Incorrectly setting the vector dimensions for Qdrant for semantic search leads to a huge degradation in search results quality.

[TurboPlanner]

Plugin Type

VSCode Extension

App Version

4.140.1 (7fe600c)

Description

The most powerful embedding provider for KiloCode is the Qwen 3 Embedding 8b model with 4096 dimensions. This model ranks significantly higher in vectorization benchmarks than embeddings from OpenAI and other vendors.

However, Qdrant settings show that Kilo Code has recently been causing significant damage to vector databases by incorrectly configuring them for the vector dimensionality. While previously the databases were correctly created with 4096 dimensions, now 768 dimensions are being erroneously selected.

Somehow, this still works, probably due to vector conversion, but it's a COMPLETE DISASTER in terms of search quality. Reducing the vector dimensionality in this way makes KiloCode's search not just stupid, but terribly stupid. I conducted various semantic tests, and I can say that the quality of semantic search is now simply destroyed, as it often produces completely nonsensical results.

Reproduction steps

1. Delete Index
2. Select Qwen 3 Embedding 8b for embeddings (reproduced with other Qwen models too)
3. Start index creation
4. Open Qdrant Collection on administration site for checking REAL vector dimension

Provider

No response

Model

No response

System Information

No response