A RAG chatbot for your company documents

Can an AI assistant answer questions from our own documents?

Yes. That is what retrieval-augmented generation (RAG) does: before the model writes an answer, the assistant searches your own documents and hands the model only the passages that answer the question. Every answer links back to the document it came from, so your team can check it in one click.

It works across drives, wikis, ticket history and product data, in plain language, for support teams answering customers and for employees looking for an internal procedure.

How does it avoid making things up?

We match the retrieval to the questions you get, rather than using one mode for everything:

• Direct search for a clean, well-kept knowledge base with single-fact questions, where speed and a low cost per answer matter most.
• A corrective pass for knowledge bases that are incomplete or go out of date: each retrieved passage is checked for relevance first, and when nothing fits, the assistant rewrites the query from the conversation or says it doesn’t know. This is our default for customer-support assistants.
• A knowledge graph for questions that span many documents, such as “what changed over the last two weeks” or “how is company X connected to project Y”, where the answer depends on how things relate.
• A second model that checks the first before the user sees the answer, for the few answers where every word matters.

Will it find exact codes, names and article numbers?

Yes. Meaning-based search alone tends to blur exact strings, so keyword search (BM25) runs alongside it. Part numbers, clause references and product codes are matched exactly, while the rest of the question is matched by meaning.

For scale and metadata filtering we use Qdrant; when the answer depends on relationships between documents, Neo4j.

Does it respect who is allowed to see what?

Yes. Permissions follow the user: the assistant only searches the documents the person asking is allowed to open. A question from sales never surfaces a document that only HR can read.

Which model does it run on?

The one that fits your budget and your data rules: OpenAI, Anthropic, Google, Qwen, DeepSeek, or an open-weight model on your own servers. All sit behind the same interface, so switching to a better or cheaper model later is a configuration change. When a cloud model is used, a small local model masks personal data before the request leaves and restores it in the answer.

When don’t you need one?

When the questions are few and the answers never change, a well-organised FAQ page is cheaper. When people mostly need one exact figure from one system, a report or a search box on that system may be enough. We’ll say so on the first call.