NotifAI Performance Improvements

[PaulTaylor]

Description

Improve performance of the model pipeline to more accurately classify documents as admail/business.

• Add a docker compose stack for local development
• Update the system prompt
• Remove text extraction and use multi-modality in llm instead
• Switch the default model to Nova-Lite
• Update evaluation framework to account for new pipeline
• Update dependencies to fix vulnerabilities (for packages used in our code, there are some dependabot issues that may remain in the wider template (eg. in docs))
• Disable the schedule for the automated evaluation as it always returns the same result from the same model/docs

Context

The current model and prompt are generating false-negatives for letters that should be classified as admail. Misclassifying letters that are appropriate for admail as business incurs unnecessary postage costs.

Type of changes

• Refactoring (non-breaking change)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would change existing functionality)
• Bug fix (non-breaking change which fixes an issue)

Checklist

• I am familiar with the contributing guidelines
• I have followed the code style of the project
• I have added tests to cover my changes
• I have updated the documentation accordingly
• This PR is a result of pair or mob programming

---

Sensitive Information Declaration

To ensure the utmost confidentiality and protect your and others privacy, we kindly ask you to NOT including PII (Personal Identifiable Information) / PID (Personal Identifiable Data) or any other sensitive data in this PR (Pull Request) and the codebase changes. We will remove any PR that do contain any sensitive information. We really appreciate your cooperation in this matter.

• I confirm that neither PII/PID nor sensitive data are included in this PR and the codebase changes.