keep on adding more questions if we want. And this is just like, you know, GitHub issues and all this. I'm not so sure about what exactly it's doing, so I still have to check on this. And yeah, there's just like requirements, just like the technical part. And then... I think it's gone to your main. Your main screen. Go to the left mouse. Yeah. Go past it. Okay, you do it. I've never done this before. Yeah. So it's connecting to both screens together. In here. Does that make sense? Okay. So if you go to the right mouse here, you will see the mouse over there. Yeah. And if you go to the left mouse, then you will see the mouse back here. Okay, just open this for me. Okay. Oh, okay, okay. Got it. And talk about this one. Okay. And so this is basically all of the evaluation metrics for RAC file. So the first part basically talks about the evaluation metrics. Remember when we talked about semantic search and the K values in the first meaning? So this is based on that. And then mean reciprocal rank. And it's also here in the file if you see it. MRR cutoff. So I've set it as 10. But ideally, a score of means one means the first result is always the correct one. So we should try to like lower the cutoff, I feel right. And then this is called hit rate, which is like a binary measure of whether at least one relevant document was found within the top K-retreat results. So there's something called, I'm gonna search for it instead. Yeah. Establish a gold standard data set. So this is basically where we would be giving a manually verified, first of all, transcript. So then it knows the format. It knows that Ted is always going to be the manager. Jenna and I are always the team members. A small example. So in the data set, we're basically going to be, it's like a baseline, a rule book on which it's going to work. Like, you know, a book on which RAC can always like refer to and know that this is what we're talking about. This is our relevant documents of like who's there in the team, who's the team lead, what projects we're working on. And if we have any other notes side by side, you know, anything like that, or we can try connecting it with Discord or something like that. And so then basically we can then retrieve all of those things from here. From the gold standard data set, it's going to retrieve the hit rate, the reciprocal rank. just to ensure that one relevant document was found when we got the meeting transcript. And then obviously the context should be relevant. And then this is to evaluate the generation it's doing, how accurate it is and how reliable it is. So there's something called faithfulness, which is basically to see if it has generated any sort of false information or just random stuff on its own, like how ChatGP does it sometimes. Then answer correctness, answer relevancy. And then this is basically all the technical things that we have to keep in mind. Like the goals turned data set. Then automated scoring. And obviously there has to be a human in loop at all times. And then establish automated testing pipelines. So are we gonna do this in the hugging phase thing? Or is that gonna be different? we're going to do this with the hugging face and other stuff. Yeah, we can implement this in the queries that we were talking about, positive retrieval and negative retrieval. Okay. And then this is just like more on a theory base now on what practices we can do to make sure that we're following the steps for accurate metric, which is faithfulness, answer correctness, the context relevancy and tailor specific domains like how I was talking about in the first meeting that it could be possible that a different team even though they might be in the technical arena but they could still be talking about something completely different. So it has to be tailored to their specific domain also. And now these are the possible test cases that we have to deal with in our code, which is basically like whether, let's say the chunks of information it's retrieving, is it losing the context or is it too small, too large, or it's just too much of irrelevant information? And then weak re-ranker, which is basically like You know, the, like, RAG file, it might lose certain key points that we discussed during the meeting, so it might not include that. And what metrics we have to deal with if we want to improve it. Okay, and then the uniform uniformity of embedding which is basically like when we get the final transcript is a uniform is it consistent is it domain aware all of those things, and this is again regarding faithfulness and then ignoring context, so we have to deal with these test cases on our code itself. And I've like added links also from where exactly I got it. Just give it a minute. So see wrap pipelines and like exactly how we can make sure. So I think we have to like work on these things also. So yeah, and then this is just like a real world application and then all the sources I've used. So what I've done is from part time to time, like you can just open any of these tools and see how, like this is from Geeks for Geeks. So how we can maintain these things and it's got like a sample code also. So yeah, this is what I've worked on so far. All right, over to you. And it's almost done.
Teddy : And what Shreya is working on is uploading to Hugging Face. And she will do this later. And you follow it in this so that it improves the time. OK? Make sense? Yeah. OK. Right now, I'm working on getting feedback from Notion. Right, so we got chunks from all the queries here, which used to generate the meeting minutes. And what if the chunks are not correct? Like Shreya said, it's not accurate. So we need to make a data set to fine tune this. system right and what we need is but how do we know if this is incorrect right so we need to get feedback right yeah and what i've done so far is that see this is a simple Notion page. And we got a transcript here. And we got error in here. Student D doing API performance while the student C doing it. Right. Oh, this is a bad example, but it's an example. So you see the button here? You can click on it, and by linking We can get the retrieved issue. What is the query of that part? What's the wrong retrieval, which is the current retrieval? And what is the correct retrieval? Right. And then this will send to a hugging face with the pipeline to create a data set. Yeah. So what I need is when you are generating the Notion page, I want the ID of a specific block. So maybe we should add a section like chunk sending and buttons.
Gina : So how are you planning to get the feedback?
Teddy : Yeah, this part is too hard. No, no, no. That's the manual.
Manually, right? No, that is not. So when we're making the block, we don't have this part. So we just have maybe a section saying. Yeah, that section that chunks to send bit. But then you have the one that you want to send, and then you manually have to send it by pressing the button. Is that what you mean? Yes. So let's say you see, oh, this one is wrong, and you said, I want this chunk to be the correct one. Put the incorrect one in here, and then you send it by clicking the button. So that would be done manually, because you have to click the button. And what I need is the ID of this block. So when you are generating the page in Notion, you will have the block, and I need the ID. And also, what I need from Shreya is what kind of chunks I need to send it into the database.
Shreya : For the feedback, right?
Teddy : For the fine-tuning.
Shreya : For the fine-tuning.
