What feature do you want to see added?

Today, the backend initializes heavy AI components during import/startup (embedding model + LLM provider). This makes the startup slow and fragile, especially in local development, CI, and environments where model files or optional dependencies are missing.
I’d like to add lazy, fault-tolerant model initialization so the API can start reliably and only initialize heavy models when they are actually needed.

Why this helps

• Prevents startup failures caused by missing model artifacts/dependencies
• Improves dev-lite and test experience (fewer import-time side effects)
• Reduces cold-start pain and makes behavior more predictable
• Supports graceful degraded mode when LLM/retrieval is unavailable

Proposed behavior

• Replace eager globals (e.g., embedding/LLM singletons created at import time) with cached lazy getters
• Initialize models on first use with thread-safe guards
• Return graceful fallback responses when model initialization fails
• Expand health/readiness info to show model availability state explicitly
• Add tests for lazy init, failure paths, and “no heavy load at import time”

Acceptance criteria

• Importing service modules does not trigger heavy model loading/downloading
• API starts even if model files are missing
• Chat endpoints degrade gracefully when LLM is unavailable
• Retrieval errors are handled without crashing the process
• The health endpoint clearly reflects the readiness of major model components

Upstream changes

No known upstream dependency yet.
If there are related Jenkins plugin or API reliability discussions/PRs, I can link and align this issue to avoid duplicate work.

Are you interested in contributing this feature?

Yes — I’m interested in contributing this.

I can start with:

1. a small design proposal (lazy getter + cache + error policy),
2. a focused PR for backend wiring changes, and
3. follow-up tests for startup/import behavior and degraded-mode responses.

If maintainers have preferences regarding module structure or the health endpoint schema, I’m happy to align on them before opening the PR.