Built a Self-Hosted Knowledge System with Cited Retrieval
Conceived and built VaultMind from scratch, combining multi-source ingestion, scoped vector retrieval, cited answers, and read-only agent access in one self-hosted application.
Context
Useful knowledge is often split across source repositories, workspace tools, drive files, and authored documents. VaultMind creates one inspectable retrieval layer while keeping deployment and source control with the operator.
What I did
- Defined the product direction, system boundaries, technical architecture, data model, interface, testing strategy, and public release.
- Built source adapters, synchronization workflows, Markdown-aware chunking, embeddings, scoped pgvector retrieval, ranking, persisted citations, and evidence-aware answer generation.
- Created knowledge spaces and a normalized document boundary so different providers feed the same indexing and retrieval pipeline.
- Added a bounded read-only MCP interface that makes the indexed knowledge available to agent workflows without exposing mutation operations.
- Implemented credential protection, prompt-injection handling, synchronization leases, input limits, deterministic demo data, automated tests, and smoke checks.
Outcomes
- Published a complete MIT-licensed implementation that can be self-hosted and inspected end to end.
- Made one cited knowledge index available through both the application and a read-only MCP interface.