Phase 3 · Embeddings & RAG

Turbopuffer's Simon Hørup Eskildsen on what retrieval looks like after classic RAG: hybrid search, agentic retrieval, and the database design decisions behind production vector search at scale.

The full pipeline — parse, chunk, embed, retrieve, rerank, generate — and the decisions that determine whether your AI gives good answers or hallucinated garbage. War stories from production RAG systems that went sideways.

Why neither keyword nor vector search works alone, how Reciprocal Rank Fusion merges incompatible scores, why re-ranking is the precision secret weapon, and why hybrid search is the non-negotiable default for serious RAG.

LlamaIndex founder Jerry Liu on the limits of naive RAG, why retrieval is a stopgap, and where data frameworks for LLMs are headed.

Bloomberg's Sebastian Gehrmann on how RAG can inadvertently degrade model safety, how to evaluate those risks, and domain-specific safety taxonomies for high-stakes generative AI.

Why classic RAG was built for chatbots, not agents; how Pinecone, PageIndex, SAP, and GraphRAG attack different shapes; and what a retrieval contract actually looks like for production AI agents.

Making agents knowledgeable about company data with agentic RAG: using a vector data store to organize information into a form an agent can work with efficiently, and the approaches available today.

Pinecone's Roie Schwaber-Cohen on where RAG breaks first as systems scale, answers that are technically correct but fundamentally wrong, and rethinking 'truth' when building enterprise AI.

AWS's Marc Brooker on how agentic workflows transform database usage — vectors, RAG, and relational — and why agents need serverless, elastic, operationally simple databases.

Shahul from Ragas on systematically measuring and improving RAG/LLM apps: specific metrics, benchmarking models vs. evaluating apps, and generating synthetic test data.

Embeddings, similarity metrics, and approximate nearest-neighbor search — how vector systems store and query high-dimensional vectors, and where pgvector, Weaviate, Pinecone, and Milvus fit.

Pinecone's VP of Product on hallucination in generative models, vector databases and their best use cases, semantic search, and the tech stack for AI applications.

Pinecone's Edo Liberty explains how vector similarity search works and its advantages over traditional database approaches — searching billions of embeddings for similar matches in milliseconds.

Qdrant's Bryan O'Grady on traditional Lucene text search vs. modern vector databases, when exact-match vs. semantic search wins, and growing into video embeddings and local-agent contexts.

The mathematical foundations of vector embeddings, how words and concepts become vectors in multi-dimensional space, and how semantic search leverages them to connect queries to relevant information.

Word embeddings in NLP: history and evolution, transformer architecture and self-attention, fine-tuning pre-trained models, and applications including information retrieval and question answering.

What RAG is, how it differs from fine-tuning and when to use each, plus vector databases and GraphRAG — adding structure to clusters to map complex relationships between concepts.

Jerry Liu from LlamaIndex on data ingestion, indexing, and query patterns tailored for LLM applications — and going beyond the realm of vector databases.

LlamaIndex co-founder Jerry Liu on the data framework for connecting custom data sources to LLMs, the emerging GenAI infrastructure stack, and navigating a fast-paced field.

Why Retrieval-Augmented Generation might be the practical AI skill every developer needs in 2026, and how to build practical AI systems with it.

An accessible walk through what RAG is and how it works under the hood — turning general-purpose models into AIs aware of your specific documents — tracing back to the 2021 Facebook Research paper.

IBM's Techsplainers on how connecting AI models to external knowledge sources creates more accurate, up-to-date responses.

A glossary-style primer on RAG: why LLM users demand greater accuracy and relevance for their specific domain, and how retrieval addresses it.

Unpacking the core idea of RAG — retrieving relevant documents to enrich a model's context — plus vector-based search, knowledge-graph search, and combining vector DBs with KG queries.

A hands-on walkthrough of building a RAG system with a local LLM, with an accompanying notebook.