01What it solves
- support teams answering the same questions that are already written in manuals nobody reads;
- new hires spending weeks learning procedures and technical documentation;
- sales and engineering searching clauses, prices or specs across hundreds of files;
- asking business data questions in plain language and getting numbers you can verify.
02How a solid RAG system works
- Collection and cleaning of documents, split into passages that make sense on their own.
- Indexing in a vector database, combined with keyword search: hybrid retrieval fails far less often.
- Retrieval and answer grounded only in the retrieved passages, with references.
- Permissions: each user only sees documents they can already access.
- Evaluation: a test set of questions with expected answers measures quality before every release.
03Where I have done it
I built a natural-language query layer over more than 2 million documents spread across dozens of indices, with a schema catalog of 2,000+ fields in a vector database. Full story: AI-augmented analytics at document scale.
04How the work runs
- Discovery. Before any code: interviews, a look at the data you already have, technical and business constraints on the table.
- Architecture. Data schema, component boundaries and failure points decided on paper, when changing them costs an hour and not a month.
- Build. Short iterations with working demos instead of one delivery at the end. AI speeds things up where it makes sense; it never replaces review and testing.
- Handoff & scaling. Documentation, runbooks and a system your team can run without depending on me. No lock-in.
05Pricing
Every project is priced after a scoping call: the same word, "chatbot" or "agent", can mean a two-week prototype or a system in production for thousands of users. What moves the price:
- volume and format of the documents (scanned PDFs, tables and images take more work);
- how many sources to connect and whether they update in real time;
- permission and logging requirements;
- where the model runs: API, EU cloud or your own servers;
- expected query volume, which sets the monthly running cost.
The first call is free and ends with a clear answer: whether the project makes sense, roughly how big it is, and what to do first.
FAQ
Are my documents used to train the model?
No. In RAG, documents are read at question time and never train a model. With business APIs from the major providers or with a private model, your data stays out of training.
Does it work in Italian?
Yes. Current models handle Italian well; quality depends mostly on how documents are prepared and split.