We built Directed Intelligence to tackle the problems we faced with ChatGPT-style applications. It's not a library (yet), but a pattern for building better apps.

Here’s our story:

Why We Built It:

We started by building a "wrapper" very similar to ChatGPT but it sucked.

- Users found responses inconsistent and unreliable.

- Conversations drifted off-topic.

- Advice was generic.

- Experts struggled to steer the model via prompts without technical skills.

Our Solution:

- Combine LLMs with focused scripts written in natural language.

- Active topic management to keep conversations on track.

- The system selects and follows the right scripts based on user input.

How It Works:

- Starts with a base LLM like GPT-4.

- Selects and follows appropriate scripts.

- Monitors context and switches scripts as needed.

- Adds external tools only when scripts call for them.

Real-World Example (https://fforward.ai):

- We built a CMS to manage all our scripts

- We use a state machine and various LLMs to select the appropriate script, monitor context, and swap scripts when needed.

Check out our demo to see it in action: fforward.ai. No signup required.

Looking forward to your feedback!

Cameron Westland, Max Cameron, Jon Friis