Demonstrate-Search-Predict: Composing Retrieval and Language Models for Knowledge-Intensive NLP
Description
In this talk, you will learn about how retrieval-augmented in-context learning has emerged as a powerful approach for addressing knowledge intensive tasks using frozen language models (LM) and retrieval models (RM). Existing work has combined these in simple “retrieve-then-read” pipelines in which the RM retrieves passages that are inserted into the LM prompt. To begin to fully realize the potential of frozen LMs and RMs, we propose Demonstrate–Search–Predict (DSP), a framework that relies on passing natural language texts in sophisticated pipelines between an LM and an RM. DSP can express high-level programs that bootstrap pipeline-aware demonstrations, search for relevant passages, and generate grounded predictions, systematically breaking down problems into small transformations that the LM and RM can handle more reliably. We have written novel DSP programs for answering questions in open-domain, multi-hop, and conversational settings, establishing in early evaluations new state-of-the-art in-context learning results and delivering 37–125%, 8–40%, and 80–290% relative gains against vanilla LMs, a standard retrieve-then-read pipeline, and a contemporaneous self-ask pipeline, re…
Description from YouTube. Full content on the video page.
More from Databricks
ReleasesDatabricks launches across the Data + AI stack in 90 seconds
Databricks announced LTAP to unify lakebased and lakehouse data, eliminating ETL and enabling a single copy of data for analytical and operational needs. They also introduced Unity AI Gateway for governance, Genie Ontology for enterprise knowledge graphs, and open-sourced Omniant for managing multiple coding agents.
ReleasesIntroducing Omnigent: The Ultimate Meta-Harness for AI Agents
Omnigent is a new open-source meta-harness for AI agents that provides a unified interface for composition, control, and collaboration across multiple models and agent workflows. It enables stateful, data-centric policies for guardrails and allows real-time sharing and steering of live agent sessions with teammates.
NewsHow DEFRA and Natural England Accelerate Peatland Restoration with AI and Databricks
DEFRA and Natural England utilize AI and Databricks to accelerate peatland restoration by automating the mapping of peatland features and peat dams across England. This technology significantly reduces the time required for mapping, enabling faster identification and restoration of these crucial carbon-storing habitats.
NewsAI Stack Explained in 3 Layers (LLM, Agent Harness, Omnigent)
The AI stack now includes a third layer, the meta harness, which sits above individual agent harnesses. This meta harness, exemplified by Databricks' open-sourced Omnigent, allows for routing queries to appropriate agents and orchestrating tasks across multiple agents, enabling seamless interaction and context sharing between them.
NewsWhat’s coming next to Free Edition
Databricks announces the availability of Genie, GPUs, Agent Hooks, Lakehouse, and Lake Flow Designer on its Free Edition. This update provides virtually all of Databricks' production platform features for free, enabling users to learn and build data and AI projects.