Case study · 04 · 2025 — Now
Knowledge & MCP Platform
A central MCP catalog and multi-source knowledge pipeline for enterprise agents, with semantic processing and hybrid retrieval.
Architect & R&D Lead
MCPLangChainElasticsearchpgvectorCelery
Impact
- Confluence, Jira, and document sources
- Vector and keyword retrieval
- Shared foundation for multiple RAG use cases
Context
- Every RAG use case rebuilt its own ingestion, indexing, and retrieval — with inconsistent permissions and freshness.
- Agents needed one catalog of tools and knowledge instead of scattered endpoints.
Approach
- Built a central MCP catalog exposing governed tools to enterprise agents.
- Created a multi-source pipeline (Confluence, Jira, documents) with semantic processing.
- Combined vector and keyword retrieval (hybrid search) with permission-aware chunks.
- Made the pipeline replayable so every indexed chunk can be explained and reproduced.
Lessons
- Corpus quality caps retrieval quality — invest in ingestion before the vector database.
- Permissions must travel with chunks, not sit in a separate layer.
- A shared knowledge foundation compounds across use cases.