Retrieval-augmented generation has a precise birthday. In 2020 Patrick Lewis and colleagues, most of them at Facebook AI Research, published the paper that gave the technique its name: before the model answers, fetch passages relevant to the question and place them in front of it.
The idea was simple enough to implement in an afternoon and useful enough that, five years later, most company assistants run on some version of it. So why do the same companies now talk about a context layer as though retrieval had left something out?
It did leave something out, though that is no criticism of retrieval. The two answer different kinds of question.
What does RAG do well, and where does it stop?
A RAG system cuts your documents into passages, stores a numerical fingerprint of each, and at question time returns the passages whose fingerprints sit closest to the question's. The model then writes its answer from those passages. For a question whose answer someone has already written down, this works very well. "What does our travel policy say about rail?" has an answer in one paragraph of one document, and retrieval will find it.
The method weakens as the answer moves away from a single passage. Ask how many projects are behind schedule and there is no paragraph to find: the answer is a count across records. Ask whether the Q4 launch is on track and the passages that resemble the question are status reports, each of which says yes. Retrieval finds text that looks like an answer. It does not check that text against the tickets.
What does enterprise search add?
Enterprise search products, Glean and Microsoft 365 Copilot among them, take the same principle and do the hard engineering around it. They connect to dozens of company tools, respect each user's access rights, and rank results using what they know about people and teams. The result is a clear improvement: one question reaches the wiki, the mail archive, the chat history and the ticket system together.
However, the unit is still the document or the message. Enterprise search can tell you what your company wrote about the launch, in every tool, sorted by relevance. Whether the launch is in trouble is a judgment that someone has to assemble from forty results, and the agent assembles it afresh, and somewhat differently, on every run. We describe that instability in Why AI Agents Get Your Company Context Wrong.
What does a context layer hold that search does not?
An AI context layer starts from the other end. Where search indexes text and waits for a question, a context layer maintains a model of the business in which the things that matter are objects with links between them, and it keeps the state of each object current as activity arrives from your tools.
| RAG | Enterprise search | Context layer | |
|---|---|---|---|
| Holds | Passages from documents you loaded | Documents and messages across company tools | Objects, links and their current state |
| Answers well | "What does the policy say?" | "Where did we discuss this?" | "Is this on track, and what is blocking it?" |
| Stays current through | Re-indexing | Connectors | Analysis of work activity |
| Same answer on each run | Depends on phrasing | Depends on phrasing | Yes, until the facts change |
Vindaris is a context layer of this kind for strategy execution. Its Work Graph holds the strategy with the goals, KPIs, projects and tasks under it, and it analyzes work activity across the tools your teams use, such as Jira, Slack, Microsoft 365 and meeting transcripts. A project in the graph has an observed state that Vindaris derives from that activity.
When an engineer says in a thread that the work is waiting on a vendor, the project becomes blocked in the graph that day, and the people who own it hear about it weeks before the next review would have shown it. The published schema lists the object types and the alerts.
Which one do you need?
Arguably the comparison in the title is unfair to all three, because a serious setup uses more than one. Keep retrieval for the large body of written knowledge that changes slowly: policies, contracts, product documentation, past decisions. An agent that drafts a customer reply needs exactly that.
Add a context layer for the questions about the plan, which are the ones executives ask most and trust the answers to least. An agent can use both in one session. Over MCP it can ask Vindaris which projects under a goal are stuck, as shown on the AI agents page, and then search the document store for the contract that explains why.
Ask of any question whether the correct answer would be different tomorrow morning even if nobody wrote a single new document today. If so, retrieval alone will keep handing your agent yesterday's truth, well ranked.
How many of the questions you put to your AI last week were of that second kind?
FAQ
Is a context layer a replacement for RAG? No. RAG suits questions whose answer someone already wrote down, such as a policy or a contract clause. A context layer suits questions about current state, such as whether a project is blocked. Most companies use both.
Is a context layer the same as a knowledge graph? A context layer is usually built on a graph, but it adds upkeep: it derives the current state of each object from activity in your tools. A static knowledge graph records relationships and leaves the state to whoever maintains it.
Where does Glean or Microsoft 365 Copilot fit? Both are enterprise search and assistant products that retrieve documents and messages across company tools with permissions applied. They are built to find and summarize what your company wrote. A context layer for strategy execution is built to hold the plan and derive each project's state from work activity.
Can an agent use retrieval and a context layer together? Yes. Over MCP an agent can ask Vindaris which projects under a goal are stuck and then search the document store for the contract or decision that explains why.