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
No comments yet.
