RAG Chatbot

Organizations collect hundreds or thousands of documents over the years: contracts, guidelines, product data sheets or internal files. Every search costs time, because it is often unclear where the relevant information is located. Knowledge is scattered across email attachments, Excel files and PDFs. This quickly becomes a problem, especially when onboarding new employees.

Chatbots are only partly suitable for this. With multiple documents, input limits are reached quickly, there is no permanent access to your content and there are no reliable source references. Sensitive data also does not belong on external servers.

A dedicated AI application indexes your documents once, answers questions with source references, can optionally run in your own infrastructure and grows with your knowledge base.

For these cases, Retrieval-Augmented Generation, or RAG, is a good fit. Instead of manually uploading documents or letting the model guess, a dedicated knowledge base is built that the AI can access in a targeted way.

In this demo, a web crawler downloaded all Pensplan information sheets. The documents were then indexed, split into text sections, so-called chunks, and transformed into embeddings. Embeddings are the numerical representation of text used to measure semantic similarity. The embeddings were then stored in a vector store on a server.

When a user asks a question, the system first searches for the relevant document passages and gives them to the AI as context. The answer is therefore based on the found sources instead of free model fantasy.