Content that

Unlike ChatGPT, RELAYTO AI Content Chat operates within a controlled environment where the content of the PDF document is the primary source of information. It does not generate responses based on external data sources or access to the internet. This controlled and contained nature may alleviate concerns related to compliance and policy restrictions.

When a particular document from a hub is opened, Content Chat will answer questions using the information given in documents from a hub. If a document is opened from the dashboard or via a direct link, only chat for that particular document will be shown. In the case of a hub, when a hub is opened, it will answer questions based on all given contents, including contents from subhubs if there are any.

By default, we use OpenAI’s GPT-4 model to power the AI chat experience. However, this can be reconfigured to use our in-house deployment of LLaMA 4, which is currently in beta. Both models are large language models (LLMs) capable of advanced natural language understanding and generation.

The AI chat is tightly integrated into our platform via a retrieval-augmented generation (RAG) architecture. Here’s how it works:

1. Content from documents is processed and converted into embeddings, which are stored and indexed within our AWS-based infrastructure.

2. When a user asks a question, we perform a similarity search to identify the most relevant content chunks.

3. These relevant chunks are then passed along with the user’s question to the language model (e.g., GPT-4) to generate a grounded response that references specific content.

While the core language model (e.g., GPT-4) is not retrained on customer data, our system uses reinforcement mechanisms to improve performance over time.

This includes monitoring incorrect answers, collecting user feedback, and fine-tuning the retrieval process (e.g., improving document chunking and prompt engineering). These enhancements help ensure more accurate citations and responses without modifying the model itself.

Customization of the initial three prompt questions is on our roadmap. We plan to release this feature in early Q3 2025, which will allow tailored onboarding prompts based on the context of each experience.

Yes. We use industry-standard security protocols to protect your content end-to-end.