Lakshya A. Agrawal is a Ph.D. student at U.C. Berkeley! Lakshya has lead the research behind GEPA, one of the newest innovations in DSPy and the use of Large Language Models as Optimizers! GEPA makes three key innovations on how exactly we use LLMs to propose prompts for LLMs, (1) Pareto-Optimal Candidate Selection, (2) Reflective Prompt Mutation, and (3) System-Aware Merging. The podcast discusses all of these details further, as well as topics such as Test-Time Training and the LangProBe benchmarks used in the paper! I hope you find the podcast useful!

Maarten Grootendorst is a psychologist turned AI engineer who has created BERTopic and authored "Hands-On Large Language Models" with Jay Alammar. The rise of LLMs and Agents are transforming many areas of software! This podcast dives deep into their impact on Topic Modeling! Maarten designed BERTopic from the start with modularity in mind -- letting you ablate embedding models, dimensionality reduction, clustering algorithms, and more. This early insight to prioritize modularity makes BERTopic incredibly well structured to become more "Agentic". An "Agentic" Topic Modeling algorithm can use LLMs to generate topics or topic descriptions, as well as contrast them with other topics. It can decide which topics to subdivide, and it can integrate human feedback and evaluate topics in novel ways... I hope you find the podcast interesting!

Hailey Joren is a Ph.D. student at UCSD! Hailey and collaborators at Duke University and Google have recently published Sufficient Context: A New Lens on Retrieval Augmented Generation Systems in ICLR 2025! There are so many interesting nuggets to this work! Firstly, it really helped me understand the difference between *relevant* search results and sufficient context for answering the question. Armed with this lens of looking at retrieved context, Hailey and collaborators make all sorts of interesting observations about the current state of Hallucination. RAG unfortunately makes the models far less likely to hallucinate, and the existing RAG benchmarks unfortunately do not emphasize retrieval adaptation well enough -- indicated by LLMs outputting correct answers despite insufficient context 35-62% of the time! However, reason for optimism! Hailey and team develop an autorater that can detect insufficient context 93% of the time! There are all sorts of interesting ideas around this paper! I really hope you find the podcast useful!

Nandan Thakur is a Ph.D. student at the University of Waterloo! Nandan has worked on many of the most impactful works in Retrieval-Augmented Generation (RAG) and Information Retrieval. His work ranges from benchmarks such as BEIR, MIRACLE, TREC, and FreshStack, to improving the training of embedding models and re-rankings, and more!

Multi-vector retrieval offers richer, more nuanced search, but often comes with a significant cost in storage and computational overhead. How can we harness the power of multi-vector representations without breaking the bank? Rajesh Jayaram, the first author of the groundbreaking MUVERA algorithm from Google, and Roberto Esposito from Weaviate, who spearheaded its implementation, reveal how MUVERA tackles this critical challenge.

Dive deep into MUVERA, a novel compression technique specifically designed for multi-vector retrieval. Rajesh and Roberto explain how it leverages contextualized token embeddings and innovative fixed dimensional encodings to dramatically reduce storage requirements while maintaining high retrieval accuracy. Discover the intricacies of quantization within MUVERA, the interpretability benefits of this approach, and how LSH clustering can play a role in topic modeling with these compressed representations.

This conversation explores the core mechanics of efficient multi-vector retrieval, the challenges of benchmarking these advanced systems, and the evolving landscape of vector database schemas designed to handle such complex data. Rajesh and Roberto also share their insights on the future directions in artificial intelligence where efficient, high-dimensional data representation is paramount.

Whether you're an AI researcher grappling with the scalability of vector search, an engineer building advanced retrieval systems, or fascinated by the cutting edge of information retrieval and AI frameworks, this episode delivers unparalleled insights directly from the source. You'll gain a fundamental understanding of MUVERA, practical considerations for its application in making multi-vector retrieval feasible, and a clear view of future directions in AI.

