LLM integration & RAG GPT, Claude and Gemini answering from your documents

RAG combines search with a language model: before the model answers, the system finds the relevant passages in your documents and hands them over together with the question. Ordinary ChatGPT does not know your procedures, price lists or contracts, so it will guess or answer in generalities. With RAG the answer is grounded in a source, you can see where it came from, and knowledge is updated by replacing a document rather than retraining a model. It is the most common way to connect a language model to organisational information. The first build is relatively quick; the steady improvement afterwards is what separates a demo from a working tool.

Yes, but Hebrew has to be tested rather than assumed. The large language models understand Hebrew well, yet the search step that finds the right passage for them is sensitive to spelling, acronyms, poor scans and tables. So we choose the search method and the model only after running real questions against your documents, including questions phrased in different ways. If some files are scanned images we add a text recognition step, and we flag in advance any documents that cannot be read at acceptable quality. The feasibility check shows you how the system handles your most awkward files before you commit to anything larger, so surprises come early and cheaply.

Zero errors cannot be promised, but they can be reduced a great deal and made visible. The assistant is instructed to answer only from the passages it retrieved, shows the source with each answer, and says it found nothing when the information is missing. Before launch we run a fixed set of questions and review the results by hand, including questions whose answer is not in the documents at all. After launch, failed questions are saved and used to improve the system. For sensitive topics a human can approve the answer before it reaches a customer. The aim is an assistant you can rely on, one that is honest about the limits of what it knows and hands over to a person when it should.

That is decided in the specification. One option is a business-grade API from a major provider, where the terms of use state whether data is used for training, and where only the passages needed for a question are sent rather than the whole archive. When data cannot leave the organisation, an open model runs on a server you control. Each option is chosen by how sensitive the data is, and for personal data the provider's terms must be checked. We prepare a written description of the data flow that you can pass to your privacy adviser. That lets you decide from a clear document rather than from a vendor's general promises, and your adviser can review the same description.

Amendment 13 to Israel's Privacy Protection Law broadened the obligations of bodies that hold personal data, so privacy enters the specification from day one. In practice we work on data minimisation: we do not load documents into the assistant that it does not need, we limit access by role, we log usage, and a document can be removed from the index. We do not give legal advice. If your database is sensitive, we describe exactly what is collected, where it is stored and who can reach it, so your adviser can decide whether more steps are needed. The topic is worked through with you at the start of the project and is not left for the launch stage, when changing course is harder.

A prototype over one set of documents and real questions can usually be shown within a few weeks, depending on the quality of the documents and the number of systems it connects to. At the end you receive an assistant working in the channel your team already uses, the code for the connection and the index, a test set that shows the accuracy level, and a document explaining how to add and update documents. The code belongs to you. After launch you can leave monitoring and ongoing improvement with us, or take a handover session and run it yourselves. In the first conversation we will check together how likely it is that your documents suit this approach, and say plainly if they do not.