Hanna Lichtenberg and Aamir Shakir describe a gap between answering with relevant documents already supplied and finding those documents in a noisy collection. Their reported benchmark comparisons motivate improving retrieval alongside model reasoning; the presentation does not independently establish a general performance guarantee.
Aamir Shakir argues that keyword-heavy habits learned from coding tools and human-facing web search can be a poor match for semantic retrieval. Hanna Lichtenberg describes four complementary tools: broad summarized overview search, detailed semantic search, metadata filtering and exact lexical matching.
Their harness starts with search results and available metadata, divides a question into separate evidence needs and allows parallel searches within at most four rounds. It deduplicates returned chunks and stops when it has enough evidence to submit a ranked list. Prompts ask the agent to describe what it needs to find before formulating a query.
Hanna Lichtenberg describes training a smaller agent through teacher-based supervised fine-tuning and on-policy reinforcement learning. Retrieval rewards assess the final ranking, while trajectory rewards assess tool choice, query quality, exploration and efficiency. The trained agent is described as unreleased at the time of the talk; preliminary results and a separate production beta are vendor-reported rather than independently verified here.
Watch on YouTube
