The notable release this week was quiet and not a model. Progress Software announced an expansion of its agentic RAG platform, reported by SD Times on Tuesday, framed explicitly around connecting enterprise knowledge across business systems. The accompanying distribution additions — a Microsoft Teams app, a WordPress plugin, and a packaged "Smart Agent" — tell you more about where this market is going than the retrieval claims do.
Read the shape of it rather than the feature list. Two years ago an enterprise RAG pitch was about chunking strategy, embedding model, reranker, maybe a hybrid BM25 stage. None of that is a differentiator now. Open-weight embedding models are good enough, hybrid retrieval is table stakes, and reranking is a line item. What remains genuinely hard is the part nobody demos: getting authoritative, permission-aware, freshly-synced content out of the eleven systems where your company actually keeps its knowledge, and putting answers in the surface where people already are.
That's the bet here. A Teams app is not a feature, it's an admission that a standalone search UI loses to the chat window that's already open. A WordPress plugin is an admission that a meaningful slice of "enterprise knowledge" is public marketing content that someone needs to keep in sync without an integration project.
If you're building internally, the lesson transfers whether or not you'd ever buy this. The expensive work in your knowledge stack is connector maintenance and ACL propagation, and it's recurring cost, not one-time. Every source system has its own deletion semantics, its own idea of what a permission grant means, and its own rate limits that determine whether your index is minutes or days stale. A retrieval pipeline that returns a document the asker shouldn't see is worse than one that returns nothing. Budget accordingly: the team that owns ingestion should be at least as large as the team tuning the retriever.
The structural tradeoff is familiar. Buy the connector layer and you inherit someone else's refresh cadence, permission model, and roadmap — plus a dependency on their coverage of your specific SaaS footprint. Build it and you own a permanent integration treadmill that produces no visible product value until it breaks. Most teams I'd expect to end up hybrid: vendor connectors for the long tail of commodity SaaS, hand-built ingestion for the two or three systems where your actual proprietary knowledge lives and where permission semantics are load-bearing.
One caveat on the announcement itself: vendor release notes at this level of detail don't establish how sync, incremental reindexing, or permission inheritance actually work. Ask for those specifics before drawing conclusions — the architecture question is whether permissions are evaluated at query time against the source, or snapshotted at index time, and that distinction is usually buried.