Hi how about Search Score What are your thoughts about it? will it represent the accuracy of the information taken from the DB which relates to the query?
- 00:00: Discussion on transforming customer support with AI and language models - 00:16: Introduction to MyAsk AI founders and their focus on customer support - 01:04: Empowering customer support using AI chatbots and gathering feedback - 02:40: Insights into founders' background and transition to entrepreneurship - 04:44: Entrepreneurial journey from travel startup to AI-focused company - 04:52: Transition from tech interest to starting own business - 05:14: Challenges faced with Pluto travel startup and its closure - 06:01: Evolution towards AI with MK AI development - 09:28: AI Chat GPT customization and usability - 09:30: Focus on reducing support ticket volume - 09:42: Tailored content for users and customer support - 09:53: Differentiation from generic models for accuracy and customization - 14:12: AI Support Features Overview - 14:16: 75% resolution by AI, 25% to human - 14:32: Focus on quality answers through data - 15:10: Insights from conversations for better focus - 18:56: Discussion on key tools for rapid business development - 19:02: Usage of Bubble for no-code development in front and back end - 19:17: Importance of fast development cycles utilizing tools like Bubble and Carbon - 20:02: Integration of Portkey for AI model requests and fallback options - 23:41: Discussion on embedding models and chunking strategies - 24:02: Migration to new open AI embedding models - 24:24: Challenges in migrating embedding models - 24:46: Performance improvement with new OPI models - 28:25: Key points on data processing challenges and AI advancements - 28:37: Suggests focusing on customer use cases to streamline data handling - 29:05: Advises identifying poor-quality data sources to enhance outcomes - 29:29: Emphasizes the need for novel strategies in tackling data processing issues - 33:09: Challenges in deploying AI technologies and navigating distribution channels - 33:34: Differentiating services in a crowded market - 33:42: Demonstrating credibility and business stability - 34:25: Struggle to find repeatable distribution channels for AI products - 37:53: Discussion on managing system prompts and user feedback in LLMS production - 37:57: Challenges in ensuring obedience to system prompts - 38:07: Importance of patience and testing in LLMS production - 39:00: Handling user feedback signals and AI-human handover for improvement - 42:37: Discussion on handling unstructured data and updating content - 42:50: Scraping processes used for web content and PDFs - 42:58: Handling tabular data and structuring it - 44:20: Updating content in Pinecone index for customers
I was traveling when the stream happened, so I couldn't attend live. I very much appreciate that you posted this recording. Excellent information and analysis here.
One question -- it seems to me like each embedding / vector would contain a different number of dimensions. Trying to establish a master-type vector template with every single conceivable dimension represented would involve mainly blank space and be a computational nightmare (hence PCA and other dimensional reduction techniques). So if something complex like "the US Constitution" has thousands of dimensions and something like "grass" has hundreds of dimensions, how can they be compared, seeing as they reside in spaces with different numbers of dimensions? Like, you can't find the distance between an object that resides in 7 dimensional space and an object that resides in 11 dimensional space, right?
What? Since when did English become an ambiguous language? From my understanding, it's the opposite. Ambiguous languages are those like Semitic languages. Just the fact you can use a different English word to clarify an ambiguous English word is unambiguous.
what about using javascript imstead of python cause i am doing a project and pinecone is suppoused to retireve the data from aws s3 and open a chat where the user chats with their pdf
I'm a guy who just wants to use the tools to get the work done. I jumped to serverless, got halfway across and then saw that langchain which I was using (and I think many others) is broke when it comes to Pinecone.from_documents(). Loading a premade dataset to pinecone.dataset does not help me as I don't need to work with a play dataset. Things are moving fast on this I know, but the transitions should be a little more coordinated. I would love to move forward but I needed some handholding which just wasn't there. Other youtube videos with Mr Briggs are tremendous BTW.
If I have currently have a POD which has a lot of data and want to migrate it to serverless, is there a simple path or would I need to write the code to transfer the data?
Great demo on the basics of image search! Thx! I do have some follow up questions: 1. Looks like the label is single-class (i.e. bird, jelly fish, dog). In the embedding-creation step, how is multi-class picture being handled? For example, if a picture has both a woman and a dog, would there be multiple labels (or perhaps a label is a list of classes)? Would squeeze net able to handle this situation? 2. Similarly, in the matching stage, would it be able to handle multiple labels? 3. Or in general, does it require a pre-processing stage to dissect the images into multiple objects, and feed each object into detection algorithm? Thx