July 27th, 2023

Eric Ciarla imageEric Ciarla

Introducing Precise Mode for Mendable.ai

A Solution for More Controlled AI Conversations

Introducing Precise Mode for Mendable.ai image

Are you tired of asking, “How can I get my model to not talk about X?” At Mendable.ai, we understand this concern. As creators and innovators in the AI space, we have heard this query more times than we can count. So, we decided to rise to the occasion and tackle this issue head-on. We are excited to unveil our latest feature - Precise Mode, developed with the assistance of our partners at LangChainAI.

Understanding How Precise Mode Works

Complexity is often seen as a mark of genius. However, sometimes the best solutions are the simplest. Our approach to refining AI conversation abides by this ethos, comprising two key components: a moderation layer and prompt templates.

The Moderation Layer

The first step in achieving more controlled AI responses is our moderation layer. It functions as a gatekeeper, flagging any requests that contain text we, or the client, would prefer the model not to attempt answering. For instance, any sensitive topics, inappropriate content, or areas outside the model’s training can be flagged. If a user request is flagged, we provide feedback through a friendly, informative message, thus maintaining the user experience’s quality while ensuring content suitability.

Prompt Templates

After the moderation layer comes the exciting part - prompt templates. These are developed in collaboration with LangChainAI, designed to further constrict the model’s response parameters. For Precise Mode, we specify instructions like “Do not make up any part of an answer. If the answer isn’t in the sources, say X”. This approach ensures the AI responses adhere closely to the information available, providing the most accurate and helpful replies.

The Advantage of GPT-4 in Following Prompts

While we have designed the Precise Mode to be effective with various AI models, it’s worth noting that GPT-4 has shown particularly promising results. GPT-4 follows these prompt directions far more reliably than its predecessor, GPT-3.5. If you’re aiming for more consistent and controlled AI interactions, GPT-4 is the recommended choice with Precise Mode.

However, we understand that not everyone has transitioned to GPT-4, and GPT-3.5 remains a popular choice. Don’t worry - we have you covered! There are specific prompt techniques that can considerably enhance the consistency of GPT-3.5. We are excited to share these tips in the future.

In conclusion, our latest feature - Precise Mode, aims to answer a longstanding question in AI conversation control. By combining a moderation layer with highly customized prompt templates, we offer a simple yet effective solution to controlling the topics your AI model can discuss. Welcome to a more precise, accurate, and controlled AI conversation experience with Mendable.ai. Stay tuned for more exciting updates and features!