In short By default Claude knows only what it absorbed in training and sees only the current conversation (model knowledge and hallucinations, working memory and context). Three mechanisms extend its reach to the real world: Connectors plug in your apps, Enterprise Search your company's knowledge, Research deep gathering across many sources, including the web.
The problem: the model is "boxed in"
On its own the model doesn't know what happened after its cut-off; doesn't see your mail, drive, calendar, databases; doesn't go online. For Claude to work with real and fresh data, it needs access — which is what the three mechanisms provide.
Connectors
A connector plugs Claude into a specific app — mail, cloud drive, calendar, task tracker, knowledge base. Once connected, Claude can read from it (and, with permissions, act). Example: "find the contractor's email and build the estimate from the attachment." Analogy: giving an assistant keys to the right cabinets — now they fetch documents themselves. Under the hood connectors are often built as MCP servers (MCP).
Enterprise Search — search over company knowledge
Enterprise Search lets Claude search an organisation's internal knowledge across many connected systems at once (wikis, drives, documents). The answer is built on your corporate data, not the model's general knowledge. Example: "what's our leave policy?" — Claude finds the answer in the internal base rather than inventing it. 🔴 Needs a Team/Enterprise plan.
Research — a deep-dive mode
Research is a mode where Claude itself runs a multi-step investigation: forms sub-questions, searches many sources (including the web), cross-checks and assembles a structured, cited write-up. Slower than an ordinary answer, but deeper and grounded in sources. Analogy: not "ask an erudite from memory," but "send an analyst to compile a dossier."
Why this cures hallucinations
All three realise one principle: give the model facts/a tool, don't rely on its memory. It's a direct cure for made-up answers (model knowledge and hallucinations) — Claude replies from real data it sees now.
| Mechanism | What it gives | Example |
|---|---|---|
| Connectors | access to your apps | mail, drive, calendar |
| Enterprise Search | search over company knowledge | internal wiki, policies |
| Research | deep dive across many sources + web | comparative analysis with citations |
Related
The source is the Claude 101 course. Related: model knowledge and hallucinations, MCP, steerability.