The more interesting retrieval news this week wasn't a model or a vector index. It was a publisher deciding that the right delivery format for its corpus is an MCP server.
O'Reilly announced Expert Intelligence on October 7, an organization-wide suite that grounds enterprise AI in its repository of practitioner knowledge and calls on job-specific agent skills, with responses verified through citations from named authors. The product surface is an "Expert MCP" that pulls context from the knowledge repository, plus agent skills, and it connects with any MCP-compatible tool. The pitch is that it lives inside the AI tools organizations already use, so there's no separate platform to roll out.
The attribution argument
The framing is worth reading closely because it's a retrieval-quality argument, not a content-licensing one. O'Reilly cites a February 2026 benchmark across 60 questions drawn from live production traffic, finding roughly one in five claims from leading AI tools unsupported by a cited source, about two-thirds of cited sources naming no author, and around half carrying no publication date. The stated fix is a direct, secure connection to the curated corpus so every response traces to a named expert and a confirmed publication date.
If you build RAG, you already know why author and date are load-bearing. A chunk from a 2019 blog post and a chunk from a current reference manual look identical in embedding space. Recency and authorship are metadata problems, and most ingestion pipelines either discard those fields during chunking or never had them because the source was scraped HTML. A curated corpus with enforced provenance fields is less a knowledge advantage than a metadata-integrity advantage — which is the part teams consistently underinvest in.
Why MCP is the right-sized abstraction here
Shipping as an MCP server rather than a bulk export or an embeddings dump is the real design choice. The content never lands in your index, so licensing stays enforceable per-query, and the retrieval logic stays with the party that understands the corpus structure. The tradeoff is familiar: you give up control over chunking, ranking, and hybrid-search tuning, and you add a network hop with someone else's latency and uptime. For a reference corpus you don't own, that's usually the correct trade. For your own Confluence and ticket history, it isn't.
The surrounding evidence
Context from October 5: MIT Technology Review Insights published a Neo4j-sponsored report surveying 300 data, AI, and technology executives on their ability to give agents full contextual understanding across semantic knowledge, episodic memory, and procedural knowledge. It reports only 34% of agentic AI projects reaching production, with data fragmentation and missing context as major blockers, and higher production rates among organizations with stronger semantic knowledge capabilities. It's vendor-sponsored, so treat the numbers as directional. The stated investment priorities — ingestion pipelines, AI-ready APIs, RAG, evaluation agents, and knowledge graphs — still describe where the work actually is.
Who should care: anyone whose agents cite sources users will check. The provenance layer is becoming a product boundary, and external corpora are arriving as protocol endpoints rather than data dumps.