AI agents are getting more complex and harder to debug. How do you know what's happening when your agent makes 20+ function calls? What if you have a Multi-Agent System orchestrating several Agents? Anand Kannappan, co-founder of Patronus AI, reveals how their groundbreaking tool Percival transforms agent debugging and evaluation. Percival can instantly analyze complex agent traces, it pinpoints failures across 60 different modes, and it automatically suggests prompt fixes to improve performance. Anand unpacks several of these common failure modes. This includes the critical challenges of "context explosion" where agents process millions of tokens. He also explains domain adaptation for specific use cases, and the complex challenge of multi-agent orchestration. The paradigm of AI Evals is shifting from static evaluation to dynamic oversight! Also learn how Percival's memory architecture leverages both episodic and semantic knowledge with Weaviate!This conversation explores powerful concepts like process vs. outcome rewards and LLM-as-judge approaches. Anand shares his vision for "agentic supervision" where equally capable AI systems provide oversight for complex agent workflows. Whether you're building AI agents, evaluating LLM systems, or interested in how debugging autonomous systems will evolve, this episode delivers concrete techniques. You'll gain philosophical insights on evaluation and a roadmap for how evaluation must transform to keep pace with increasingly autonomous AI systems.

How do you ensure your AI systems actually do what you expect them to do? Leonard Tang takes us deep into the revolutionary world of AI evaluation with concrete techniques you can apply today. Learn how Haize Labs is transforming AI testing through "scaling judge-time compute" - stacking weaker models to effectively evaluate stronger ones. Leonard unpacks the game-changing Verdict library that outperforms frontier models by 10-20% while dramatically reducing costs. Discover practical insights on creating contrastive evaluation sets that extract maximum signal from human feedback, implementing debate-based judging systems, and building custom reward models that align with enterprise needs. The conversation reveals powerful nuggets like using randomized agent debates to achieve consensus and lightweight guardrail models that run alongside inference. Whether you're developing AI applications or simply fascinated by how we'll ensure increasingly powerful AI systems perform as expected, this episode delivers immediate value with techniques you can implement right away, philosophical perspectives on AI safety, and a glimpse into the future of evaluation that will fundamentally shape how AI evolves.

Ben walks us through Box's three-layer infrastructure puzzle: First, the mind-boggling base infrastructure (think millions of interactions per second and trillions of files). Second, their unique multi-tenant security challenge - unlike most SaaS platforms, Box users share content across company boundaries, making traditional tenant isolation impossible. And third, ensuring AI respects all these complex permissions while still delivering value. The podcast then dives further into how vector embeddings can balloon file sizes - a few hundred bytes of text can require 4-6KB of vector data storage! We also dig into why RAG remains essential despite growing context windows, and how Box is developing AI agents that transform painful enterprise processes like RFP responses.

Hey everyone! Thanks so much for watching another episode of the Weaviate Podcast! Dive into the fascinating world of structured outputs with Will Kurt and Cameron Pfeiffer, the brilliant minds behind Outlines, the revolutionary open-source library from .txt.ai that's changing how we interact with LLMs. In this episode, we explore how constrained decoding enables predictable, reliable outputs from language models—unlocking everything from perfect JSON generation to guided reasoning processes.Will and Cameron share their journey to founding .txt.ai, explain the technical magic behind Outlines (hint: it involves finite state machines!), and debunk misconceptions around structured generation performance. You'll discover practical applications like knowledge graph construction, metadata extraction, and report generation that simply weren't possible before this technology.Whether you're building AI systems or curious about where the field is heading, you'll gain valuable insights on how structured outputs integrate with inference engines like vLLM, why multi-task inference outperforms single-task approaches, and how this technology enables scalable agent systems that could transform software architecture forever. Join us for this mind-expanding conversation about one of AI's most important but under appreciated innovations—and discover why the future might belong to systems that combine freedom with structure.

Synthetic Data: The Building Bocks of AI's Future! Hey everyone! I am SUPER EXCITED to publish the 118th episode of the Weaviate Podcast featuring David Berenstein and Ben Burtenshaw from HuggingFace! This podcast explores the intricacies of synthetic data generation, detailing methodologies such as data augmentation, distillation, and instruction refinement. The conversation delves into persona-driven synthetic data, highlighting applications like Persona Hub, and discusses algorithms to enhance diversity, complexity, and quality of generated data. Additionally, they cover integration with Hugging Face’s ecosystem, including Argilla for annotation, AutoTrain for fine-tuning, and advanced data exploration tools like the Data Studio and SQL console. The podcast also touches upon the potential for synthetic image data generation and the exciting future of AI education and accessibility.