Teddy : Yeah, for the fine-tuning. And yeah, another thing. You don't need to look into the code. I just need the information. Oh. Yeah. Do you mean like the information in? No. The information is in the chunk. So this block, it will have an ID. And it will have an ID, yeah. And I want the ID. So you just need to extract that ID from there. But I still want to know like for feedback, what questions are you planning to ask the members to rate how accurate the Yeah, that is also a problem. Because as you can see here, the template is quite a bit messy. And I still haven't figured out how to add in that chunk to sentence and buttons inside that template so that people can give feedback. Adding that chunk at the bottom. At the bottom. But which one is referring to? I mean we can add more text in here so that people can better identify. What is that any function is like? Do you mean like for the user to differentiate? Yeah.
Gina : Do you mean the user or for the code to know?
Teddy : both because if you add in at the end then the code will never know which chunks is referring like which section is referring to and also for the user if you just accept chunks to send it's not descriptive enough yeah exactly so what i'm facing the issue is what how to design that section so that users can specify which sections they want to change.
Shreya : What we can do is we can add the link of a Google Form. We can make a Google Form where we can ask them questions. Because on Google Form, we can add drop down questions, we can add MCQ, stuff like that. Rate how accurate you found the transcript, what improvements we can do. So we give options for everything. Because if you just add everything on Notion, there is a possibility it could get a little bit cluttered. So I would say first get it on Google Form. So then we also have the IDs of each and every person, right? Because when you fill out a Google Form, you have to do it with your Google ID and stuff, right? And then when we have done it and we feel that, okay, this is what we're getting, then we can try to merge it with Notion. Because if you do it on Notion, it's like even if it messes up, you know, it's going to be like all over Notion and it might create a hurdle for like other things as well. In case if there's a problem in the code or if it doesn't fetch things correctly or if it messes up the ID or stuff like that. Right. Because in Notion, sometimes it's like it also gives you the option to go and make edits for somebody else. Like I can open your profile on Notion and change your profile picture for you. Right. So I mean, just to like avoid all of those things, we can first try it on like Google Form, where we know that the person who is taking the Google Form, like if Gina is giving me feedback on the transcript we got, then it is from Gina and, you know, we can say that it creates more like privacy and everything.
Teddy : I see. What do you think?
Gina : I'm kind of confused. Isn't the point of this one for like... So this is the meeting minutes that we designed. Or like, not designed, that's outputted. And then this part here is the... So you're reading it and then you notice that there's an incorrect output. Because it says student D but it should be student C. So you want to tell... You want to send that bit... to tell like whatever here is that it's incorrect and they need to like yeah because right now the chunks we're sending it's just like even if we add a few more buttons to it and even if we give like we can't add questions to it right like because we're just sending the chunk we only want the chunk that's incorrect right okay so let's say it's incorrect the manager can change it themselves it doesn't have to change through the code But the process is not robust enough, meaning, right?
Gina : You mean like currently, like in Notion itself at the moment?
Teddy : What do you mean induction?
Gina : Do you mean like this process at the moment?
Teddy : No, I mean the RAG system is not robust enough. Oh, okay, yeah. Yeah, because the RAG system provides the answer. Oh, yeah, because it's giving you the incorrect answer, yeah. Yes, so what we want to do is to make a data set to improve the rack system.
Gina : Mm-hmm.
Teddy : So that when you click in that button, it will send that into FastAPI. And we will get the query wrong chunks and correct chunks. Wrong chunks we know because it's the chunks that used to generate that. But correct chunks we don't know. So we also need the manager to give us that information.
Gina : Yeah, so that's why you're sending this, like sending the correct, what should have been outputted.
Teddy : And the problem is, how are we going to design that so that But the system knows it's that session that is wrong and also I think just for like if you want for now you can just add a few questions to it and like I said you know that's what I'm saying on Notion we are bounded We can't add a lot of diversity into the feedback. But I feel like if on Google Form, you know how we take polls and stuff as well on Google Form? So it generates bar graphs for you also to tell you what percentage of people said. We can get it. Let's say if we hand this out. like this DAPI know to other DAPI teams and they attempted and they're like yeah you know 40% people found the meeting transcript to be accurate 20% people said that it got the ID wrong you know first we can start asking like basic questions because we were at the initial stage right so we need to ask the basic questions like did it get the topic right did it get the student right did it get the notes right we have to ask specific questions also Right. I see what you... Yeah. Because until the time we don't try it out, like the hidden trial method. If you do it through Google Forms, then how are you going to send that? Like, for... Because you want to send this to... It doesn't have to be in FastAPI. That is just triggering a signal into the backend so that we can store the information. Yeah, I think she's trying to say how are we going to link Google Form with our code or Notion and stuff. That's the other thing. I think this is fine.
Gina :But then if we add the Google Docs and stuff, you have to research more about how to link it all together. Or we can try to look for some other tool that is already doing that type of work. We can make a form here so that they can do it within Notion. That also works. Oh, they have forms. I don't know. I don't need to know. Same. That's why. We can just make a block that has all the questions and the response we want and make a button so that we can extract that in one go. That also works.
Teddy L It's just not aesthetic and elegant, I guess. But who cares? I think the notino is to... I'm fine with that. That's fine. It's just like the way Notion is developed, it just kind of makes it difficult.
Gina : Yeah, you're right. You're right.
Shreya : Okay, so it turns out we can connect Google Forms to Notion.
Teddy : How?
Shreya : Yep. Oh, so it's like... So it's like you take the data from Google Form and you connect it to Notion and it generates the same thing for you. And then you collect data inside Notion. But I want to, because now you are having Google Form to Notion, right?
Shreya: Yeah, it's saying you can just do it directly. It's called Google Form Add-on.
