ReleaseBytes
databricks Databricks Blog ·

Databricks details Adaptive Instructed-Retriever for efficient AI search

blogaidatabricksengineer
announcement

Databricks details its Adaptive Instructed-Retriever, a model designed to optimize search for enterprise data agents by combining parallel and sequential search. It adaptively decides computation requirements for user requests, delivering frontier-quality retrieval with 2x lower latency than leading models for complex, multi-hop questions. The post explains the training methodology using online reinforcement learning and its applicability for customers developing specialized models. This approach allows for customizable quality-latency tradeoffs to fit various production workloads.

  • →Adaptive Retrieval for Enterprise AI Agents
  • →Achieving Frontier-Quality Search at 2x Lower Latency
  • →Customizable Quality-Latency Tradeoffs with ORL Training
Notes (3) ›
  • Adaptive Retrieval for Enterprise AI Agents

    The post explains Adaptive Instructed-Retriever, designed for enterprise data agents, which combines parallel single-step retrieval and sequential search to balance accuracy and speed for complex, multi-hop questions.

  • Achieving Frontier-Quality Search at 2x Lower Latency

    The model showcases performance comparable to leading third-party and open-source models while achieving 2x lower latency, significantly improving retrieval quality over single-step search on benchmarks.

  • Customizable Quality-Latency Tradeoffs with ORL Training

    The article describes the training process using online reinforcement learning (ORL) with a reward design that balances trajectory quality and search cost, allowing for a customizable quality-latency tradeoff. This approach, trainable with AI Runtime (AIR), is practical for customers developing specialized models for their domains.

Read the original announcement →

https://www.databricks.com/blog/adaptive-instructed-retriever-frontier-quality-search-2x-lower-latency

Related releases