AI Workbench

AI Workbench is an open-source, self-hosted interface for teams building retrieval-backed AI applications on DataStax Astra. It brings the operational pieces into one product surface: workspaces, knowledge bases, document ingest, chunking, embedding and reranking services, API keys, retrieval experiments, agents, and MCP.

The public repository describes the TypeScript runtime as the production ship path today. It bundles the UI and implements the full /api/v1/* contract in one Docker image. Python and Java runtimes are present as preview scaffolds so the cross-runtime API contract can be tested as those implementations mature.

Product surface

The core model is workspace-first. A workspace owns the knowledge bases, documents, jobs, credentials, services, agents, and API keys that belong together. A knowledge base owns its Astra collection end to end, including the chunking, embedding, and optional reranking services that feed it.

That gives the UI a practical shape:

• create or attach Astra-backed knowledge bases;
• ingest raw text or files and track synchronous or asynchronous job state;
• run text, vector, hybrid, and reranked retrieval in the browser;
• configure workspace-scoped services and API keys;
• define retrieval-grounded agents and expose the workspace through MCP when enabled.

What I work on

My contributions have spanned the runtime, UI, test harness, and production hardening work. Recent public PRs include file ingest support (#192, #193), authentication and observability work (#196, #197), conformance coverage (#198), and control-plane cleanup (#199).

The work is close to the kind of developer product I like: a clear UI, a stable HTTP contract, direct Astra integration, and enough conformance testing that the implementation can keep evolving without surprising its users.