RAG systems: answers from your company documents you can trust

We build systems that find what you need in your documents, answer with a link to the source, respect access rights, and deliver measurable answer quality.

In short

RAG (retrieval-augmented generation) connects search over your documents with a language model: the model answers not from memory but from the passages it found. A demo is quick to build. The difficulties start on a real corpus: scans, tables, outdated versions, documents not everyone may see. RAG quality depends on search and corpus preparation more than on the choice of model, and that is where we focus.

Where RAG helps

  • An assistant for employees on procedures, instructions and the internal knowledge base.
  • Customer support: answers from product documentation and terms of service.
  • Work with contracts, regulations, and technical documentation, where a reference to the exact clause matters.
  • Search across accumulated correspondence, reports, and minutes that nobody has time to read.

If the answer is not in the documents, RAG will not create it. Tasks that require calculations or actions in systems need a different architecture, and we will say so in the first conversation.

What the system is made of

The corpus and its life cycle

Which sources go into the corpus, who owns them, how versions are updated and how withdrawn documents are removed. Without this, the system answers confidently from an outdated procedure.

Document processing

Extracting text from PDFs, scans, tables, and presentations, and splitting it into passages that follow the structure of the document: sections, clauses, table headers. This stage most often determines the final quality.

Search

Hybrid search: vector search by meaning plus full-text search by words, numbers, and names. Hebrew has prefixes and attached prepositions, Russian has cases and endings, so the full-text component must account for the language's morphology. Re-ranking the retrieved passages with a separate model noticeably improves accuracy.

Models

For vector search, multilingual embedding models in the cloud (OpenAI, Google, Cohere) or open models deployed in your environment. For answer generation, models from OpenAI, Anthropic, Google, or an open model on your own servers. The choice depends on the language of the documents, data requirements, and the cost per request.

Storage

PostgreSQL with pgvector, Qdrant, Milvus, or OpenSearch, in your environment or in the cloud. The choice depends on the size of the corpus and what your infrastructure already includes.

Access rights

An employee should receive answers only from documents they are allowed to see. The permission filter is applied during search, not after the answer is generated.

Quality evaluation

A set of real questions with verified answers, used to measure search and answer accuracy. Every change to the corpus, the splitting, or the model is tested against it before going live.

How the work goes

  1. Understanding the task and the corpus. Who will ask, about what, from which documents. We collect a set of real questions with your specialists.
  2. A pilot on part of the corpus. Document processing, search, and an initial quality evaluation. Even at this stage, the main difficulties become clear.
  3. Refining search and corpus preparation to reach the agreed quality level on the control questions.
  4. Integration and launch. Integration with your systems, access rights, monitoring, user feedback.
  5. Ongoing support. Corpus updates, quality control, SLA support, or handover to your team.

If RAG already runs but answers poorly, we start by analysing quality on your questions.

Personal data: Amendment 13 and GDPR

If the documents contain personal information, the requirements for protecting it apply not only to the source files but also to the passages in the vector database, the query logs and the model's answers. In Israel these are the requirements of the Privacy Protection Law with Amendment 13; for data of EU residents, the requirements of GDPR. Sending passages to an external model API means passing data to the provider, which needs a contract with the provider and clarity on where the data is stored and for how long.

We design the system so these requirements can be met: storage and models where your rules require them, masking personal data the answer does not need, and control over logs. The legal assessment and the decisions are for your company's lawyers to make. What Amendment 13 changes in software systems is covered on a separate page.

How we work

  • Remotely. Meetings with your team and subject-matter experts, work in your repository and your environments.
  • The documents stay with you. Documents without personal data, or anonymised documents, are suitable for the pilot. Access to the rest is decided by your security team.
  • In-person meetings. Possible by arrangement, if the task requires it.
  • The code and infrastructure are yours. The repository, servers, and API keys are in your name from day one.

Why us

We have been building and supporting software systems since 2004: 150+ projects for clients in 14 countries. Our main system has been running in production without interruption since 2008. We build RAG systems ourselves and know where they break: on documents, access rights and the lack of measurement.

Frequently asked questions

Why RAG rather than fine-tuning a model on our documents?

For answers from documents, RAG is usually simpler and more reliable: documents can be updated without retraining, the answer refers to its source, and access rights are respected during search. Fine-tuning makes sense for style, format or narrow terminology, and is often combined with RAG.

Can the system make up answers?

It cannot be ruled out completely. The risk is reduced by accurate search, a requirement to answer only from the passages found with a reference, declining to answer when data is insufficient, and regular checks on control questions.

Can we avoid sending documents to an external service?

Yes. The storage, the embedding models, and the generative model can be deployed inside your environment. This requires computing resources, and we estimate in advance what answer quality the models that fit those resources will deliver.

How soon will we know whether it works on our documents?

After a pilot on part of the corpus. It shows the real accuracy on your questions, not on demo questions, and the main difficulties with the documents.

Which document formats work?

PDF, Word, Excel, presentations, and scans, as well as exports from internal databases and document management systems. New versions of documents enter the index on a schedule or based on an event, and withdrawn documents are removed from it.

How fast and accurate are the answers?

Speed depends on the model, the amount of passages found and where the model runs. It is measured in the pilot together with accuracy, on your questions. Every answer comes with a link to the document and the place in it, so it can be checked.

Where to start

Tell us which documents the answers should come from, who will be asking the questions, and whether the documents contain personal data. Based on that, we will propose a pilot.

Talk to us

Related: AI solutions, Amendment 13 for systems

SLAtech LTD