RAG
RAG stands for Retrieval-Augmented Generation. Behind the acronym is a simple idea: look it up before you answer.
The problem it solves
An AI model knows what it was trained on — a huge amount of general knowledge, frozen at some point in the past. It knows nothing about your organisation. It has never seen your case files, your policies, or last week's decision.
Ask it something specific to you and it does the worst possible thing: it answers confidently and wrongly. It has no way to know it does not know.
RAG fixes this by changing the order of operations. Instead of ask → answer, it becomes:
- Retrieve — search your files and records for material relevant to the question.
- Augment — put what it found in front of the model along with the question.
- Generate — answer from that material.
The answer is grounded in your actual documents rather than the model's recollection. Which is exactly what you would want from a colleague: check the file, then speak.
RAG vs context vs memory
All three put material in front of the model. The difference is when and who chooses:
| Who chooses | When | |
|---|---|---|
| Context | You, in advance | Attached to the agent, always present |
| RAG | The system, per question | Looked up fresh, based on what was asked |
| Memory | The agent, from experience | Accumulated over time |
Context is the binder you hand over. RAG is the filing cabinet they walk to. Memory is what they picked up along the way.
Use context when the agent always needs the material. Use RAG when the relevant material depends on the question and there is far too much to hand over up front.
Configuring RAG on an agent
RAG is off by default, and enabling it on an agent gives you a few choices:
| Setting | What it decides |
|---|---|
| Enable RAG | Whether the agent looks things up at all. |
| Search mode | Keyword matches words literally; semantic matches meaning; hybrid does both. |
| Number of sources | How many results to pull in. More is not better — each one spends budget. |
| Include files | Search your Nextcloud files. |
| Include objects | Search your OpenRegister records. |
Search mode is the one worth understanding. Keyword search finds "permit" when you write "permit". Semantic search finds a document about permits when you ask about "planning applications", because it matches on meaning. Hybrid does both, which is usually the right default when you are unsure.
Number of sources is a budget decision, not a quality dial. Every retrieved source competes for the same finite space as the prompt, skills, context and memory. Ten mediocre sources can crowd out the one good one.
What it can see
RAG searches as the person the run is acting for. It cannot surface a file or a record that person could not open themselves — your existing permissions apply unchanged. Retrieval does not become a way around access control.
Related
- Context — material handed over in advance.
- Memory — material the agent accumulated itself.
- Tools & MCP — the other way an agent reaches your data, by taking action rather than searching.