Descriptions:
At the AI Engineer conference, Leonie Monigatti — ML developer advocate at Elastic, the company behind Elasticsearch — presents a hands-on workshop built around a provocative claim: context engineering is roughly 80% agentic search. The talk opens by tracing the evolution from fixed RAG pipelines, where a user query was passed verbatim to a vector search and retrieved chunks were injected into the context window, through to agentic RAG, where a language model decides autonomously when and how to invoke retrieval tools and can issue multiple sequential queries for multi-hop reasoning tasks.
Monigatti then extends this framework to the multi-source context landscape that real-world agents operate across: local file systems, relational and vector databases, the web, working memory like scratchpad files, and long-term memory stores. For each source type she walks through the native search tools typically used to retrieve from it, including semantic vector search, SQL-style general-purpose query executors, and web search APIs. The live coding sequences are particularly instructive — one demonstrates an agent generating a syntactically plausible but semantically broken ESQL query (using % as a wildcard instead of *), returning zero results, and then self-correcting when given an error-handling wrapper that returns failure messages back to the model rather than crashing.
Practical guidance covers when to upgrade from lightweight models like GPT-4o Nano to more capable variants for complex query generation, how to write tool descriptions that reduce query generation errors, and when agent skill files are a better design choice than in-prompt instructions. Slides and code are available via the QR code shared during the session.
📺 Source: AI Engineer · Published May 08, 2026
🏷️ Format: Tutorial Demo