Teddy : But what I want is that Notion links to Google Form and Google Form back to the backend. No, we can do Google Form to Notion and then Notion to backend. Because we would anyway, like DAPI is going to be inbuilt in Notion, right? So Notion is our primary interface. Google Form is not the primary interface. I'm just suggesting Google Form because it has other features also, but Notion doesn't have that. Notion is mainly for making notes and structuring and organizing everything. So we get the data from Google Form to Notion, and then at Notion we can do whatever we want. We can get it in backend, we can keep updating DAPI, all of those things. How are you going to trigger that through Notion? Yeah, that is the main thing. I mean, this is a good situation. Either we connect it with our gold standard data set where we keep updating the documents and they have to be manually verified, like I researched in the metric thing. Because I still feel like I've not used Notion that much, but I still don't think Notion is that accurate when it comes to feedback, like the feedback that we want.
Teddy : I'll think about it, and also I will make...
Shreya : Yeah, you can also have a look at this, right, and see what we can do. Do you want to copy the link and just put it in the...
Teddy : Yeah, okay. I'll have a look at it. Just copy the link. All right. OK. Should be fine then. Is it too much work?
Shreya : No, it's OK. So Gina will finish that. And on Hugging Face, I have to do how the RAG file is going to get positive and negative retrievals. Yeah, and also what kind of positive and negative retrieval you're getting. Right, OK.
Shreya : So before you came, we were trying out Clue.ly, and turns out it is doing exactly what we're doing. It's giving you the transcripts. It can also give me individual transcripts, like what each one of us is talking about. It can also give you follow-up questions and everything. So I'm thinking we'll get today's meeting and the past meeting, because we only have two recordings so far, get their transcript and have it laid out. Because the RAC file that I've made, I've not even once tested out how it's generating the transcript, right? Because the RAC file that I have, like, if I give it, like, the main thing that we're trying to do, how am I going to test out my RAG file? Like, I've made it, but how are we going to test out the implementation it's doing? So for that, we need a transcript, right? So... How am I going to connect it with the data pipeline? That's what I'm wondering.
Teddy : Oh, you're not connecting that. I will do the connection. Okay. Yeah.
Shreya : Okay. So I'll just make the data pipeline thing. Yeah. Give you the rag file, the data, and everything, and you're going to connect everything. Yeah. Okay. And you're going to deal with the errors? That's cool. Deal with the errors. Because this file is still showing a lot of errors. I don't know why. So I still have to deal with it. And it's just showing, like, stupid, like, module error and, like, import errors for some reason. Oh. I don't know why.
Gina : Yeah. So that's why. Well, I fixed one, so my imports are good. Oh, because she's not using my requirement.txt. Yeah, no, I edited it. Yeah. Yeah, because my PC is not working. I just removed some of the, what is it? I forgot what I removed. Tendency. Yeah, just a few of them. We can try out the requirement.txt if you want. That should be all of them. All of the libraries that we need? Yes, and you need to remove some of them though. Yeah, you told me there's like errors in the file. Oh, okay. Hugging face, got it. Yeah, that's it. Okay.
Teddy : Then, next week we actually need a proper demo. Right. So when are we planning to have the meeting? Meeting? You mean the... Yeah, the DAPI meeting. Is it decided yet? No. It's not decided yet. Okay. Yeah, should we have some... J, J, J. Yeah. Is he back from Miami yet? Or no? Oh, no, not J. It's J. Oh, J. Okay. J is always good. Did he say week nine? No, week 10. Oh, week 10. So that's after the break. That's right. Yeah, yeah, yeah. I believe it's week 10. Yep, it's after the break. Oh, that's when we have our showcase. Let me ask him. He will be here soon. If it's going to be week 10, there's nothing to showcase at that point. If we haven't gone through it all together. Then do you want to make the deadline next week? Yeah, I think we should set our internal demo without the DAPI team to have everything ready by next week. Yeah, can you guys do it? Yeah. No, you don't need to force yourself. I can help. Just ask for help, okay? Just ask for it. So stubborn. Yeah. I want to take the whole credit, you see. No, no, no. I'm just kidding. Yeah, I'll make... I'll figure it out next week. And don't work more than 10 hours. I don't, I don't. You don't? I don't. Okay, that's good. The dream is to make your work look like 10 hours, but you only spend like 5 hours.
Gina : That's like a corporate life.
Shreya : Yeah, that's how you want to do it. Like you just get your AI agent to do the whole thing for you, but they're like, oh, I'm still doing it, I'm still doing it. Okay, you're sure about that? Next week, deadline. Okay, so next week, let's keep two deadlines. So first deadline is just basically, like you said, you're going to do the whole integration, right? Yep. So you also need time. Because when you're going to integrate it, it's going to take you time.
Shreya : So first deadline is to just complete the whole code from my end, from Gina end, and we're going to hand over everything to you. And then you're going to sit down and do the integration. And then obviously we will have to sit down and fix a lot of other things. And in the meantime, we can work on the feedback thing for the notion and figure out what we're going to do. And then how we're going to connect with the backend, right? So then it works both ways. And then the second deadline of the week is the demo part, where we're done with the integration. And if we need to do it at the showcase, we would be able to do it. Let's keep it like that.
