Knowledge infrastructure

senior-mcp

319k chunks of specialized knowledge behind an MCP server in production.

Active

The problem

Enterprise AI systems need to answer from specific knowledge — processes, tables, rules, support history. Without a unified index, every agent re-scans on its own, misses recency and hallucinates vocabulary.

The core

senior-mcp indexes structured knowledge (HCM/ERP data dictionaries, screens, tables) and unstructured knowledge (tickets, forum) in a single PostgreSQL base with pgvector. Search combines three stages:

  1. Vector (75%) — semantic similarity over embeddings.
  2. Text (25%) — term matching, catching the exact vocabulary embeddings miss.
  3. Rerank — the top-25 goes through a reranker (Voyage) before returning to the agent.

The catalog

0 indexed chunks 261,803 HCM/ERP · 46,470 LSP · 10,267 tickets · screens/tables/forum
+5.4pp MRR@10 gain from reranking measured over pure hybrid
0 MCP tools exposed search, chunk, catalog, stats and source tracing

Access and integration

The service exposes 8 MCP tools — among them senior_search (hybrid search with rerank), senior_get_chunk (full context of a result), senior_catalog and senior_stats. It runs on Node.js behind a Cloudflare Tunnel (seniormcp.gbbragadev.com) with token authentication — secrets never reach the client.

Real consumption

Coding agents (Claude Code, Codex) use senior-mcp as a grounding tool daily — the query becomes part of the agent’s flow, not a documentation tab. In Braga Suite, it is the knowledge layer designed into the BFF, server-side and allowlist-gated — live querying is under controlled validation.

My role

Hybrid indexing concept, rerank architecture, PostgreSQL/pgvector integration, MCP contract design and service operation.

senior-mcp is independent infrastructure. Brands mentioned belong to their respective owners.

Want to build something like this?

No strings attached: tell me the context and I'll tell you what can be proven first.

Start a conversation