Critino
Critino is revolutionizing AI training by replacing abstract prompts with concrete examples and semantic search, enabling more precise and reliable AI behavior.
Example-based Training
Concrete examples over abstract prompts
Semantic Search
Intelligent matching of examples to match the scenario
Large Scale Data
Vector database to handle the large amounts of examples
The Challenge
While working with a client to create a life coach AI, we hit a wall. The traditional method of guiding LLMs. Prompt engineering was cumbersome and ineffective. Crafting detailed instructions consumed hours, yet still left critical edge cases unresolved. Worse, minor changes to prompts often derailed past improvements, leaving us without confidence in the AI's behavior.
We realized the market was saturated with companies reusing existing solutions instead of innovating. To stand out, we needed a fresh approach that addressed these issues head-on.
The Solution: Critino
Critino is our solution to making LLMs behave in ways consistent with human expectations. By combining two fundamental concepts: few-shot learning with semantic search, Critino is able to use concrete examples ("critiques") to guide AI responses. This shifts the focus from abstract rules in prompts to real-world examples, improving both consistency and alignment.
With Critino, we were also able to enable reinforcement learning of the AI agents that used it, like in Reletino, to allow users to train their agents over time.
Process: How Critino Works
Concrete Examples Over Abstract Rules:
Critiques replace vague, high-level instructions with specific examples of desired behavior. For instance, if a user's AI handles customer complaints, a critique might include a sample query, context, and the optimal empathetic response.
Structured Critique Format:
Every critique includes:
- Context: Detailed context if required to set the scene for the query.
- Query: The input the AI needs to answer.
- Optimal Response: The ideal reply to the query.
- Situation: A general description of the situation for future use in semantic search.
These are subject to, and likely will change as we continue to explore improvements to the platform. But fundamentally, the concept of semantic search to give an LLM relevant examples remains.
Semantic Search Integration:
Critino's semantic search ensures that LLMs generalize effectively by matching new queries with relevant critiques.
Scalability:
As users add more critiques, they aren't limited by context size. Each new critique adds to the available examples that can be referenced, enabling the system to handle an ever-growing range of situations and use cases.
Results
In a quick proof of concept (PoC), Critino demonstrated its ability to address key challenges and deliver meaningful results. After refining the platform and implementing core features like authentication and API integration, we launched an MVP that's ready to transform AI training.
Currently, Critino is being used in all our projects, including Reletino. Because the principles behind Critino are so fundamental in LLM applications, we're excited to use Critino in future projects to save us time and efforts while improving the ability to train agents from feedback.
Summary: Lessons Learned
Master the Fundamentals:
A deep understanding of how LLMs work, "predicting the next word", enabled us to identify that examples often outperform instructions in teaching desired behavior.
Build Fast, Iterate Faster:
Shipping a PoC quickly allowed us to validate Critino's potential and refine its capabilities based on real feedback.
Focus on Unique Value:
Critino's concept isn't novel, it's just a "few-shot example" wrapper. What's unique about Critino is the plug and play ability to start training your agents with intuitive features and APIs.
What's Next for Critino
We're just getting started. Critino is evolving to include:
- Automated Critique Generation: Leveraging AI to create critiques from historical data or user feedback, like YouTube videos or blog articles.
- Deeper Fine-Tuning Integration: Combining critiques with fine-tuning to achieve unparalleled model performance.
Conclusion
Critino represents a bold step forward in making AI more aligned with the desired goals. By focusing on concrete examples, semantic search, and rapid scalability, we've created a platform that not only solves today's alignment challenges but also sets the stage for future innovation.