2 hours
a day
5 people, 4 questions a day each, 6 minutes to search and answer.
Technology: search with RAG
The answers are already in your documents. RAG finds them and quotes the source.
Procedures, manuals, product sheets, contracts: the answers exist, but only one person knows where, and every question goes through them. With RAG, whoever asks gets a written answer built from your documents, with the exact passage it comes from.
At a glance
- Used for
- Answering questions about procedures, manuals and internal documents
- Not for
- Processes with fixed rules, like checking an invoice
- Where it runs
- On your servers, with open models: documents do not leave
- How to start
- With a test on a set of documents and real questions
How it works
From the question to the answer, with its source.
The problem
One person knows where things are. Everyone asks them.
How do we handle a return for a foreign customer? What is the procedure for a damaged parcel? What warranty does that product have? The answers are written down somewhere: in a PDF, in a manual, in an email from two years ago.
Usually only one person finds them. Every question interrupts them, and when they are on holiday the answers stop.
In detail
RAG in four steps.
RAG stands for retrieval-augmented generation: first the right passages are found, then the answer is written using only those.
- 01
Prepare the documents
Procedures, manuals and sheets are split into short passages and indexed, so they can be searched by meaning, not only by word.
- 02
Find the right passages
The question is compared with the index, and the system retrieves the few passages that really answer it.
- 03
Write the answer
A language model writes the answer using only those passages, and shows which document they come from.
- 04
Say when it does not know
If the answer is not in the documents, the system says so instead of making it up.
When RAG is not the answer
RAG is not for automating a process with fixed rules: comparing an invoice with a price list is done with rules, not with a model that writes text. And it does not replace your ERP: to know how much stock you have, you query the ERP, not a document.
When it pays off
It pays off when the same questions come back every week.
It makes sense if
- You have many internal documents: procedures, manuals, sheets, contracts
- The same questions come back every week
- The answers already exist, written down somewhere
- Today only one person knows where to find them
You do not need it if
- The process has fixed rules: that calls for rules
- The data lives in an ERP or a database
- The documents are outdated or contradict each other: tidy them first
An example with numbers
Twenty questions a day to the person who knows.
Reference figures, to redo with your own.
Over 40 hours
a month, from one person
Over 21 working days: the time of whoever knows where things are, taken from their own work.
With the source
every answer
Whoever asks also sees the passage, and checks it in one click.
How to start: a test on a set of documents and about twenty real questions, counting how many answers are right. Only then do you decide whether to extend it, and on which server.
The limits
What to know before starting.
It can be wrong
That is why every answer quotes the passages it comes from, and can be checked in one click.
It repeats the documents
If a procedure is outdated, the answer will be outdated. Sources have to be kept in order.
Who sees what
Someone who cannot read a document must not get answers built on it. Permissions follow the documents.
It needs a suitable server
Open models also run on premises, but they need memory and power in line with documents and users.
Frequently asked
Questions on this topic.
Do the documents end up on ChatGPT or other outside services?
Not necessarily. The system can run on your servers with open models, and documents stay in the company.
Can it make answers up?
The risk exists with any language model. That is why the system answers only from the retrieved passages, quotes the source and says when it cannot find the answer.
Which documents can be used?
PDFs, Word documents, manuals, product sheets, exported emails. Scanned documents go through OCR first.
What does it take to start?
A test on a small set of documents and real questions. You measure how many answers are right before extending it to everything.
Do we need to buy hardware?
It depends on how many documents and users. It is decided after the test, with real numbers, not before.
Start with the problem
Procedures and manuals nobody can find?
Tell us about the documents and the questions that come up most often. We will tell you whether RAG is the right route, or whether something simpler will do.
- The first check is free and commits you to nothing
- Every message gets read, there is no call centre
- Fixed price and payment by milestone
- Italian and English, VAT invoices
- No phone call unless you want one
- If nothing should be built, we will say so
Prefer to write directly? [email protected]
Message received.