Answers from your own knowledge, with the passage behind each one
Your procedures, your tickets, your contracts, your wiki. A model that reads them answers in your own words and cites the page it read, so the person asking can check it.
Munito AI serves the reading and the writing, on European infrastructure, with zero data retention by default. Your content stays yours, and it stays where you put it.
Be clear about the split
We serve no embeddings endpoint. That is a real limitation and it changes how you build. You own the index; we answer from what it returns.
| Step | Who does it |
|---|---|
| Store and chunk your content | You, in your own system |
| Build the index | You, with your own embeddings or a keyword index |
| Retrieve the passages | You |
| Read them and answer | Munito AI |
| Return the source | Your application, from the passages it sent |
If you want one vendor for every step, say so early: Scaleway, OVHcloud and Blade all serve embeddings and we do not. The retrieval recipe builds the whole loop with a keyword index that runs on your machine, so you can measure the answer quality before you choose a vector store.
A long context changes the design
DeepSeek-V4-Flash, DeepSeek-V4.1-Flash-1017 and GLM-5.3-Flash each read a million tokens. That is enough to stop cutting documents into small pieces and hoping the right one comes back.
Send fewer, larger passages. Fewer chunks mean fewer retrieval misses, and a model that sees a whole section writes a better answer than one that sees three fragments of it.
A repeated prefix costs almost nothing: cached input on DeepSeek-V4.1-Flash-1017 is $0.0060 per million tokens, against $0.30 uncached. A fixed instruction block in front of every question is close to free.
Make the answer checkable
Use structured outputs and the model returns the answer and the identifiers of the passages it used, in a shape your application can render as links. An answer no one can check is a rumour with good grammar.
The semantic search recipe covers ranking, and the retrieval recipe covers the generation step.
Internal knowledge is why the region matters
A support wiki holds customer names. A procedure holds how a control works. These are the documents a security team asks about first.
Send to https://api.eu.munito.ai and the request runs in the EU; the response names the region in the munito-region header. Munito AI keeps no prompts and no outputs by default, and trains on nothing you send. The DPA applies with no signature.
For content that may not leave your perimeter at all, run the platform inside your own cloud account or on-premises. See Enterprise.
Prove it on questions you already know the answer to
Write fifty questions your team answers today, with the correct source for each. Measure how often the system finds the right passage, how often the answer is right when the passage is, and how often it says it does not know.
The second number is the model's job. The first is your index's, and it is usually the one that is wrong.
Start
- Build the retrieval loop
- Rank the passages
- Fix the response shape
- Talk to an engineer about your content and your accuracy bar