Architecture center

Patterns and best practices for designing document intelligence pipelines with Bundata.

Core architecture

Bundata acts as a document intelligence layer between your raw documents and your applications. Typical flow:

1. Ingest — Documents enter via source connectors (S3, SharePoint, Confluence, etc.) or direct upload.
2. Extract — Context-aware extraction produces schema-aware smart bites with metadata and optional enrichment (NER, image/table descriptions).
3. Chunk and embed — Content is chunked and optionally embedded for retrieval.
4. Deliver — Output goes to the Vector Catalog, your vector store, or downstream systems. Source lineage is preserved for grounded answers and audit.

Key patterns

• Schema-first design — Define schemas in Schema Studio before scaling extraction. See Schema design and Extraction best practices.
• Workflow orchestration — Use Workflows for scheduled runs, triggers, and retries so the Vector Catalog stays fresh. See Workflows overview and Triggers & scheduling.
• Grounding for agents — Query the Vector Catalog for semantic retrieval and pass chunks to LLMs with source lineage. See Grounding from catalog and Vector Search overview.

Use-case guides

• First extraction — End-to-end extraction run
• First agent — Agent grounded on the catalog
• Workflows — Automation and scheduling
• Integrations — Connectors and cloud

Reference

• Glossary — Smart bites, Vector Catalog, extraction confidence, source lineage
• Limits and quotas — Rate limits and capacity