Events

This is an in-person event! Registration is required to get in.  Github will email you a form the day before the event, which you will need to complete for your access pass.  Registration will close 2 days before the event.

 Connecting your unstructured data with Generative LLMs

Have some food and refreshments. Hear three exciting talks about LLMs and unstructured data.

5:30 - 6:30 - Welcome/Networking/Registration

Anyone interested in talking and learning about Unstructured Data and LLM Apps.

 How to extract Tabular Data from PDFs for RAG Pipelines using Milvus

: With the continuous advancement in data processing and extraction, delving into table extraction from PDFs has become increasingly significant. Jack will shed light on innovative solutions being built for table extractions using RAG pipelines. As his team ventures into training and testing a new model, Jack will discuss the challenges faced, the need for such a solution, and provide insights into their current progress.

 Vector Search in the Age of OpenAI Assistants using Milvus

: The OpenAI Assistants API allows you to build AI assistants within your own applications. An Assistant has instructions and can leverage models, tools, and knowledge to respond to user queries. The Assistants API currently supports three types of tools: Code Interpreter, Retrieval, and Function calling.  I will talk about and show a few examples what you can and cannot do just invoking the OpenAI assistant directly to get final answers, versus using custom execution loops for Retrieval. 

Tech Talk 3: Are CPUs Enough? A Review Of Vector Search Running On Novel Hardware

: The Cambrian explosion of artificial intelligence has ushered in a new Golden Age for computer hardware. While Nvidia steals the lion share of headlines due to their dominance running deep learning workloads, several other hardware alternatives are starting to appear in the market. AI-adjacent technologies like vector search will also reap the benefit from this hardware renaissance. In this talk, I’ll talk about the intersection of vector search and advanced hardware, hardware that goes far beyond traditional CPU design which has dominated the computing landscape for over 60 years. I’ll review the results from the first NeurIPS BigANN hardware competition, we’ll discuss the benchmarks that are key to differentiate hardware alternatives, and we will look at new chip architectures that will fuse vector search and transformer inference in one device. No prior knowledge about hardware is necessary for this talk. I will describe at a high level the fundamental differences among CPU alternatives such as GPUs, Vector Processors, Compute-In-Memory microprocessors, AI Accelerators, Neuromorphic Chips, and Field Programmable Arrays.

This is an in-person event! Registration is required to get in.  Registration will close 2 days before the event.

Unstructured Data Meetup SF

This is an in-person event! Registration using this luma page is required in order to get in. Github will email you a form the day before the event, which you will need to complete for your access pass.

 Connecting your unstructured data with Generative AI

Have some food and refreshments. Hear three exciting talks about unstructured data and generative AI with images.

Anyone interested in talking and learning about Unstructured Data and Generative AI Apps.

This is an in-person event! Registration is required in order to get into the event.  Registration in advance will close the day before the event.

Tech Talk 2: Using Vector Search to Better Understand Computer Vision Data

: Jacob Marks, MLE/DevEvangelist, Voxel51

: These days, the most popular use case of vector search is retrieval augmented generation (RAG), giving large language models relevant context with which to generate text or code. But did you know that vector search is also an incredibly powerful tool for visual data understanding?

In this talk, you will learn how to combine vector search engines with the FiftyOne open source computer vision library for unstructured data curation and visualization, so you can interactively explore and find hidden structure in your data. From standard applications like similarity search and reverse image search, to multimodal applications such as semantic search and concept interpolation, you’ll see your data like never before.

Tech Talk 3: Evaluating Multimodal RAGs in practice

: Josh Reini, Data Scientist/DevRel, TruEra

: How do you evaluate multimodal RAGs? The RAG triangle of evals still applies! In multimodal RAGs, we often start with an image as our initial query. Think of a RAG that uses similar x-ray images to recommend a likely diagnosis. To measure our retrieval quality, we can measure the embedding similarity between the query image and retrieved images. In this example, after retrieving images, we will take the linked diagnoses to each X-ray to pass to our LLM. For groundedness, we can use our standard text-based evaluations to measure the entailment of the LLM's diagnosis. Last, we can use answer relevance to ensure the provided diagnosis is relevant to the user's request.

Please make sure you are registered using this luma page. Github will email you a form before the event, which you will need to complete for your access pass.

Tech Talk 1: How to Detect and Correct Logical Fallacies in LLM model output

: Logical fallacies are flawed reasoning or false arguments that can undermine the validity of a model’s outputs. LM models are optimized to perform well on specific metrics like cosine similarity, safety, or helpfulness. However, optimizing for metrics alone does not guarantee logically sound reasoning. Users cannot depend on such outputs - if left unchecked, propagating logical fallacies can spread misinformation, confuse users, and lead to harmful real-world consequences when models are deployed in products or services. Eliminating fallacies ensures model outputs remain logically valid and aligned with human reasoning. This maintains user trust and mitigates risks.

How to stream data with Kafka into your RAG application

In this talk, Filip will delve into the integration of Kafka (Confluent Cloud) with Zilliz Cloud (Hosted Milvus), showcasing how this synergy enables real-time data ingestion, parsing, and processing. The primary objective is to enhance the accuracy and performance of your Retrieval Augmented Generation (RAG) applications. Additionally, Filip will provide a comprehensive demonstration, offering practical insights into the capabilities and benefits of this integration.

Tech Talk 3: AI-Powered Pair Programming with GitHub Copilot

: Rob Crystal-Ornelas and Kasia Sitkiewicz, GitHub

: Learn how GitHub’s AI-powered code assistant can help you save time and innovate at any career stage. We’ll demo GitHub Copilot and show the newly beta-released Copilot Chat. Also, we’ll share tips on how to prompt Copilot to make the most of this AI pair-programmer.

 Connecting your unstructured data with LLMs

6:00 - 6:30 - Welcome/Networking/Registration

6:30 - 6:50 - Yi Ding, Developer Advocate, LlamaIndex

6:50 - 7:10 - David Garnitz, Founder, VectorFlow

7:10 - 7:30 - Rachel Hu, Founder, CambioML

Tech Talk 1: Advanced Strategies for Production RAG

: Yi Ding, Developer Advocate, LlamaIndex

: As we pass the second anniversary of ChatGPT's knowledge cutoff, it's clearer than ever that retrieving the right data for the LLM is crucial for the majority of applications. What we've seen at LlamaIndex is that once you get past the demo, tuning your RAG strategy involves a lot of domain specific optimization and tradeoffs. I will be discussing a number of more advanced strategies including SubQuestionQueryEngine, "Small to Big," Document Hierarchies, and Milvus's Hybrid Search.

Tech Talk 2: Ingesting Chaos: Handling Unstructured Data Reliably at Scale for RAG & Beyond

: The wide range of scenario and edge cases to account for makes ingesting and processing unstructured data into vector databases difficult. You can offload some of the complexity by using a vector embedding pipeline. This, in combination with an automated evaluation system, will allow you to experiment with different ingestion techniques to see what works best for your data and use-case.

Christy Bergman

Events