Teddy OK. Then next meeting will be on Friday. I will expect you guys to finish your features.
|
{
"action_items": [
{
"assigned to": "Shreya",
"deadline": "Next week",
"status": "Not Started",
"task description": "Implement Google Form to collect feedback on meeting transcripts"
},
{
"assigned to": "Shreya",
"deadline": "Next week",
"status": "Not Started",
"task description": "Connect Google Form to Notion"
},
{
"assigned to": "Teddy",
"deadline": "Next week",
"status": "Not Started",
"task description": "Design a section in Notion for users to provide feedback on meeting transcripts"
},
{
"assigned to": "Teddy",
"deadline": "Next week",
"status": "Not Started",
"task description": "Implement a system to send feedback from Notion to the backend"
},
{
"assigned to": "Shreya",
"deadline": "Next week",
"status": "In Progress",
"task description": "Finish implementing the RAG file and data pipeline"
},
{
"assigned to": "Shreya",
"deadline": "Next week",
"status": "Not Started",
"task description": "Test the RAG file with meeting transcripts"
},
{
"assigned to": "Teddy",
"deadline": "Next week",
"status": "Not Started",
"task description": "Connect the RAG file to the data pipeline"
},
{
"assigned to": "All",
"deadline": "Week 10",
"status": "Not Started",
"task description": "Prepare a demo for the DAPI meeting"
},
{
"assigned to": "Teddy",
"deadline": "Next week",
"status": "Not Started",
"task description": "Complete the integration of all features"
},
{
"assigned to": "Gina",
"deadline": "Next week",
"status": "In Progress",
"task description": "Fix errors in the code and ensure all libraries are installed"
}
],
"agenda": [
{
"notes": "The team discussed the generation of meeting minutes using the RAG system, including the need for a data set to fine-tune the system and improve its accuracy.",
"presenter": "Teddy",
"time allocated": "",
"topic": "Meeting Minutes Generation"
},
{
"notes": "Shreya presented her work on the RAG file, including evaluation metrics such as faithfulness, answer correctness, and context relevancy, and discussed the need for a gold standard data set.",
"presenter": "Shreya",
"time allocated": "",
"topic": "RAG File and Evaluation Metrics"
},
{
"notes": "The team discussed the need for a feedback mechanism to improve the RAG system, including the possibility of using Google Forms to collect feedback and connect it to Notion.",
"presenter": "Teddy",
"time allocated": "",
"topic": "Feedback Mechanism"
}
],
"discussion_summary": [
{
"decisions": [
"The team decided to use Google Forms to collect feedback, with the option to later integrate it with Notion.",
"Shreya will work on connecting Google Forms to Notion using the Google Form Add-on.",
"Teddy will work on integrating the RAG file with the data pipeline and dealing with errors.",
"The team set a deadline for the next week to complete their features and have a demo ready."
],
"issues": [
"The team needs to design a section for users to provide feedback and decide on the best approach for collecting feedback."
],
"key_points": [
"The team discussed the current status of their project, including the parsing of markdown and the integration of Notion and Hugging Face.",
"Shreya is working on creating a data pipeline and has updated the RAG file to include evaluation metrics.",
"Teddy is working on getting feedback from Notion and is exploring ways to design a section for users to provide feedback.",
"The team discussed the possibility of using Google Forms to collect feedback and the potential benefits and drawbacks of this approach."
]
}
]
}
|
Transcription
Welcome to the second last meeting. How are you guys feeling so far?
Teddy : Gina, are you doing fine.
No, I'm okay. I'm getting there. I think after this week I should be alright.
Teddy :No, I didn't expect you to finish last week.
Honestly, the parsing bit is the hardest part. I'm so stuck on that bit. I don't know... So I kind of followed what you gave me, right?
Teddy : Yep.
Gina : But I also, like, edited some bits myself. Like, moving to the notions of this line, but the problem is, like, parsing, going through and trying to parse the, like, for example, like, the markdown like that, right?
Teddy : Yep.
Gina : And, like, I don't know if you want it, like, exactly, like, you want things to be, like, separated, or it doesn't matter, like, as long as you can parse it and put it into the... the file? Separate it as in,
Teddy : what do you mean?
Gina : Because you know how you have, so for example, there's a date object in Notion, and then, or you can just parse it as a string. I was thinking, do I need to make it into a date time object? Because Notion has their own date thing, right? And then... For attendees, you can also have it as actual people that are in the Notion workspace.
Teddy : No, but for now, just parse it as string. That's the easiest.
Gina : Well, it's fine now. It's parsing it as a string and putting it into the page. So that's fine. But I just wasn't sure if you needed to separate it into separate... specific block type?
Teddy : Nah, too difficult. We will leave it to the next generation.
Gina : Okay. Well, the next generation is not you.
Teddy : It's not you either. You don't need to be in the team. You can go on.
Gina : I feel like I should... I don't like doing things halfway. I like to see things through.
Teddy : Okay. If you really want to do the block thing, then do it. But for now... string is fine if it's too much then yeah yeah but it goes in like this is like I parse like that file and it just goes in like that but then yeah so it I can make the database as a database yeah it's just a random like That looks nice. Thanks, that's one of my favourite. K-pop?
Gina : No, it's Chinese. Chinese! So this is just like a random page on my thing, right? And there's the database which I made through Python. And then the objects that I just saved in there. These are just empty because I was just trying to test out. But this last one I was able to put through. yeah just as like the blocks myself it's really good it's not like specific but I think I can like still clean it and stuff a little bit but yeah it works it works
Teddy : yeah that's nice um but the only thing is uh I didn't because you know how you gave me
Shreya: all these yeah I feel like what you can do is you already have the categories out like meaning information agenda you can create that as one big list and then create sub list into it where you can add your categories like you're talking about yeah that's fine yeah I think that would be able like I would be able to do that if we needed to but like the um extracting of the info like the markdown into the notion is working basically um but i also i didn't really use the because you know how you gave me like all this like uh fast api stuff yeah i ended up i didn't end up using it oh it's cool it's just like using yeah don't use that just the um notion yeah the notion client yeah yeah Yeah. That's cool. Like, I spent so much time because I thought... Wait, how long did you spend? Oh, not, no, not that much. Please don't be more than 10 hours. No, no, it wasn't more than 10 hours, don't worry. But I spent a lot of time, like, thinking about, like, I forgot that you told me that you just wanted, like, just a page is fine. Because I was thinking, like, oh, do I have to, like, make a date? And then, like, people, and then, like, all the information, like, as, like, a table. I was like, oh, like, you said it was just a page, so it's fine. Because I started doing it right here. I was doing it here and I was like, oh, wait. Oh, shit. That looks pretty neat, so I think that's fine. If you should prove the next... Yeah, maybe for the future, yes. This part, yeah. That's really good. I didn't get to end up doing the grok stuff though yet. Like, implement the grok stuff. Yeah. OK.
