Some of an organization’s most valuable data remains hard to reach. Documents, slides, PDFs, and internal systems hold important information, yet teams spend more time finding, cross-checking, and stitching together references than making decisions. Traditional retrieval-augmented generation (RAG) required complex parsing, chunking, and embedding of unstructured data, plus constant tuning to maintain quality — overhead that kept many teams and systems locked out of their own knowledge.
Databricks’ Agent Bricks Knowledge Assistant aims to change that. Now generally available and expanded to 10 new regions, it turns documents into accurate, grounded answers in minutes. As part of the Agent Bricks platform, it offers a fully managed experience across the agent lifecycle — ingestion, continuous updates, retrieval, and inference — behind a scalable endpoint. Databricks reports that it achieves up to 70% higher answer quality than a simplified RAG approach, without the operational burden.
A different approach to enterprise retrieval
Most retrieval systems were built for a simple world: a single index, a uniform schema, and similarity search as the primary signal. Enterprise knowledge does not look like that — it spans multiple systems, each with different structures, metadata conventions, and expectations about freshness and authority. Treating them all as one homogeneous source limits both quality and control.
Knowledge Assistant instead uses what Databricks calls an “instructed retriever,” developed by its AI research team. Rather than relying on similarity search alone, the assistant learns how each source is organized and how it should be queried. When a user asks a question, it translates the request into precise, source-aware sub-queries — prioritizing recent content, emphasizing specific metadata, or favoring authoritative sources — then synthesizes across systems while staying grounded in each one’s structure. The result includes page-level citations on every response, reducing hallucinations and letting users jump straight back to the source. Because it is delivered as a managed service, teams inherit ongoing research improvements automatically: Databricks continuously evaluates new models and techniques against its evaluation suite and folds the winners into deployed agents. The general RAG pattern this builds on is illustrated in this production RAG case study.
Connecting subject-matter experts and developers
Most enterprise AI systems rely on static labels or coarse feedback that capture individual answers but fail to generalize. Knowledge Assistant instead uses Agent Learning from Human Feedback (ALHF) to turn expert guidance into durable, repeatable behavior changes. Rather than one-off corrections, ALHF generalizes natural-language feedback across a conversation: experts supply questions and guidelines for good answers, and the assistant applies them broadly. Every response is captured in an end-to-end trace and logged as structured, governed data, and with native MLflow integration teams can evaluate changes and track quality with the same rigor used for production ML systems.
One early user, Ben Bartholik, principal data engineer at FinThrive, described Agent Bricks as turning documentation “into conversations, an always-available AI teammate” that helps the team move faster and focus on core work. The organizational discipline behind deploying agents like these is explored in this summary of enterprise AI agent trends.
Getting started
Setup is designed to be simple: upload documents and the assistant handles ingestion, retrieval, and updates. Typical uses include agents over market research, support documentation, or policies and procedures.
Limitations and what to watch
- The 70% quality figure is Databricks’ own, measured against a “simplistic RAG” baseline it defined; real-world gains depend heavily on document quality, domain, and how the baseline is configured.
- A fully managed service reduces overhead but increases dependence on a single vendor’s platform and pricing; portability of the resulting agents is worth evaluating up front.
- Page-level citations reduce but do not eliminate hallucination risk; answers on high-stakes topics still warrant human verification against sources.
- Feedback-driven systems can encode the biases of the experts who train them, so governance of who provides guidance matters as much as the tooling.
