fix: prioritize user-configured embedding dimension over model default #8106
Conversation
dosubot
bot
added
size:M
hannesrudolph
added
the
Issue/PR - Triage
![roomote[bot]](https://avatars.githubusercontent.com/in/1546624?s=60&v=4){kind=small}
| /** | ||
| * Gets the current model dimension being used for embeddings. | ||
| * Returns the model's built-in dimension if available, otherwise falls back to custom dimension. | ||
| * Returns the user-configured custom dimension if set, otherwise falls back to model's built-in dimension. |
There was a problem hiding this comment.
Nice fix! The logic correctly prioritizes user-configured dimensions now. Though I wonder if the JSDoc comment could be even clearer by explicitly stating "User-configured dimensions override any model defaults" to make the priority crystal clear?
| // Only use manual dimension if model doesn't have a built-in dimension | ||
| if (!vectorSize && config.modelDimension && config.modelDimension > 0) { | ||
| // Prioritize user-configured custom dimension if explicitly set | ||
| if (config.modelDimension && config.modelDimension > 0) { |
There was a problem hiding this comment.
Good consistency with the config-manager implementation. The if-else structure makes the priority clear.
| }) | ||
|
|
||
| it("should return model's built-in dimension when available", async () => { | ||
| it("should prioritize user-configured dimension over model's built-in dimension", async () => { |
There was a problem hiding this comment.
Excellent test coverage! This test name could be even more descriptive though. How about "should always use user-configured dimension when set, ignoring model defaults" to really emphasize the override behavior?
| expect(configManager.currentModelDimension).toBe(1536) | ||
| expect(mockedGetModelDimension).toHaveBeenCalledWith("openai", "text-embedding-3-small") | ||
| }) | ||
|
|
There was a problem hiding this comment.
Great addition! This test specifically addresses the issue #8102 scenario where user sets 1536 but model default is 1024. This ensures the bug won't regress.
This PR attempts to address Issue #8102. Feedback and guidance are welcome.
Problem
When a user sets an embedding dimension (e.g., 1536), the system was still producing vectors of a different size (e.g., 1024) while the vector database creates a collection with the configured size. This mismatch caused a vector dimension error and blocked indexing and search.
Solution
Updated the dimension priority logic to respect user-configured dimensions as the single source of truth:
config-manager.ts: Now checks user-configured dimension first before falling back to model defaultsservice-factory.ts: Uses the same priority logic for consistencyTesting
Fixes #8102
Important
Prioritize user-configured embedding dimension over model default to prevent dimension mismatch errors.
config-manager.tsandservice-factory.ts.config-manager.spec.tsto reflect new priority logic.This description was created by
for 0c35f22. You can customize this summary. It will automatically update as commits are pushed.