Fix: Ensure proper embedding dimensions

[Daniel-Vaz]

Notes

• Fixes silent INSERT failures by replacing NULL with empty strings for branch and repo in vec0 TEXT columns.
• Embedding dimension is now set dynamically based on model name (e.g., 1536 for text-embedding-3-small).

[djannot]

@Daniel-Vaz I'm already fixing it in #53

[Daniel-Vaz]

@djannot
Thank you !
I was just today noticing this behaviour. Could you please also address then the static dimension sizes in that same PR ? If not I can at least change this PR to just focus on that instead.

Locally I'm testing using Azure Open AI text-embedding-3-small model, and the dimension size difference from the one statically defined makes all my queries fail even when using that same model to build the db.

[djannot]

I'll ping you as soon as the other PR is merged, so you can rebase yours from main at that time, ok ?

[djannot]

Done, you can rebase

[Daniel-Vaz]

@djannot Done. Feel free to have a look whenever you can.

[djannot]

npm test is returning errors

[Daniel-Vaz]

My bad. Updated the test accordingly:

 Test Files  7 passed (7)
      Tests  456 passed (456)
   Start at  17:00:21
   Duration  6.72s (transform 1.48s, setup 0ms, import 7.33s, tests 7.30s, environment 1ms)

[djannot]

Thanks. I used AI to analyze it and got some good feedback:

1. Default value discrepancy - getEmbeddingDimension defaults to 1536 for unknown models, but all the function signatures default embeddingDimension parameter to 3072. This inconsistency could cause mismatches if someone passes a default through one path vs. the other. I'd recommend making the default consistent everywhere, or removing defaults from the function signatures to force callers to always pass the value explicitly.
2. console.warn in a library function - utils.ts:115 uses console.warn directly instead of the project's Logger class. This is inconsistent with the rest of the codebase and won't be captured by the logging infrastructure.
3. Hardcoded model knowledge is fragile - The model-to-dimension mapping will need updates whenever new models appear. Consider:
   • Making this configurable in the YAML/config (e.g., an embedding_dimensions field)
   • Falling back to a test API call to determine dimensions at startup
   • At minimum, documenting which models are supported
4. getMetadataValue in Qdrant makes an extra API call (database.ts:281-290) - Every metadata read now calls getCollection to discover the dimension. This adds latency. It would be better to pass the dimension as a parameter here too, like the other methods.

[Daniel-Vaz]

Thanks will look into this next week !