Technology
Below is a working model of the system we build for clients. Every point is a fragment of this company's documentation, placed by meaning. Ask a question and watch the search: the question becomes a vector, the closest fragments are found and ranked, and a language model running on our own servers writes the answer from them.
/01 RAG
Documents are split into short fragments: contracts, manuals, policies, product sheets, emails. Text is extracted from PDFs, Word files and scans, with OCR when needed.
An embedding model turns each fragment into a vector. Fragments with similar meaning end up close together, even when they are written in different languages.
The question becomes a vector and is compared with every fragment. The closest ones are selected, and filters can limit them by department, client or permissions.
The language model receives only those fragments and writes the answer, citing where each part comes from. If the answer is not in the documents, it says so.
/02 STACK
We choose the components for each project according to the client's data, budget and privacy needs.
Open-weight models such as Llama, Qwen, Mistral or gpt-oss run on the client's own servers or in a private cloud. Data stays inside the company's infrastructure and there is no cost per question. The assistant on this page runs this way.
When a task needs the most capable models, we integrate OpenAI, Anthropic Claude or Google Gemini through their APIs, with spending limits, logging and a second provider as fallback.
Retrieval-augmented generation connects a language model to a company's documentation. The model answers from the fragments it retrieves and cites them, instead of answering from general knowledge.
We extract text and tables from PDF, Word, Excel and scanned documents, classify them and keep them in a searchable knowledge base that updates when the documents change.
Agents and workflows that read and write in the tools a company already uses (CRM, ERP, accounting, email, WhatsApp, spreadsheets, databases), with a person approving the steps that need it.
Containers, a reverse proxy with TLS, backups, monitoring and access control. People sign in with their company accounts, questions can be logged for auditing, and the system grows from one server to several as usage grows.
/03 ENTERPRISE
The same technology works inside a company. Employees ask in plain language and get answers taken from the organization's manuals, contracts, procedures and files, with a link to the source. It is similar to NotebookLM, but installed for your company, connected to your systems and limited to what each person is allowed to see.
How do I process a refund? What is on the onboarding checklist? Answers from the current version of each procedure.
Which contracts renew this quarter? What does the supplier agreement say about penalties? Clauses found and quoted in seconds.
Support and engineering teams search manuals, tickets and specifications by meaning instead of exact words.
Vacation, benefits and expense rules answered the same way for everyone, at any hour.
Past proposals, prices and case studies found in seconds to answer a client faster.
Questions about invoices, statements and reports, with the figures taken from the documents themselves.
We map the document sources, the users, the permissions and the questions that matter most.
We connect the sources (shared drives, Google Drive, SharePoint, email, databases) and build the knowledge base.
A working assistant for one team within a few weeks, measured with that team's real questions.
More teams and sources, permissions per group, usage reports and monthly improvements.
In a free 30-minute call we review which documents and questions matter most to your team and propose a pilot.