Two things stood out this week if you build production RAG and agent infrastructure:
- Microsoft added Knowledge Graphs and
Microsoft-Retriever-1to Foundry IQ (InfoWorld, Nov 19), specifically to stop developers from forcing large frontier models to plan multi-hop retrieval loops over flat vector chunks. - Starburst launched Context Studio (Nov 18), exposing governed semantic context over MCP so agents don't hallucinate conflicting business definitions.
Both validate the same shift: vector search alone isn't enough, and stuffing everything into a frontier LLM prompt is too slow and expensive.
At the same time, building an open-source, self-hosted alternative at PipesHub (https://github.com/pipeshub-ai/pipeshub-ai) has taught us a few practical lessons that vendor announcements usually gloss over:
- Knowledge graphs shouldn't rely purely on expensive LLM triplet extraction: Automated entity extraction over messy internal jargon creates noisy edges and huge indexing bills. Combining deterministic structural edges (document -> section/block -> linked Jira ticket / GitHub PR / Slack thread) with selective entity linking gives you multi-hop grounding and traceable citations without unpredictable cloud indexing fees.
- Agent identity != human user identity: When an autonomous agent queries your context layer in the background, treating its token like a generic admin key breaks both auditability and billing. This week we implemented explicit service-account token marking (PR #6048) so RBAC and org-level credit/usage tracking work cleanly for non-human actors.
- MCP OAuth Dynamic Client Registration (DCR) is still rocky in the wild: While expanding our MCP server catalog (adding Miro today and testing Dropbox's official MCP server), we hit a classic real-world spec gap: Dropbox's
.well-knowndiscovery endpoint advertises a registration URL, yetPOST /oauth2/registerreturns403 Forbidden(dcr_failed). Supporting pre-registered OAuth clients alongside CIMD/DCR is mandatory if you want MCP integrations to actually work in production.
If you want a self-hosted, Apache 2.0 context layer (Qdrant + Neo4j/ArangoDB + 50+ connectors + MCP + Python/TS/Go SDKs) that ranked #1 on EnterpriseRAG-Bench (93.24) without Azure or lakehouse lock-in, check out the repo:
For those running self-hosted MCP servers in production—how are you handling OAuth registration and service-account permissions?
