Refactor/embedding service cleanup

[hinxcode]

Description

Make the offline embedding script use the new BaseEmbeddingService (it was bypassing it), stop the factory
from loading a model at import time, and fix SigLIP's scoring formula.

Motivation and Context

1. Abstraction not used in the offline path. generate_embeddings.py reached into service.model / .processor directly, so adding a new model would require touching the batch script. Now it calls service.encode_image().

2. Module-level singleton in embedding_service_factory. Importing the module triggered model download and collided with the existing vector-store embedding_service singleton.

3. SigLIP scoring used CLIP's formula (logit_scale * sim) instead of SigLIP's (sigmoid(logit_scale * sim + logit_bias)).

4. transformers>=5.0.0 forced CLIP-only users to upgrade. SigLIP needs only >=4.42.0.

Type of Change

• Bug fix (non-breaking change that fixes an issue)
• New feature (non-breaking change that adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Documentation update
• Code refactoring (no functional changes)
• Performance improvement
• Research contribution (new models, evaluation methods, etc.)
• Other (dependency cleanup):

Component(s) Affected

• Backend (Python/FastAPI)
• Frontend - Photographs
• Frontend - Maps
• Frontend - Documents
• CLIP/ML models
• Configuration
• Documentation
• Tests
• Build/deployment

Changes Made

• Use BaseEmbeddingService.encode_image() in offline pipeline; drop src/models/clip/ layer
• Remove module-level singleton from factory; search.py owns its instance
• Add transform_score abstract method; SigLIP applies sigmoid(logit_scale * sim + logit_bias)
• EmbeddingService.search() takes optional score_transform callable instead of scalar logit_scale
• Stop re-exporting embedding_service from services/__init__.py (avoids name collision)
• transformers>=4.42.0; restore config.json default to clip + cuda
• Add pyproject.toml with black + isort config
• Remove dead or "clip" fallback in factory
• Update test_search_with_logit_scale → test_search_with_score_transform to match new callable API
• Bump CI to Python 3.12 and remove continue-on-error from the pytest job (was masking failures)

Testing

How Has This Been Tested?

• pytest tests/ passes
• Manual: text and image search via FastAPI on both CLIP and SigLIP

Test Configuration

• Collection type tested: photographs
• Python version: 3.12
• Node version: 22
• OS: macOS
• CLIP model: openai/clip-vit-base-patch32
• SigLIP model: google/siglip-base-patch16-224

Screenshots (if applicable)

Checklist

Code Quality

• My code follows the project's coding standards
• I have run black . and isort . on Python code
• I have run npm run lint on frontend code (if applicable)
• I have performed a self-review of my own code
• I have commented my code, particularly in hard-to-understand areas
• My changes generate no new warnings or errors

Testing

• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes
• I have tested this locally with actual data

Documentation

• I have updated the documentation accordingly
• I have updated the README if needed
• I have added docstrings to new functions/classes
• I have updated config.json documentation if config changes were made

Dependencies

• I have updated requirements.txt (if Python dependencies changed)
• I have updated package.json (if Node dependencies changed)
• I have documented any new configuration options

Research (if applicable)

• I have included references to relevant papers or research
• I have shared evaluation results or benchmarks
• I have included information about datasets used
• I have documented model training procedures

Breaking Changes

Two internal-API changes; both updated in this PR:

• EmbeddingService.search(): parameter logit_scale: Optional[float] replaced with score_transform: Optional[Callable[[Tensor], Tensor]].
• services/__init__.py: no longer re-exports embedding_service (it pointed to the factory's encoder, conflicting with the vector-store singleton). Import from embedding_service_factory directly.

Additional Notes

The transform_score abstraction makes explicit that absolute scores differ across models. Benchmarking should rely on rank-based metrics rather than score thresholds.

Reviewers Checklist (for maintainers)

• Code quality and style compliance
• Test coverage adequate
• Documentation complete
• No security concerns
• Performance implications acceptable
• Breaking changes documented

[ktamaral]

@hinxcode Thank you for making these updates, everything looks good. I was able to generate new embeddings and run the unit tests successfully.