How does AI work with company data? Train it or search it?
“How do we use AI with our company data?” has two common answers: train the model on the company’s documents, or search the documents and have the model answer from what it finds. They can look like the same result, but they differ a great deal in where the knowledge lives, how updates are made, whether a source can be shown and who can see what. This guide explains how AI works with company data, the two routes and when each fits.
Why doesn’t the model know your documents?
A large language model is a model pre-trained on large text datasets. The guide of the Turkish Personal Data Protection Authority on generative AI (November 2025) explains that once training is complete the model produces output according to the patterns it learned. Your company’s contracts, regulations and reports were not part of that training; the model sees them only if you show them to it. There are two ways to show them.
Route 1: fine-tune the model with more training
Fine-tuning is the further training of a base model for a specific task (the definition in the Authority’s guide glossary). In practice, examples are prepared from company documents and the model is retrained on them; the model changes. Methods such as LoRA keep the pretrained weights fixed and add small trainable matrices to each layer, which greatly reduces the number of parameters to train; the result is still an adapted model.
The limits of this route show up when you work with documents:
- The knowledge sits in the model’s weights. When a document changes, a new version of the model needs retraining.
- It is hard to trace which document an answer came from. The authors of the 2020 RAG paper said that providing provenance for decisions and updating knowledge were still open research problems for models that carry knowledge only in their parameters.
- Permissions live at the document level; weights do not. Once a document’s content has been worked into a model, it is harder to tell whether it ends up in an answer given to a user who is not allowed to see that document.
Fine-tuning suits setting the format and style of a specific task: answers that always come in the same template, for example.
Route 2: search the documents and answer from them
In the RAG paper, retrieval-augmented generation (RAG) is described as connecting a pre-trained language model (parametric memory) to a document index (non-parametric memory). The flow goes like this:
- The documents are uploaded, split into passages and put into a searchable index.
- A user asks a question.
- The system finds the passages related to the question.
- The model writes the answer from those passages alone and can show which passages it relied on.
The model does not change; the knowledge stays in the documents and the index.
The two routes compared
| Fine-tuning | Searching documents | |
|---|---|---|
| Where the knowledge lives | In the model’s weights | In the documents and the index; the model does not change |
| When a document changes | Retraining is needed | Update the document and re-index it |
| Showing sources | The document behind an answer cannot be traced directly | The answer rests on retrieved passages, and a passage can be shown |
| Who sees which document | Hard to separate at the document level | Results can be filtered by the user’s permissions at search time |
| First effort | Preparing training data, training, evaluating | Uploading and indexing the documents |
| Keeping it current | Training a new version | Keeping the documents current |
Who can see which document? The permission question
Company documents are not written for everyone: only HR sees a pay table, only legal sees a contract. Searching documents is a way of building permissions into the search and the answer: if a passage the user is not allowed to see never enters the search, it cannot get into the answer either. Making that separation in a fine-tuned model is much harder. So when you pick a document-based assistant, the first question should be “does it check permissions at search time?”
What does showing a source solve, and what not?
Seeing the document and page behind an answer makes checking easier; it is not proof of correctness. NIST’s generative AI profile notes that models can also produce confabulated logic or citations that purport to justify an answer, which can push users into misplaced trust, and it recommends reviewing sources and citations in outputs. So open the source and read the document; for how, see the guide on checking an AI answer (in Turkish).
When to pick which
- Search the documents if they change often, answers need to show a source, and not everyone should see every document.
- Fine-tune if you want to set the format, style or terms of a specific task and you have ready examples for it.
- The two can be combined. NIST’s profile also treats fine-tuning and grounding with retrieval (RAG) as separate methods; for questions about documents, search is usually the simpler and more traceable place to start.
If the documents contain personal data
According to the Authority’s guide, personal data can be processed at several stages of a generative AI system’s life cycle: when training datasets are built and the model is trained, when further information is derived after the model is in use, and through the data entered while the system runs and the output it produces. The guide also says that Law No. 6698 applies, regardless of technology, wherever personal data is processed. So training and searching are both processing if the documents hold personal data; whichever you pick, the same questions come up (which data, which purpose, whose access, for how long). For details see the Turkish guide KVKK ve yapay zekâ. This article is not legal advice.
How AURA works
AURA indexes your documents for search only. AURA does not train or fine-tune any model on customer documents; the models are open-weight and are used as they are. Many document types, Word and Excel among them, can be loaded into the knowledge base. The knowledge base’s owner decides who sees it: they can open it to everyone in the organisation or grant read and add-and-delete-document rights person by person (per-person rights are granted through the API today); permissions are checked both in the application and in the database, and an administrator cannot open other people’s conversations or documents. For an answer drawn from the knowledge base, when the model writes a citation, the document and the place it relies on are shown; clicking the source opens the page (the quoted section for formats that have no pages). In the Company Server and This Computer setups the models run on the company’s own hardware. Web search starts off on Company Server and This Computer and works if your organisation turns it on; on Voiser Cloud it comes on, and your organisation can turn it off. When it is on, only the search query and the addresses of the pages opened leave. Voiser runs the servers of Voiser Cloud.
For more, see AI for company documents and the on-premise AI assistant, or request a demo.
Frequently asked questions
Does AI learn my company's documents on its own?
No. A model produces answers from the patterns it learned before its training ended; your company's documents reach it only if you give them to it or train it with them. Giving them to it is what searching your documents does.
What is RAG?
Retrieval-augmented generation: before answering, the language model finds relevant passages in your documents and answers from them. The model does not change; the knowledge stays in the documents and in an index.
Can fine-tuning and document search be used together?
Yes. Fine-tuning can set a model's style or the format of a task, and document search can supply current, sourced information; for questions about documents, search is usually the simpler place to start.
What should I watch for if my documents contain personal data?
Personal data can be processed when training data is built and the model is trained, when data is entered while the system runs, and in the output, and the law applies whatever the tool. For details see the Turkish guide on KVKK and AI and ask your legal counsel.
Sources
The outside facts in this guide rest on these sources. The date is the source’s publication date; for a live document with none, it is the date it was accessed.
- Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks (Lewis et al.)Publisher: arXiv · Date:
- LoRA: Low-Rank Adaptation of Large Language Models (Hu et al.)Publisher: arXiv · Date:
- Üretken Yapay Zekâ ve Kişisel Verilerin Korunması Rehberi (15 Soruda) [Generative AI and the Protection of Personal Data: a Guide in 15 Questions]Publisher: Kişisel Verileri Koruma Kurumu (Turkish Personal Data Protection Authority) · Date:
- Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile (NIST AI 600-1)Publisher: National Institute of Standards and Technology (NIST) · Date:
Related guides
On-premise vs cloud AI: what changes for your data
On-premise AI runs the models on your own hardware. Where your data is processed, the trade-offs, and the questions to ask before you choose.
Updated
See AURA with your own documents.
Request a demo