Shreya : All right, should I show up my stuff as well? You want me to connect mine as well? You can.
teddy : If this is too much, I can help. That's all right. I should be all right. We can work on it in the next two weeks. It should be fine.
Shreya : How does this thing work? I think first I have to change the display settings. Oh, the screen mirroring. Wait. Do you mind if I? Yeah. So yeah, just drag it like that. Oh, OK. Wait. Why is it not opening? Well, it's not opening. Wait, wait. OK. Oh, yeah.
Teddy : In the meantime. What is that Chinese actress? Is it an actress?
Gina : Yeah, actress.
Teddy : I don't know. But my sister would know. She might know. She loves watching Chinese dramas.
Gina : Oh, I like Chinese dramas as well.
Teddy : I'm surprised it's not K-pop.
Gina : I do like Koreans as well, but the last few years, I've been into Chinese dramas. Yeah, last few years.
Gina : So, how do you, why? Why? I feel like Chinese dramas are so diverse. it's like so interesting i think
Shreya : have you guys seen that korean web series uh when life gives you tangerine uh
Gina : i started watching it and then
Shreya : oh my god it's like i was too scared to continue because i knew i was gonna cry watching it
Shreya : i watched it during my mid-semester break last semester i was crying the whole time it was like so emotional so i'm telling you even if you watch it it's really emotional and i'm like it's it's based on a true story what was that called When life gives you 10 dreams. Oh, wait. It's a really popular one. Okay, I cannot open this. The guy dies 10 times, yeah. And the last time he went against the bad guy. No, no, no. That might be a different drama. This is like a family drama. Not even a drama. It's more of like their life story. I cannot open this file on my Mac anymore. Okay, what file? So it's like I have it open in VS Code. Yep. Where's my cursor? Oh, wait. No. See? Okay. Oh, yeah. Okay. Oh, this is how it works. Okay. So I can't open it on my computer, but I can open it on the monitor. Oh, okay. All right. So I just updated the RAG file a little bit more because I'm still facing a lot of errors. And...
Shreya : I've updated this. This would be like a code walkthrough when we have like the internal meeting with the whole DAPI team. Then instead of obviously like because it's really difficult to cover the code and all of this, then I can just show this. And it basically talks about what the rag file is doing, the components it's using, and we're doing data pre-processing and then embedding, then vector search, evaluation metrics. And I've created a separate file called metrics configuration. So we're adding the retrieval metrics like the K values, the MRR cutoff, And then this is a template where we can just keep on adding more questions if we want. And this is just like, you know, GitHub issues and all this. I'm not so sure about what exactly it's doing, so I still have to check on this. And yeah, there's just like requirements, just like the technical part. And then... I think it's gone to your main. Your main screen. Go to the left mouse. Yeah. Go past it. Okay, you do it. I've never done this before. Yeah. So it's connecting to both screens together. In here. Does that make sense? Okay. So if you go to the right mouse here, you will see the mouse over there. Yeah. And if you go to the left mouse, then you will see the mouse back here. Okay, just open this for me. Okay. Oh, okay, okay. Got it. And talk about this one. Okay. And so this is basically all of the evaluation metrics for RAC file. So the first part basically talks about the evaluation metrics. Remember when we talked about semantic search and the K values in the first meaning? So this is based on that. And then mean reciprocal rank. And it's also here in the file if you see it. MRR cutoff. So I've set it as 10. But ideally, a score of means one means the first result is always the correct one. So we should try to like lower the cutoff, I feel right. And then this is called hit rate, which is like a binary measure of whether at least one relevant document was found within the top K-retreat results. So there's something called, I'm gonna search for it instead. Yeah. Establish a gold standard data set. So this is basically where we would be giving a manually verified, first of all, transcript. So then it knows the format. It knows that Ted is always going to be the manager. Jenna and I are always the team members. A small example. So in the data set, we're basically going to be, it's like a baseline, a rule book on which it's going to work. Like, you know, a book on which RAC can always like refer to and know that this is what we're talking about. This is our relevant documents of like who's there in the team, who's the team lead, what projects we're working on. And if we have any other notes side by side, you know, anything like that, or we can try connecting it with Discord or something like that. And so then basically we can then retrieve all of those things from here. From the gold standard data set, it's going to retrieve the hit rate, the reciprocal rank. just to ensure that one relevant document was found when we got the meeting transcript. And then obviously the context should be relevant. And then this is to evaluate the generation it's doing, how accurate it is and how reliable it is. So there's something called faithfulness, which is basically to see if it has generated any sort of false information or just random stuff on its own, like how ChatGP does it sometimes. Then answer correctness, answer relevancy. And then this is basically all the technical things that we have to keep in mind. Like the goals turned data set. Then automated scoring. And obviously there has to be a human in loop at all times. And then establish automated testing pipelines. So are we gonna do this in the hugging phase thing? Or is that gonna be different? we're going to do this with the hugging face and other stuff. Yeah, we can implement this in the queries that we were talking about, positive retrieval and negative retrieval. Okay. And then this is just like more on a theory base now on what practices we can do to make sure that we're following the steps for accurate metric, which is faithfulness, answer correctness, the context relevancy and tailor specific domains like how I was talking about in the first meeting that it could be possible that a different team even though they might be in the technical arena but they could still be talking about something completely different. So it has to be tailored to their specific domain also. And now these are the possible test cases that we have to deal with in our code, which is basically like whether, let's say the chunks of information it's retrieving, is it losing the context or is it too small, too large, or it's just too much of irrelevant information? And then weak re-ranker, which is basically like You know, the, like, RAG file, it might lose certain key points that we discussed during the meeting, so it might not include that. And what metrics we have to deal with if we want to improve it. Okay, and then the uniform uniformity of embedding which is basically like when we get the final transcript is a uniform is it consistent is it domain aware all of those things, and this is again regarding faithfulness and then ignoring context, so we have to deal with these test cases on our code itself. And I've like added links also from where exactly I got it. Just give it a minute. So see wrap pipelines and like exactly how we can make sure. So I think we have to like work on these things also. So yeah, and then this is just like a real world application and then all the sources I've used. So what I've done is from part time to time, like you can just open any of these tools and see how, like this is from Geeks for Geeks. So how we can maintain these things and it's got like a sample code also. So yeah, this is what I've worked on so far. All right, over to you. And it's almost done.
