fix(gateway): correct Codestral embedding dimension#3197
fix(gateway): correct Codestral embedding dimension#3197marius-kilocode wants to merge 5 commits into
Conversation
![kilo-code-bot[bot]](https://avatars.githubusercontent.com/in/2193792?s=60&v=4){kind=small}
| const contentType = response.headers.get('content-type') ?? ''; | ||
| if (!contentType.includes('application/json')) return response; | ||
|
|
||
| const body = await response.clone().json(); |
There was a problem hiding this comment.
WARNING: Unnecessary .clone() leaks the original response body stream
The function calls response.clone().json() to read the body, then constructs a brand-new Response from the re-serialized JSON. The original response body (uncloned) is never consumed or returned, so its readable stream is abandoned — a resource leak for every successful base64 conversion.
Since the original response is not returned or used after this point (a new Response is always returned on success), call response.json() directly to consume the stream:
| const body = await response.clone().json(); | |
| const body = await response.json(); |
| if ( | ||
| !item || | ||
| typeof item !== 'object' || | ||
| !Array.isArray((item as { embedding?: unknown }).embedding) |
There was a problem hiding this comment.
SUGGESTION: as casts here violate project coding standards
Both (item as { embedding?: unknown }) on this line and (item as { embedding: number[] }) on line 62 use the as operator, which the project rules flag as something to avoid in favor of flow-sensitive typing or satisfies.
Since item is already narrowed to object (non-null) by the typeof item !== 'object' guard above, you can use 'embedding' in item to narrow further without any cast:
if (!('embedding' in item) || !Array.isArray(item.embedding)) {
return item;
}
return {
...item,
embedding: floatEmbeddingToBase64(item.embedding as number[]),
};This still needs one cast for item.embedding as number[] (since Array.isArray widens to unknown[]), but that's a much narrower, more defensible cast than the current two.
Code Review SummaryStatus: No Issues Found | Recommendation: Merge Changes Since Last ReviewThe new commit reverted
Files Reviewed (5 files)
Reviewed by claude-sonnet-4.6 · 245,185 tokens |
marius-kilocode
changed the title
1 participant
Kilo indexing uses
/api/gateway/embedding-modelsas the source of truth for local vector-store dimensions. After #3196 stopped forwardingdimensionsto upstream providers,mistralai/codestral-embed-2505returns its native 1536-dimensional vectors.The catalog still advertised Codestral as 256 dimensions, so the client created a 256-dimensional Qdrant collection and then upserted 1536-dimensional vectors. Qdrant rejects that with
Wrong input: Vector dimension error: expected dim: 256, got 1536, which surfaced in the extension as a genericBad Request.This updates the Codestral catalog entry to 1536 so the vector store is created with the dimension the gateway actually returns.