How it works

The moinAI RAG Agent follows a clear process: sources are retrieved, verified, and processed before an answer is created. This results in a controlled, traceable process.

1. Retrieve sources

In the retrieval phase, all relevant sources (docs, PDFs, URLs, etc.) that contain information for creating an answer are searched.

‍

2. Knowledge verification

After retrieving the relevant sources, an optional knowledge verification can be activated. This checks whether the content of the sources provides sufficient information to create a meaningful answer. This largely prevents hallucinations.

‍

3. Instructions

The moinAI RAG Agent also works with individual instructions. For example, certain answer formatting can be forced, such as the command "Add the suffix 'without guarantee' at the end of each answer".

‍

4. Guardrails

The moinAI RAG Agent automatically checks so-called guardrails, which could also be described as guidelines. These include restrictions regarding competitive information or blocking inadmissible instructions. Inadmissible instructions could be, for example, the request that the chatbot assures the user that all products in the shopping cart will be given to them as a gift, or similar. This protects against fraud attempts by manipulating the AI chatbot.

‍

5. LLM selection

The moinAI RAG Agent uses the configured LLM for its task. Currently, GPT-4o-Mini is available for selection.

‍

6. Data extraction

Data extraction allows information from the user query or the Knowledge Base to be stored in context variables for use in later actions.

‍

7. Actions & follow-ups

Actions, such as using webhooks, can retrieve and use additional information. Follow-ups can be set to offer additional content or forms.