Instructed Retriever: Unlocking System-Level Reasoning in Search Agents 🚀

Om_Jha — Fri, 09 Jan 2026 15:09:01 GMT

Retrieval-based agents drive mission-critical enterprise workflows, but traditional RAG fails on complex constraints (e.g., recency, exclusions, source priority).

Instructed Retriever is a retrieval architecture for the agent era that carries full system context—instructions, examples, and index schema—across query generation, retrieval, and response, not just the raw user query.

What makes this exciting for Databricks users?

Instructed Retriever turns natural-language constraints into schema-aware, multi-part search plans.
Delivers large recall gains over RAG on the StaRK-Instruct benchmark.
Small offline-RL-tuned models match or beat much larger LLMs for instruction-following retrieval.
In Agent Bricks: Knowledge Assistant, it produces higher-quality answers than RAG and RAG + rerank.
Performs especially well as a tool for multi-step agents.

Building search-heavy or agentic workloads on Databricks? Instructed Retriever brings agents closer to truly understanding—and correctly executing—complex enterprise instructions.

Read the full → Blog for more details!

topic Instructed Retriever: Unlocking System-Level Reasoning in Search Agents 🚀 in Announcements

Instructed Retriever: Unlocking System-Level Reasoning in Search Agents 🚀