1. RAG is not magic search
Retrieval-augmented generation combines document retrieval with generated answers. The quality of the response depends on what was retrieved, how documents were chunked, whether sources are current and how the answer is reviewed.
Poor documents create poor answers, even with a strong model. Prepare the sources before indexing anything.
- Prepare and de-duplicate documents before indexing them.
- Remove obsolete files and ambiguous drafts.
- Keep document titles, dates and owners visible.
- Prefer answers that cite retrieved sources.
- Treat missing answers as documentation gaps.
2. RAG or tool-use and MCP: choose by the question
Not all context lives in documents. Many questions depend on data that changes constantly: order status, balance, availability, an open ticket. Indexing a document does not solve that; the answer is to give the AI a tool that queries the live system. That is tool-use, and the Model Context Protocol (MCP) standardizes how models connect to those tools and data sources.
An AI that queries systems through tools inherits every integration concern plus AI-specific risk. Permissions, logging and action limits still apply.
- Use RAG when the answer is in relatively stable text: manuals, policies, procedures.
- Use tool-use or MCP when the answer depends on live system data.
- Many solutions combine both: RAG for knowledge, tools for current data.
- Treat each tool as a door with its own permission.
- Never index or expose secrets, credentials and private keys.
3. Permissions, boundaries and a feedback loop
A company knowledge base may include HR, finance, customer, technical and operational content. Not every user should retrieve every document, and not every tool should reach every system. The OWASP Top 10 for LLM Applications 2025 highlights prompt injection, sensitive-data disclosure and excessive agency.
The first version will expose missing or confusing documentation. That is useful: each failed answer becomes a source or tool improvement task.
- Classify collections and apply role-based access to retrieval and tools.
- Log high-impact usage for audit and improvement.
- Separate internal assistant use from public website content.
- Collect unanswered or low-confidence questions and update sources.
- Measure whether users resolve tasks faster with the assistant.
How to use this article
Treat this page as a decision aid. Use it with the related hub, checklist or service route when the topic affects production, customer experience, deployment, security or business continuity.