Services

Retrieval-augmented generation

RAG development: assistants that answer from your documents and cite their sources.

Gorilli builds RAG (retrieval-augmented generation) systems: assistants and search that answer from your documents, manuals and knowledge bases, with the source cited in every answer. We are an AI product studio in Milan and run the same approach in production in our own product, Kommander. We handle the whole chain, from document ingestion to evaluating answers in production, and we respect permissions so each person only sees the documents they can access.

What you get

RAG development on solid technical foundations.

  • Assistants for customers or employees that answer from manuals, procedures, contracts and FAQs, with the source of every answer
  • Search across internal documents that understands plain-language questions, across PDFs, Word files, web pages and tickets
  • Per-document permissions: answers only use content the user can already read
  • An evaluation set of real questions to measure accuracy before and after every change

How we work

  1. 01

    Source inventory: formats, update frequency, duplicates and conflicting versions

  2. 02

    Ingestion and indexing with hybrid semantic and keyword search, and chunking suited to each document type

  3. 03

    An evaluation set built from the questions people really ask, to measure answers and sources

  4. 04

    Release with citations, a confidence threshold, handoff to a person and monitoring of answers without a source

Timeline

  1. Source inventory and preparation

    1-2 weeks

  2. First assistant on part of the sources

    2-4 weeks

  3. Evaluation, permissions and channels

    2-6 weeks

  4. New sources and automatic updates

    Ongoing

Typical budget

  • RAG assistant on a defined document set, one channel

    Up to about €20,000

    Like a simple agent. Indicative market range.

  • Many sources, per-document permissions, connectors and continuous evaluation

    €20,000-80,000

    Indicative market range: the project price is set after discovery.

Why Gorilli

  • Kommander, our customer support product, answers from company documents with the source of every answer in 15+ languages
  • Evaluation is part of the code: every change to prompts, models or indexing is measured
  • We integrate RAG into existing systems through APIs or MCP, not only as a separate chatbot
  • Data can stay in Europe, with providers that do not train models on your documents

Guide: Build an AI product: timeline, costs and FAQ

Comparison

OptionWorks whenMain risk
Gorilli (custom RAG)Many mixed sources, per-document permissions, integration with existing systems and channelsNeeds access to the sources and time to build an evaluation set of real questions
Chatbot with file uploadA few stable documents and one person or a small teamNo control over permissions, updates or answer quality
AI features of your suite (Microsoft 365, Google Workspace)The documents already live there and standard features are enoughHard to include sources outside the suite or to customise answers and channels
Fine-tuning a modelYou need to teach a style or format, not knowledge that changesExpensive to update and cannot cite sources: a poor fit for documents that change often

Frequently asked questions

How much does RAG development cost?
As an indicative market range, a RAG assistant on a defined document set and one channel costs up to about €20,000, like a simple agent. With many sources, per-document permissions, connectors to company systems and continuous evaluation it rises to €20,000-80,000. The first call with Gorilli is free.
How long does it take to put a RAG system in production?
A first assistant on part of the sources usually takes 3-6 weeks, inventory included. Permissions, new channels and full evaluation take another 2-6 weeks, depending on how many sources and systems need connecting.
What is the difference between RAG and fine-tuning?
RAG retrieves information from documents at question time, so it stays up to date and can cite the source. Fine-tuning changes the model: useful for style and format, a poor fit for knowledge that changes often.
How do you stop the system from making answers up?
The model answers only from the retrieved passages and shows the source. When confidence is low it says so or hands over to a person. An evaluation set of real questions measures how many answers are correct and supported by sources, before and after every change.
Which documents can be used?
PDFs, Word files, web pages, internal wikis, tickets and FAQs, plus data from a CRM or ERP through APIs. Quality matters more than quantity: duplicate or conflicting documents should be fixed before indexing.

Next step

Tell us what you want to build.

We help you understand scope, technical risks and the fastest path to a product people can use. The first call is free.

RAG Development Company | Gorilli