Build Retrieval Agents That Evaluate, Adapt, and Improve Mid-Query

​Most retrieval agents run the same pipeline for every query: wasted compute on easy questions, under-served hard ones, and guesses when the evidence is not there. In this session you build a loop that adapts. What you'll learn to build:

- An agent that picks its retrieval strategy per question, instead of forcing every query down one fixed pipeline.
- Cheap in-loop signals that flag weak retrieval early, plus how to benchmark which signals actually separate good results from weak ones on your own data (most candidates do not).
- A gate that routes each query to the right fix, including late-interaction reranking (ColBERT) for ranking errors, query decomposition (IRCoT) for missing multi-hop evidence.
- A STOP decision, so the agent abstains when the evidence is insufficient rather than guessing.