You ask a chatbot about your company's vacation policy. It answers instantly, with flawless confidence, three neatly written paragraphs. The catch: it made them up. It never read your internal handbook. It filled the gap with what a generic vacation policy "usually says", and delivered it with the same face of certainty it would use for something true. That is a hallucination, and understanding why it happens is the first step to understanding what is RAG and why a private vault changes everything.
01 · the gapWhy an AI alone makes things up
A language model does not store documents. It stores a gigantic average of all the text it read during training, compressed into billions of parameters. When you ask it something, it does not look up a file: it predicts, word by word, the most likely continuation. It is astonishingly good at that. But it has two hard limits. The first: it only knows what was in its training, which ended on a fixed date. The second, graver for a company: it never saw your documents. Your contract, your handbook, the minutes of last Tuesday's meeting. None of that exists in its memory.
And here is the uncomfortable detail. A model does not know what it does not know. When the fact is missing, it does not go quiet: it generates the most plausible continuation anyway, because its job is to continue text, not to verify. The fluency is identical for the true and the invented. That is why a hallucination sounds so convincing: it is built with the same machinery as a correct answer.
A model does not know what it does not know. When the fact is missing, it does not go quiet: it fills the gap with the same confidence.
02 · the vaultA store of your knowledge, yours
A knowledge vault is, at its core, a private repository where what your organization knows lives: documents, transcripts, emails, notes, policies, history. It is not a drawer of dead files. It is a store prepared so a machine can traverse it by meaning, not just by exact words. Each text is split into manageable fragments, and each fragment becomes an embedding, a list of numbers that captures what it is about. So two fragments that speak of the same thing in different words end up close together in that space, and can be found even when they share not a single syllable.
The word "private" is not decorative. The vault is yours: it lives where you decide, readable by whom you control, and it does not dissolve inside a public model. It is the difference between lending your documents to an AI for a moment so it can answer, and handing them over forever to a training run you no longer control.
| Train the model | Retrieve from a vault (RAG) | |
|---|---|---|
| Where the fact lives | Diluted in the parameters | Intact in your documents |
| Updating | Retrain, costly and slow | Add a file to the vault |
| Where the answer comes from | Blurry memory, no source | Citable, traceable fragments |
| Risk of making things up | High when the fact is missing | Low: answers over what it retrieved |
| Your private data | Absorbed into the model | Stays in your vault |
03 · what is RAGRetrieve before answering
RAG means Retrieval Augmented Generation. The name is technical, the idea is common sense. Instead of asking the model to answer from memory, you first go to the vault, find the fragments that truly address the question, and place them in front of the model. Only then do you ask it to write the answer, with those fragments as raw material. The technique was formalized by a research team in 2020, in a paper that named it and showed that anchoring a generator to retrieved documents produced more accurate, verifiable answers [1].
Think of it as the difference between a closed-book exam and an open-book one. Closed-book, the model recites what it remembers and, when it does not remember, improvises. Open-book, it first finds the right page and then answers while looking at it. The same intelligence, very different results. And with an advantage a company cares about: the answer can cite where it came from. If it says your vacation policy gives twenty days, it can point to the exact fragment of the handbook that says so. The answer stops being an oracle and becomes a lookup with a receipt.
- Question. You write in your own words: "how many vacation days do I have?".
- Retrieval. The question becomes a vector and travels to the vault to find the fragments closest by meaning.
- Context. The fragments found, your real handbook, are placed in front of the model.
- Generation. The model writes the answer using those fragments, not its memory.
- Citation. The answer points to the source, so you can verify it in one click.
04 · train vs retrievePut it inside or leave it outside
Here it helps to undo a common confusion. Many people believe that for an AI to "know" something about your company you must retrain it on your data. Sometimes that is done, but it is rarely what you want. Training is slow, costly, and dilutes your documents into a mass of parameters from which they no longer come out intact: the model learns an average of your knowledge, not your knowledge. And when the handbook changes on Monday, the trained model keeps living in the past until you retrain it.
Retrieval is the opposite. Your documents stay outside the model, in the vault, whole and up to date. When a policy changes, you edit a file and that is it: the next answer already reflects it, retraining nothing. The knowledge lives somewhere you control, and the model only borrows it for the instant a question lasts. For most enterprise cases, retrieval wins by a landslide: cheaper, fresher, more auditable, and more respectful of your data.
This is where Qirava parts ways with the idea of "an app with a search box". A vault that answers well is not a single piece of software: it is the stitching of several layers. The capture that turns a spoken meeting into stored text. The embeddings that order that text by meaning. The retrieval that finds what is relevant. The model that writes. And, when the decision matters, a human layer that reviews before the answer goes out. Qirava does not sell a loose feature: it integrates agentic and human layers to deliver a service that gives real value, and sometimes that service does not even look like software. The value is not in a button, it is in all those things happening together and well.
05 · inside QiravaWhere what you captured lives
In Qirava the vault is the heart where the knowledge that comes in through Transcribe is deposited. Someone records a meeting, an interview, a voice note; Transcribe turns it into faithful text; and that text is not lost in a folder: it lands in the vault, gets split, becomes embeddings and stands ready to be retrieved. From then on, asking Qirava about what was said in that Tuesday meeting stops being an exercise of human memory and becomes a query to the vault, with the answer anchored to what was actually said, not to what someone thinks they recall.
That is the whole point. An AI alone is brilliant and forgetful, and fills its gaps with elegant invention. A private vault plus retrieval ties it to the ground: to your documents, your meetings, your history. It leaves it, at a stroke, less of a poet and far more reliable. And that reliability, woven with capture, retrieval and the human judgment that reviews it, is what turns a chatbot that dazzles in a demo into a service a company can trust Monday to Monday.
An AI alone is brilliant and forgetful. A vault ties it to the ground: to your documents, your meetings, your history.
Sources
- Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V., Goyal, N., Küttler, H., Lewis, M., Yih, W., Rocktäschel, T., Riedel, S., & Kiela, D. (2020). Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. Advances in Neural Information Processing Systems (NeurIPS) 33. arXiv:2005.11401. (The paper that introduces and names the RAG approach.)
- Karpukhin, V., Oğuz, B., Min, S., Lewis, P., Wu, L., Edunov, S., Chen, D., & Yih, W. (2020). Dense Passage Retrieval for Open-Domain Question Answering. Proceedings of EMNLP 2020, pp. 6769-6781. arXiv:2004.04906. (Dense retrieval by embeddings, the basis of how a vault finds what is relevant.)
- Ji, Z., Lee, N., Frieske, R., Yu, T., Su, D., Xu, Y., Ishii, E., Bang, Y., Madotto, A., & Fung, P. (2023). Survey of Hallucination in Natural Language Generation. ACM Computing Surveys, 55(12), art. 248. (A survey of why models hallucinate and how grounding to sources mitigates it.)