Skip to content

RAG systems for your own documents

Connect your PDFs, contracts, manuals, shared drives and wiki to an AI assistant that answers in plain language and shows the document and page it used. It runs on your own server, offline if needed, or in a German cloud, so your data stays under your control.

Why ChatGPT cannot answer from your files

How a RAG system works

RRetrieval: find
The system finds the passages in your sources that fit the question best.
AAugmented: hand over
Those passages are added to the question, only from documents the person asking may see.
GGeneration: answer
A language model writes the answer from them and names the document and page. If nothing fits, it says so instead of guessing.

Inside: from document to vector database

  • Extract and splitText is pulled from every file, scans through OCR, then split into passages that keep title and page.
  • EmbedAn embedding model turns each passage into a vector: a list of numbers that stands for its meaning.
A book page turns in the wind by the harbour while a cargo ship passes behind
  • Store in a vector databaseVectors, text, page and access rights are stored together, indexed to find the nearest vectors in milliseconds.

Start with one use case

Pick one question your team asks every week. Once those answers are right, add the next.

  • Support answersFrom manuals, product sheets and past tickets, for your team or your customers.
  • HR and policy questionsHolidays, travel costs and onboarding, answered from your own handbook.
  • Contract and clause searchNotice periods, liability and prices across all your contracts.
  • Offers and product knowledgeSales finds past offers, specs and product details without asking around.

RAG data sources

  • Files and PDFsPDF (also scanned), Word, Excel, PowerPoint and CSV, on a file server or NAS.
  • DrivesSharePoint, OneDrive, Google Drive, Nextcloud and Dropbox.
  • Wikis and intranetConfluence, Notion, your intranet or your own website.
  • Mail, tickets and chatOutlook or Gmail mailboxes, Jira tickets, Teams and Slack channels.
  • DatabasesPostgreSQL, MySQL or SQL Server: your question becomes a database query, and the answer shows it.

Hosting and GDPR

GDPR, point by point

A RAG system reads your documents, and they often hold personal data. These four points cover what the GDPR asks of the setup.

Technical groundwork, not legal advice. Your data protection officer has the final word.

Art. 28Agreements
A data processing agreement with every provider that handles personal data.
Art. 17Deletion
A document deleted in your system leaves the index at the next sync.
Art. 5Your logs
Questions and answers are logged only if you want, on your own system, for as long as you decide.
Art. 30Documentation
The data flows written down for your records of processing, and a DPIA where one is needed.

RAG pilot in 4 to 6 weeks

  1. Free first callWhich questions, from which sources?
  2. Use case and planOne use case, its sources, and what counts as a good answer.
  3. PilotYour sources on your setup, tested with your real questions.
  4. Measured resultHow often the right source was found, on your own test questions.
  5. Rollout and careMore sources and users, team training, new documents indexed automatically.

Languages and tools

The basics, chosen per project. myRAG below is one combination.

Languages
  • Python
  • TypeScript
Frameworks
  • LangChain
  • LlamaIndex
Vector databases
  • PostgreSQL with pgvector
  • Qdrant
Embedding models
  • bge-m3
  • OpenAI embeddings
Language models
  • Ollama (local)
  • OpenAI
  • Claude
myRAG with five indexed PDF reports, a question about the Hudson River ditching, the answer with a source marker and the cited report pages

Demo: myRAG

What it is
Belambo’s own open-source RAG assistant for PDF documents. A demo, not client work.
Measured
The right source in the top five for 85 % of 46 test questions; 98 % of answers true to the source.
Runs on
One machine: Python and FastAPI, PostgreSQL with pgvector, bge-m3 embeddings, Qwen3 through Ollama, reranked by the language model.

myRAG on GitHub

RAG questions

How is this different from ChatGPT with file uploads?

Uploads cover a few files at a time and leave your company. A RAG system searches all your sources at once, follows your access rights and can run on your own server.

What hardware do I need on-premise?

For a team, usually one workstation or small server with a modern graphics card. You get an exact recommendation in the offer.

Does my data go to ChatGPT?

Only if you choose that. EU-hosted or self-hosted models are the default options, decided per project.

Does Belambo keep any of our data?

No. Belambo stores no user data from your RAG system. Documents, the search index, questions and any logs stay on your server or in your own cloud, for your admins.

Abdul Hannan sitting in an old armchair in front of a mural of a sparrow

Contact

A short message is enough.

Replies within one working day