Skip to main content

Projects

'Project page'

Each Codex Project serves as your control center for managing the safety and reliability of a deployed AI application through Cleanlab. Projects integrate real-time failure detection (Logs) with expert-driven fixes (Remediations) to create a complete AI safety management platform.

View your Projects in the Web App. Create a new Project via our Create/Configure a Project tutorial.

Use separate Codex Projects for different AI applications. Each Project maintains its own isolated safety system, allowing you to manage multiple applications independently while leveraging shared expertise across your organization.

What’s managed in each Project?​

Each Codex Project contains key components for enforcing the safety of your AI application:

Logs: Real-time detection and logging of potential AI failures.

Learn more: Codex Logs

Remediations: Expert-curated fixes and answers that automatically protect against failures.

Learn more: Codex Remediations

Analytics: Complete audit trail of your AI application’s safety improvements over time.

Learn more: Codex Analytics

How it works​

When a Codex Project is integrated with your AI application, it provides continuous safety monitoring and automatic protection through three key mechanisms:

  • Real-time Safety Monitoring: Automatic detection and logging of potential failures with detailed metadata
  • Intelligent Remediation System: SME-created fixes that automatically apply to similar queries and failure patterns
  • Semantic Matching and Protection: Expert answers are returned when similar queries are detected

Optional Project configurations (for advanced developers)​

access_key: API key used by Developers to integrate this Codex Project with an AI application.

max_threshold_distance: Semantic distance (between 0 and 1, inclusive) between a user query and a remediation within the Project. The query and remediation are considered a match if their semantic distance is lower than this threshold. Decrease this value if you find Codex remediations are being applied to queries where a different response is actually desired. To maximize the benefits of Codex, this threshold should be set to the largest value where you never observe the aforementioned problem. We generally recommend values between 0.1 - 0.2.

Remediation Status​

Each Project maintains two key status indicators for its safety system:

  • Remediation Usage: Tracks when expert-provided fixes are actively preventing failures in production
  • Addressed Status: Indicates whether each detected failure (log) has been addressed by an existing remediation