Databricks explains how workplace documents feed AI agents

Databricks’ October 26 overview describes bringing scattered workplace knowledge into governed tables with managed Lakeflow Connect connectors. For teams building internal AI agents, this provides a common starting point for document preparation. The company distinguishes incoming text from raw files: SharePoint documents and Google Drive files need text extraction before their contents can be indexed for retrieval. Loading those files is therefore one stage of preparing an agent’s knowledge.

The SharePoint connector documentation for AWS makes the access sequence concrete. In the multi-user workflow, an administrator configures authentication and creates a connection; a non-admin then creates a pipeline using that connection. Supported authentication uses OAuth, rather than basic username/password or API-key authentication. The connector can load unstructured files as binary content, turn structured formats into Delta tables, or capture metadata without loading file contents. The pipeline schedule depends on latency and cost requirements.

The downstream processing guide starts after the SharePoint pipeline exists. Its ai_parse_document function extracts text, tables, layout information and metadata from binary documents. The example applies it to content.inline_content in a materialized view, incrementally parsing newly arrived files. This produces structured output for subsequent document processing. A separate FILE-type path passes a governed file reference directly to the function without materializing its bytes; that path is explicitly Beta. It is an alternative input representation, not a required replacement for binary ingestion.