Teddy : And what Shreya is working on is uploading to Hugging Face. And she will do this later. And you follow it in this so that it improves the time. OK? Make sense? Yeah. OK. Right now, I'm working on getting feedback from Notion. Right, so we got chunks from all the queries here, which used to generate the meeting minutes. And what if the chunks are not correct? Like Shreya said, it's not accurate. So we need to make a data set to fine tune this. system right and what we need is but how do we know if this is incorrect right so we need to get feedback right yeah and what i've done so far is that see this is a simple Notion page. And we got a transcript here. And we got error in here. Student D doing API performance while the student C doing it. Right. Oh, this is a bad example, but it's an example. So you see the button here? You can click on it, and by linking We can get the retrieved issue. What is the query of that part? What's the wrong retrieval, which is the current retrieval? And what is the correct retrieval? Right. And then this will send to a hugging face with the pipeline to create a data set. Yeah. So what I need is when you are generating the Notion page, I want the ID of a specific block. So maybe we should add a section like chunk sending and buttons.
Gina : So how are you planning to get the feedback?
Teddy : Yeah, this part is too hard. No, no, no. That's the manual.
Manually, right? No, that is not. So when we're making the block, we don't have this part. So we just have maybe a section saying. Yeah, that section that chunks to send bit. But then you have the one that you want to send, and then you manually have to send it by pressing the button. Is that what you mean? Yes. So let's say you see, oh, this one is wrong, and you said, I want this chunk to be the correct one. Put the incorrect one in here, and then you send it by clicking the button. So that would be done manually, because you have to click the button. And what I need is the ID of this block. So when you are generating the page in Notion, you will have the block, and I need the ID. And also, what I need from Shreya is what kind of chunks I need to send it into the database.
Shreya : For the feedback, right?
Teddy : For the fine-tuning.
Shreya : For the fine-tuning.
Teddy : Yeah, for the fine-tuning. And yeah, another thing. You don't need to look into the code. I just need the information. Oh. Yeah. Do you mean like the information in? No. The information is in the chunk. So this block, it will have an ID. And it will have an ID, yeah. And I want the ID. So you just need to extract that ID from there. But I still want to know like for feedback, what questions are you planning to ask the members to rate how accurate the Yeah, that is also a problem. Because as you can see here, the template is quite a bit messy. And I still haven't figured out how to add in that chunk to sentence and buttons inside that template so that people can give feedback. Adding that chunk at the bottom. At the bottom. But which one is referring to? I mean we can add more text in here so that people can better identify. What is that any function is like? Do you mean like for the user to differentiate? Yeah.
Gina : Do you mean the user or for the code to know?
Teddy : both because if you add in at the end then the code will never know which chunks is referring like which section is referring to and also for the user if you just accept chunks to send it's not descriptive enough yeah exactly so what i'm facing the issue is what how to design that section so that users can specify which sections they want to change.
Shreya : What we can do is we can add the link of a Google Form. We can make a Google Form where we can ask them questions. Because on Google Form, we can add drop down questions, we can add MCQ, stuff like that. Rate how accurate you found the transcript, what improvements we can do. So we give options for everything. Because if you just add everything on Notion, there is a possibility it could get a little bit cluttered. So I would say first get it on Google Form. So then we also have the IDs of each and every person, right? Because when you fill out a Google Form, you have to do it with your Google ID and stuff, right? And then when we have done it and we feel that, okay, this is what we're getting, then we can try to merge it with Notion. Because if you do it on Notion, it's like even if it messes up, you know, it's going to be like all over Notion and it might create a hurdle for like other things as well. In case if there's a problem in the code or if it doesn't fetch things correctly or if it messes up the ID or stuff like that. Right. Because in Notion, sometimes it's like it also gives you the option to go and make edits for somebody else. Like I can open your profile on Notion and change your profile picture for you. Right. So I mean, just to like avoid all of those things, we can first try it on like Google Form, where we know that the person who is taking the Google Form, like if Gina is giving me feedback on the transcript we got, then it is from Gina and, you know, we can say that it creates more like privacy and everything.
Teddy : I see. What do you think?
Gina : I'm kind of confused. Isn't the point of this one for like... So this is the meeting minutes that we designed. Or like, not designed, that's outputted. And then this part here is the... So you're reading it and then you notice that there's an incorrect output. Because it says student D but it should be student C. So you want to tell... You want to send that bit... to tell like whatever here is that it's incorrect and they need to like yeah because right now the chunks we're sending it's just like even if we add a few more buttons to it and even if we give like we can't add questions to it right like because we're just sending the chunk we only want the chunk that's incorrect right okay so let's say it's incorrect the manager can change it themselves it doesn't have to change through the code But the process is not robust enough, meaning, right?
