The first route is the easiest start. With Hermes, you often connect a source through MCP, an open standard for attaching tools and data sources to an agent. The Hermes MCP guide recommends starting with one server, such as access to a single project folder, and allowing only the tools you need right away. Read which connections make sense in connecting AI to your CRM, CMS, and project tools.
RAG stands for retrieval-augmented generation: retrieve first, then answer. It is the common solution for large archives, but the quality depends on how you split the documents into chunks. A clause that refers to an appendix can easily lose its meaning that way.
The third route is newer. In April 2026, Andrej Karpathy described the LLM Wiki pattern. Instead of searching for fragments again with every question, the agent builds a set of linked markdown notes and keeps them up to date. Karpathy sums up the difference with RAG like this: "The knowledge is compiled once and then kept current, not re-derived on every query." In practice, the agent processes a new source once, flags contradictions, and updates the summary.
Hermes ships that pattern by default as the llm-wiki skill. The division of labor is clear: people choose the sources, and the agent summarizes, connects, and keeps everything consistent. You can also combine the routes. An agent can start in the knowledge base and still open the original document for the details.
What does such a knowledge base look like at Voltti?
At Voltti, our own Hermes installation keeps a knowledge base in plain markdown files, following that LLM Wiki pattern: one note per topic, with its sources.
There is also an index with one line per note and a log of what changed. For example, one note holds facts from the Hermes documentation, including the date they were read, and another holds decisions that have already been made.
Before an audit or an article, the agent first reads the matching note. Decisions we confirm go into that note, so the agent does not ask about them again next time. Every change goes into the version history separately, so you can look back at when something changed. A note does not replace the source: anyone who uses a fact in a text checks it in the original source.
How do you keep an AI knowledge base safe?
Start with access. Give the agent read access to the sources it needs and nothing more. Write access belongs only where the agent saves its results, such as the project documentation.
Also think about who may see what. The agent reads with the access you give it, not with the permissions of the person asking. An agent that searches the whole company can therefore answer from a document the person asking was not allowed to open. So work with a separate environment per team or department. In Hermes, each profile has its own sessions, memory, and credentials, and profiles do not share them.
Documents can also contain instructions that were not meant for the agent. OWASP calls this indirect prompt injection: text in a file or on a web page that changes how the model behaves. So do not let the agent act on what it reads in a document without a person checking.
And where does your data go? With a cloud model, the text needed for a question goes to the model provider. With a local model, that stays inside your network too. Read what stays local in a hybrid AI setup and how we handle security and privacy. If you want to keep everything in-house, look at local AI for businesses.
If those sources contain personal data, such as HR files or customer emails, the GDPR applies as usual. It asks you not to use more personal data than you need, so give the agent only the sources a question needs. If a cloud provider processes that text for you, put that in a data processing agreement, or let a local model do the work. Our AI safety checklist for business data covers the other arrangements.
Why should every answer point to its source?
Because internal knowledge is often contradictory or out of date. With the location and the date, you can judge which answer is right.
There may be three versions of the same procedure, or a pricing agreement sits in an email that was changed later. An agent that only gives an answer hides that problem. So always ask for the location, and have the agent report when two sources contradict each other.
For example, you ask what delivery time you agreed on with a customer. The agent gives the delivery time, the contract it comes from, and the date. It also notes that a later email mentions a different delivery time. You decide which one applies.
Checking stays human work. When the agent summarizes a contract, someone reads the flagged passages before anything is decided. You do not read everything again, just the passages that matter.
How do you start searching internal knowledge?
Start small, with one area of knowledge your team searches often today.
Pick one area, such as quotes, manuals, or customer agreements.
Put the sources together and give the agent read access to that place only.
Write down twenty real questions your colleagues ask, with the right answer next to each one.
Have the agent answer those questions, with the location for every answer.
Check the answers and note where the agent gets it wrong.
Only expand once that first set works reliably.
A test like this fits well in a pilot. Read how to prepare one in prepare your AI agent pilot. You will find other tasks that keep coming back in automating recurring tasks, and what an agent can do more broadly in Hermes Agent explained.
Voltti installs Hermes for businesses in Belgium and the Netherlands and helps set up the connections to your documents. You receive a custom quote, because sources and permissions differ for every business. Our page on AI automation explains how we work.
