Frontier AI Search: From Hybrid Retrieval to Agentic Deep Search

Almost every LLM application relies on retrieval, yet most RAG systems still stop at a single dense-vector lookup. This hands-on workshop takes participants from that baseline to the current frontier of AI search in three steps. Every technique is evaluated on one shared corpus, so all methods are compared on the same queries. We start with advanced retrieval. Participants extend a basic dense pipeline in Qdrant with hybrid search using BM25 and Reciprocal Rank Fusion, learned sparse retrieval with SPLADE, and metadata fusion. They then add multi-vector late interaction with ColBERT and MUVERA. For each method we discuss when it helps and what it costs in memory and latency. Next, we move to agentic search, where an LLM drives the retrieval loop itself: it queries, reads, reformulates and decides when to stop. Small models trained for this with reinforcement learning now outperform much larger general models. We walk through recent training recipes such as Search-R1 and ASearcher. We then show how to put these models to work on your own index and how to fine-tune them with GRPO. Finally, we scale the loop to deep research agents that plan, browse, backtrack and write cited reports. Participants build a compact deep-search agent on top of their own retriever, using CodeAct actions to drive complex queries and data processing. Attendees leave with working notebooks and a practical map of which technique fits their latency, cost and quality budget. They also get an overview of where agentic approaches pay off and where a well-tuned hybrid retriever is enough.