Gina : You mean like currently, like in Notion itself at the moment?
Teddy : What do you mean induction?
Gina : Do you mean like this process at the moment?
Teddy : No, I mean the RAG system is not robust enough. Oh, okay, yeah. Yeah, because the RAG system provides the answer. Oh, yeah, because it's giving you the incorrect answer, yeah. Yes, so what we want to do is to make a data set to improve the rack system.
Gina : Mm-hmm.
Teddy : So that when you click in that button, it will send that into FastAPI. And we will get the query wrong chunks and correct chunks. Wrong chunks we know because it's the chunks that used to generate that. But correct chunks we don't know. So we also need the manager to give us that information.
Gina : Yeah, so that's why you're sending this, like sending the correct, what should have been outputted.
Teddy : And the problem is, how are we going to design that so that But the system knows it's that session that is wrong and also I think just for like if you want for now you can just add a few questions to it and like I said you know that's what I'm saying on Notion we are bounded We can't add a lot of diversity into the feedback. But I feel like if on Google Form, you know how we take polls and stuff as well on Google Form? So it generates bar graphs for you also to tell you what percentage of people said. We can get it. Let's say if we hand this out. like this DAPI know to other DAPI teams and they attempted and they're like yeah you know 40% people found the meeting transcript to be accurate 20% people said that it got the ID wrong you know first we can start asking like basic questions because we were at the initial stage right so we need to ask the basic questions like did it get the topic right did it get the student right did it get the notes right we have to ask specific questions also Right. I see what you... Yeah. Because until the time we don't try it out, like the hidden trial method. If you do it through Google Forms, then how are you going to send that? Like, for... Because you want to send this to... It doesn't have to be in FastAPI. That is just triggering a signal into the backend so that we can store the information. Yeah, I think she's trying to say how are we going to link Google Form with our code or Notion and stuff. That's the other thing. I think this is fine.
Gina :But then if we add the Google Docs and stuff, you have to research more about how to link it all together. Or we can try to look for some other tool that is already doing that type of work. We can make a form here so that they can do it within Notion. That also works. Oh, they have forms. I don't know. I don't need to know. Same. That's why. We can just make a block that has all the questions and the response we want and make a button so that we can extract that in one go. That also works.
Teddy L It's just not aesthetic and elegant, I guess. But who cares? I think the notino is to... I'm fine with that. That's fine. It's just like the way Notion is developed, it just kind of makes it difficult.
Gina : Yeah, you're right. You're right.
Shreya : Okay, so it turns out we can connect Google Forms to Notion.
Teddy : How?
Shreya : Yep. Oh, so it's like... So it's like you take the data from Google Form and you connect it to Notion and it generates the same thing for you. And then you collect data inside Notion. But I want to, because now you are having Google Form to Notion, right?
Shreya: Yeah, it's saying you can just do it directly. It's called Google Form Add-on.
Teddy : But what I want is that Notion links to Google Form and Google Form back to the backend. No, we can do Google Form to Notion and then Notion to backend. Because we would anyway, like DAPI is going to be inbuilt in Notion, right? So Notion is our primary interface. Google Form is not the primary interface. I'm just suggesting Google Form because it has other features also, but Notion doesn't have that. Notion is mainly for making notes and structuring and organizing everything. So we get the data from Google Form to Notion, and then at Notion we can do whatever we want. We can get it in backend, we can keep updating DAPI, all of those things. How are you going to trigger that through Notion? Yeah, that is the main thing. I mean, this is a good situation. Either we connect it with our gold standard data set where we keep updating the documents and they have to be manually verified, like I researched in the metric thing. Because I still feel like I've not used Notion that much, but I still don't think Notion is that accurate when it comes to feedback, like the feedback that we want.
Teddy : I'll think about it, and also I will make...
Shreya : Yeah, you can also have a look at this, right, and see what we can do. Do you want to copy the link and just put it in the...
Teddy : Yeah, okay. I'll have a look at it. Just copy the link. All right. OK. Should be fine then. Is it too much work?
Shreya : No, it's OK. So Gina will finish that. And on Hugging Face, I have to do how the RAG file is going to get positive and negative retrievals. Yeah, and also what kind of positive and negative retrieval you're getting. Right, OK.
Shreya : So before you came, we were trying out Clue.ly, and turns out it is doing exactly what we're doing. It's giving you the transcripts. It can also give me individual transcripts, like what each one of us is talking about. It can also give you follow-up questions and everything. So I'm thinking we'll get today's meeting and the past meeting, because we only have two recordings so far, get their transcript and have it laid out. Because the RAC file that I've made, I've not even once tested out how it's generating the transcript, right? Because the RAC file that I have, like, if I give it, like, the main thing that we're trying to do, how am I going to test out my RAG file? Like, I've made it, but how are we going to test out the implementation it's doing? So for that, we need a transcript, right? So... How am I going to connect it with the data pipeline? That's what I'm wondering.
Teddy : Oh, you're not connecting that. I will do the connection. Okay. Yeah.
Shreya : Okay. So I'll just make the data pipeline thing. Yeah. Give you the rag file, the data, and everything, and you're going to connect everything. Yeah. Okay. And you're going to deal with the errors? That's cool. Deal with the errors. Because this file is still showing a lot of errors. I don't know why. So I still have to deal with it. And it's just showing, like, stupid, like, module error and, like, import errors for some reason. Oh. I don't know why.
Gina : Yeah. So that's why. Well, I fixed one, so my imports are good. Oh, because she's not using my requirement.txt. Yeah, no, I edited it. Yeah. Yeah, because my PC is not working. I just removed some of the, what is it? I forgot what I removed. Tendency. Yeah, just a few of them. We can try out the requirement.txt if you want. That should be all of them. All of the libraries that we need? Yes, and you need to remove some of them though. Yeah, you told me there's like errors in the file. Oh, okay. Hugging face, got it. Yeah, that's it. Okay.
