Databricks has released a new retrieval model called Adaptive Instructed-Retriever, designed to match the answer quality of frontier AI agents like Claude and GPT-5.6 Luna while operating in half the time. The system, introduced in research published this week, addresses a core inefficiency in agentic search: the fixed number of steps required to answer questions. By determining how many search iterations a query actually needs, the model aims to reduce latency without sacrificing accuracy. This development comes as enterprises struggle to balance speed with the reliability of AI-generated insights from internal documents.
Most agentic search tools face a binary choice in how they handle queries. If the search stops too early, it misses complex questions that require multiple hops of reasoning across different files. If it never stops, it wastes resources running the same number of steps on simple queries that only needed one pass. Databricks built the original Instructed Retriever in January to outperform traditional RAG by up to 70% on instruction-heavy enterprise questions. That initial system worked well for curated workspaces with smaller, defined document sets.
The challenge arose when teams connected large language models to entire enterprise workspaces rather than curated ones. The new Adaptive Instructed-Retriever is built to decide the search depth before execution begins. Databricks states that the model matches the quality of Claude Sonnet 5, GPT-5.6 Luna, and DeepSeek-V4-Flash while answering more than twice as fast. The average response time reported in the research is 5.8 seconds. These figures come from Databricks’ internal testing and have not been independently verified by third parties.
Breaking Away from Fixed Query Plans
To understand the shift, it helps to look at how traditional databases work. A standard query plan is a fixed set of steps a database follows to execute a request, such as which index to scan or the order to join tables. The same query always produces the same plan. Michael Bendersky, research director at Databricks, used this comparison to explain the new architecture.
“The query plan setup was deterministic and rule based,” Bendersky said. The new system shares the goal of speed with a query plan but breaks from its rigid structure. “It’s different because the flow here is actually non-deterministic, meaning an agent can take one path for one query and a very different path for a different query,” he explained. This flexibility is critical for multi-hop questions, where an answer requires more than one round of retrieval. For example, finding revenue figures scattered across two documents requires pulling a pointer from the first file to issue a second, specific search.
In the past, specialized retrieval models often lagged behind general-purpose frontier models in both accuracy and flexibility. By training specifically for the search task, Databricks appears to be narrowing that gap significantly. The approach suggests that for narrow, high-stakes tasks like document retrieval, a smaller, specialized agent can outperform larger, generalist systems that are not optimized for step efficiency. This aligns with a broader trend where enterprises prioritize specific workflow efficiency over general model capability.
Related Post: Masie Becomes ABC CEO Amid Exec Fine Battle
Training for Efficiency with Step Penalties
The research describes two distinct retrieval modes that the model uses. In parallel thinking, the system issues several rewritten versions of a query simultaneously, merging the results at the agent level. In sequential thinking, if the initial documents contain a clue but not the full answer, the model uses that information to issue a second round of queries. This process repeats up to a fixed step limit.
Central to the training is a technique called CISPO, or Clipped Importance Sampling Policy Optimization. Using online reinforcement learning, Databricks trained the model with a specific penalty for taking extra steps. “We train the agent to search more, but only when needed,” Bendersky said. The system balances the time spent searching against the reward of finding the right documents. This creates a family of model checkpoints, each landing at a different point on a quality-latency curve.
In one benchmark example from the research, the model matched Sonnet’s two step search on a customer account question while returning 50% higher recall.
Context Gaps Remain a Major Enterprise Issue
The improved speed of retrieval does not automatically solve the problem of incorrect business context.
Response correctness is still the top metric enterprises use to grade agents. However, buying criteria are shifting toward access control and ingestion ease rather than pure retrieval accuracy. How many steps a system takes to reach an answer is one variable. Whether that answer rests on consistently defined business context is a separate, distinct challenge. Databricks frames its results as evidence that a small, specialized model can match frontier accuracy on narrow search tasks without requiring a larger general-purpose model.
The company notes that the initial Instructed Retriever work was more similar to a fixed query plan because the structure of how the agent called the retrieval system was fixed. The new work moves to a more adaptive approach where the agent adapts to the question and changes the plan based on the specific input. This shift allows the system to handle the variability of enterprise data more effectively than previous rigid architectures.
