Artificial intelligence
AI chatbots that answer with your data
An AI chatbot is an assistant that holds a conversation in natural language using a language model (LLM). When it is built with RAG — Retrieval-Augmented Generation — it searches the company's documents before answering and writes from them, citing where each fact came from. That difference is what separates a chatbot that informs from one that makes things up.
Almost every chatbot that fails does not fail because of the model: it fails because nobody designed where the answer comes from.
Why a chatbot without your data invents
A language model does not look things up: it predicts. It learned patterns from billions of texts and, when you ask it something, it generates the sequence of words that statistically fits. If the right answer was in its training, it gets it right. If it was not, it invents one with exactly the same confidence.
The problem is not that it is wrong. It is that it has no way of knowing it is wrong, so it does not warn you — and the person reading cannot tell either.
What RAG changes, concretely
RAG changes the question. Instead of «what is most likely?», the system asks «what do these documents say?». Before answering it searches your sources — manuals, PDFs, your knowledge base, your catalogue — for the fragments that apply, and hands them to the model together with the question.
The model stops writing from its memory and starts writing from what you put in front of it. Every answer can cite its source so anyone can open it and check. And when the fact is not there, a well-built RAG says so instead of filling the gap with something plausible.
Where the chatbot lives
On the channel your people already use, not a new one you have to teach them. WhatsApp is the most common case in Colombia because that is where the conversation with the customer already happens; Telegram, Slack or the website work too.
The channel changes the shape of the answer, not the engineering: the same retrieval system answers on all of them, which is why the channel should be decided by where your people are and not by where a demo is easiest to set up.
What decides whether it works is the boring part
RAG is not a button, it is a pipeline, and every stretch of it can break. Cut the documents into pieces of the wrong size and the right fragment never arrives. Have the search retrieve ten similar paragraphs but none with the fact, and the model goes back to inventing.
And if nobody measures how often it answers well, you do not have a system: you have a demo that got lucky. That is why every project carries continuous evaluation from the start — a model without measurement is a demo, however well it speaks.
Frontier models or open models
It can be built on frontier models (Anthropic, OpenAI) or on open models fine-tuned and running on your own infrastructure. The decision is not a matter of fashion: it depends on how much each query can cost and how sensitive the data passing through it is.
If your documents cannot leave your network, that constraint decides the whole architecture — and it is worth knowing before building, not after.
Questions
What people usually ask us
- What is RAG and how is it different from a normal chatbot?
- RAG (Retrieval-Augmented Generation) is a technique where the system searches your documents before answering, and the model writes from those fragments. A normal chatbot answers from what the model memorised during training; one with RAG answers from your sources and can cite them.
- Can an AI chatbot answer using my company's internal documents?
- Yes, and that is exactly what RAG is for. The chatbot is not trained on your documents: it consults them at the moment of answering. That means updating the information is a matter of changing the document, with nothing to retrain.
- Can it be connected to WhatsApp?
- Yes. WhatsApp is the most requested channel in Colombia because that is where the conversation with the customer already happens. It also works on Telegram, Slack or embedded in the website.
- Does my data leave my company?
- It depends on the architecture chosen, and that is a decision taken before building. With frontier models, queries travel to the provider. With open models running on your own infrastructure, they never leave your network. If your data is sensitive, that constraint defines the entire design.
- How do I know the chatbot is answering well?
- With continuous evaluation. You build a set of questions with their correct answers and measure against it systematically, so you have a number rather than an impression. Without that measurement there is no way to know whether it works, or to notice when it gets worse.
Not sure which one you need?
If you're not sure, that's what the assessment is for: we look at it with you and tell you what's needed — even if it's less than you thought.