Teddy : Then, next week we actually need a proper demo. Right. So when are we planning to have the meeting? Meeting? You mean the... Yeah, the DAPI meeting. Is it decided yet? No. It's not decided yet. Okay. Yeah, should we have some... J, J, J. Yeah. Is he back from Miami yet? Or no? Oh, no, not J. It's J. Oh, J. Okay. J is always good. Did he say week nine? No, week 10. Oh, week 10. So that's after the break. That's right. Yeah, yeah, yeah. I believe it's week 10. Yep, it's after the break. Oh, that's when we have our showcase. Let me ask him. He will be here soon. If it's going to be week 10, there's nothing to showcase at that point. If we haven't gone through it all together. Then do you want to make the deadline next week? Yeah, I think we should set our internal demo without the DAPI team to have everything ready by next week. Yeah, can you guys do it? Yeah. No, you don't need to force yourself. I can help. Just ask for help, okay? Just ask for it. So stubborn. Yeah. I want to take the whole credit, you see. No, no, no. I'm just kidding. Yeah, I'll make... I'll figure it out next week. And don't work more than 10 hours. I don't, I don't. You don't? I don't. Okay, that's good. The dream is to make your work look like 10 hours, but you only spend like 5 hours.
Gina : That's like a corporate life.
Shreya : Yeah, that's how you want to do it. Like you just get your AI agent to do the whole thing for you, but they're like, oh, I'm still doing it, I'm still doing it. Okay, you're sure about that? Next week, deadline. Okay, so next week, let's keep two deadlines. So first deadline is just basically, like you said, you're going to do the whole integration, right? Yep. So you also need time. Because when you're going to integrate it, it's going to take you time.
Shreya : So first deadline is to just complete the whole code from my end, from Gina end, and we're going to hand over everything to you. And then you're going to sit down and do the integration. And then obviously we will have to sit down and fix a lot of other things. And in the meantime, we can work on the feedback thing for the notion and figure out what we're going to do. And then how we're going to connect with the backend, right? So then it works both ways. And then the second deadline of the week is the demo part, where we're done with the integration. And if we need to do it at the showcase, we would be able to do it. Let's keep it like that.
Teddy OK. Then next meeting will be on Friday. I will expect you guys to finish your features.
|
{
"action_items": [
{
"assigned to": "Shreya",
"deadline": "Next week",
"status": "Not Started",
"task description": "Implement Google Form to collect feedback on meeting transcripts"
},
{
"assigned to": "Shreya",
"deadline": "Next week",
"status": "Not Started",
"task description": "Connect Google Form to Notion"
},
{
"assigned to": "Teddy",
"deadline": "Next week",
"status": "Not Started",
"task description": "Design a section in Notion for users to provide feedback on meeting transcripts"
},
{
"assigned to": "Teddy",
"deadline": "Next week",
"status": "Not Started",
"task description": "Implement a system to send feedback from Notion to the backend"
},
{
"assigned to": "Shreya",
"deadline": "Next week",
"status": "In Progress",
"task description": "Finish implementing the RAG file and data pipeline"
},
{
"assigned to": "Shreya",
"deadline": "Next week",
"status": "Not Started",
"task description": "Test the RAG file with meeting transcripts"
},
{
"assigned to": "Teddy",
"deadline": "Next week",
"status": "Not Started",
"task description": "Connect the RAG file to the data pipeline"
},
{
"assigned to": "All",
"deadline": "Week 10",
"status": "Not Started",
"task description": "Prepare a demo for the DAPI meeting"
},
{
"assigned to": "Teddy",
"deadline": "Next week",
"status": "Not Started",
"task description": "Complete the integration of all features"
},
{
"assigned to": "Gina",
"deadline": "Next week",
"status": "In Progress",
"task description": "Fix errors in the code and ensure all libraries are installed"
}
],
"agenda": [
{
"notes": "The team discussed the generation of meeting minutes using the RAG system, including the need for a data set to fine-tune the system and improve its accuracy.",
"presenter": "Teddy",
"time allocated": "",
"topic": "Meeting Minutes Generation"
},
{
"notes": "Shreya presented her work on the RAG file, including evaluation metrics such as faithfulness, answer correctness, and context relevancy, and discussed the need for a gold standard data set.",
"presenter": "Shreya",
"time allocated": "",
"topic": "RAG File and Evaluation Metrics"
},
{
"notes": "The team discussed the need for a feedback mechanism to improve the RAG system, including the possibility of using Google Forms to collect feedback and connect it to Notion.",
"presenter": "Teddy",
"time allocated": "",
"topic": "Feedback Mechanism"
}
],
"discussion_summary": [
{
"decisions": [
"The team decided to use Google Forms to collect feedback, with the option to later integrate it with Notion.",
"Shreya will work on connecting Google Forms to Notion using the Google Form Add-on.",
"Teddy will work on integrating the RAG file with the data pipeline and dealing with errors.",
"The team set a deadline for the next week to complete their features and have a demo ready."
],
"issues": [
"The team needs to design a section for users to provide feedback and decide on the best approach for collecting feedback."
],
"key_points": [
"The team discussed the current status of their project, including the parsing of markdown and the integration of Notion and Hugging Face.",
"Shreya is working on creating a data pipeline and has updated the RAG file to include evaluation metrics.",
"Teddy is working on getting feedback from Notion and is exploring ways to design a section for users to provide feedback.",
"The team discussed the possibility of using Google Forms to collect feedback and the potential benefits and drawbacks of this approach."
]
}
]
}
|
README.md exists but content is empty.
- Downloads last month
- 6