Manage knowledge bases
The Knowledge page is the central place to manage the knowledge bases your flows use for retrieval-augmented generation (RAG). From here you can browse every knowledge base in your workspace, inspect how it was built, add new sources, and drill into the underlying chunks.
For the flow component that reads from and writes to a knowledge base, see Knowledge Base component.
Browse knowledge bases
- In the Orbis Studio sidebar, open the Knowledge page.
- Use the search box to filter by name.
- Select one or more knowledge bases. Hold Shift to select a range.
Each knowledge base lists key statistics so you can understand its size and configuration at a glance:
| Field | Description |
|---|---|
| Embedding provider / model | The provider and embedding model used to vectorize the content |
| Size | Total storage used by the knowledge base |
| Words / characters | Volume of ingested text |
| Chunks | Number of chunks the content was split into |
| Avg chunk size | Average size of each chunk |
| Chunk size / overlap / separator | The splitting configuration used during ingestion |
| Source types | The kinds of sources ingested (for example, files or tabular data) |
| Status | Whether the knowledge base is ready, processing, or failed |
If a knowledge base fails to build, the page shows the failure reason so you can correct the source or configuration and retry.
Inspect sources and chunks
Open a knowledge base to view its details in a side drawer. From there you can review the individual sources and open the chunks view to see exactly how the content was split and what will be retrieved at query time. This is useful for debugging retrieval quality — if an agent returns poor answers, the chunks view shows whether the relevant text was captured and how it was segmented.
Add content
Upload new sources to a knowledge base to expand what it can answer. Orbis Studio embeds the new content with the knowledge base's configured embedding model and adds it to the existing index, so you can grow a knowledge base over time without rebuilding it from scratch.
For column-based (tabular) sources, each column can be configured independently — you choose which columns are vectorized for semantic search and which serve as identifiers.