BRK300 Breakout Thu 19 Nov 14:45 PST
Building agentic retrieval: From DIY to Foundry IQ
Pamela Fox, Matt Gotteiner
Summary
From the recording
The session compared hand-built agentic retrieval in Python with Foundry IQ's built-in approach. Microsoft Retriever One, described as now in public preview, is a fine-tuned model that searches first and then reflects. Microsoft reported about 8 extra points of quality, and about a third fewer tokens when GPT 6.1 SOL drives a search tool with it. Auto retrieval reasoning effort, introduced in August, now escalates from search to Retriever One to the LLM. Pricing is 10 cents per million input tokens, with output tokens free. Caveat: figures are Microsoft's own evaluations; auto mode may return less complete evidence. GA date and regions were not stated.
Session description
Microsoft's catalogue description
Why does agentic retrieval matter, and what does it take to build it yourself? In this session, we will first code an agent loop that decomposes a question, runs parallel searches, inspects evidence, and decides whether to search again. Then we'll see how Foundry IQ can replace that plumbing with a managed agentic knowledge base. Learn how we made the managed service cheaper, faster, and better, thanks to serverless execution, compression, fine-tuned task-specific models, and built-in ranking.