Within about four days of each other, Meta added enterprise software connectors to its Muse agent as part of a business-focused launch (reported Sept 29–30 by Axios and others), and Progress Software expanded its agentic RAG platform with a Microsoft Teams app, a "Smart Agent," and a WordPress plugin (SD Times, Oct 2). Different companies, different market tiers, same move: neither announcement is about retrieval quality. Both are about where the data comes from and where the answer gets rendered.
That's worth sitting with if you build these systems. For most of the last two years the interesting work in a knowledge stack was in the middle — chunking strategy, embedding choice, hybrid sparse/dense fusion, reranking, query decomposition. That middle has flattened. It hasn't stopped mattering, but the gap between a competent implementation and a state-of-the-art one is now smaller than the gap between a system wired into Slack, the CRM, the ticketing system, and the CMS, and one that isn't. Coverage beats cleverness when the baseline failure mode is "the document wasn't in the index."
What this costs you architecturally
Connectors are where retrieval systems accumulate their real complexity, and it's operational rather than algorithmic. Each one brings a permission model you have to mirror, or you leak. ACL-aware retrieval is the unglamorous hard problem: you either filter at query time against a live permissions check (correct, slow, and dependent on the source system's uptime) or you denormalize permissions into the index at sync time (fast, and wrong the moment someone's access changes). Most teams do the second and discover the staleness window during an audit.
Each connector also brings its own change-detection semantics. Some sources give you a reliable cursor; some give you a modified timestamp that lies; some give you nothing and you re-crawl. Incremental sync correctness, not embedding quality, is what determines whether your answers are current.
And the surfacing side — Teams apps, chat plugins, CMS integrations — changes retrieval requirements, not just UI. A query arriving inside a Teams thread has conversational context, a channel scope, and a participant list, all of which are retrieval signals you now have to decide whether to use.
Who should care
If you're building internal knowledge tooling, the competitive read is that platform vendors are converging on connector breadth as the differentiator, which means your build-vs-buy calculus should be run on integration maintenance cost, not retrieval benchmark scores. If you're on the model side, note the counterweight: Aleph Alpha published Kolibri's full weights to Hugging Face under Apache 2.0 this week. Open weights keep the generation layer substitutable. The connector layer is what actually locks you in.