Vector dimension for code index only accepts 1536

[mkdir700]

App Version

v3.23.6

API Provider

Google Gemini

Model Used

text-embedding-004

Roo Code Task Links (Optional)

No response

🔁 Steps to Reproduce

1. setup(macOS, roo code v3.23.6)
2. Select Google Gemini as the embedding model provider
3. Choose text-embedding-004 as the embedding model
4. Qdrant can be selected locally or in the cloud
5. Save and start building the index
6. The number of discoveries is normal, but the number of completed ones remains at 0.

💥 Outcome Summary

I tried both the ollama and Google Gemini embedding models, but both failed. Checking the logs revealed that the vector dimension was required to be 1536. Eventually, I switched to OpenAI's text-embedding-3-small model, and everything worked fine.

I hope to use other embedding modes instead of being limited to models with a vector dimension of 1536.

📄 Relevant Logs or Errors (Optional)

Qdrant Logs:


2025-07-10T13:39:33.516083Z  WARN collection::shards::replica_set::update: Failed to update shard ws-7d00c287d41ff44f:0 on peer 4890243171819253, error: Wrong input: Vector dimension error: expected dim: 1536, got 768
2025-07-10T13:39:33.516336Z  INFO actix_web::middleware::logger: 192.168.65.1 "PUT /collections/ws-7d00c287d41ff44f/points?wait=true HTTP/1.1" 400 111 "-" "Roo-Code" 0.012018
2025-07-10T13:39:33.734444Z  WARN collection::collection_manager::collection_updater: Update operation declined: Wrong input: Vector dimension error: expected dim: 1536, got 768
2025-07-10T13:39:33.734513Z  WARN collection::shards::replica_set::update: Failed to update shard ws-7d00c287d41ff44f:0 on peer 4890243171819253, error: Wrong input: Vector dimension error: expected dim: 1536, got 768
2025-07-10T13:39:33.734693Z  INFO actix_web::middleware::logger: 192.168.65.1 "PUT /collections/ws-7d00c287d41ff44f/points?wait=true HTTP/1.1" 400 111 "-" "Roo-Code" 0.008853
2025-07-10T13:39:33.754102Z  WARN collection::collection_manager::collection_updater: Update operation declined: Wrong input: Vector dimension error: expected dim: 1536, got 768
2025-07-10T13:39:33.754166Z  WARN collection::shards::replica_set::update: Failed to update shard ws-7d00c287d41ff44f:0 on peer 4890243171819253, error: Wrong input: Vector dimension error: expected dim: 1536, got 768
2025-07-10T13:39:33.754347Z  INFO actix_web::middleware::logger: 192.168.65.1 "PUT /collections/ws-7d00c287d41ff44f/points?wait=true HTTP/1.1" 400 110 "-" "Roo-Code" 0.008354

[daniel-lxs]

Hey, thanks for providing this detailed issue report! After investigating, I've confirmed that the problem is a vector dimension mismatch. The application is trying to use a Qdrant collection created for a 1536-dimension model with a new 768-dimension model, which Qdrant correctly rejects.

While there is logic in src/services/code-index/vector-store/qdrant-client.ts to delete and recreate the collection when the dimension changes, it seems this process is failing silently. The generic error handling is likely masking the root cause.

Workaround:
For now, you can work around this by manually clearing the index data. This will force the creation of a new, correctly-sized collection when you start indexing again.

A more permanent solution would be to wrap the recreation logic in a dedicated try...catch block. This would allow us to throw a more specific error and halt indexing if the collection can't be recreated.

For example, we could modify the initialize() method in qdrant-client.ts like this:

// ... inside the else block where existingVectorSize !== this.vectorSize
} else {
    // Exists but wrong vector size, recreate
    try {
        console.warn(
            `[QdrantVectorStore] Collection ${this.collectionName} exists with vector size ${existingVectorSize}, but expected ${this.vectorSize}. Recreating collection.`,
        );
        await this.client.deleteCollection(this.collectionName);
        await this.client.createCollection(this.collectionName, {
            vectors: {
                size: this.vectorSize,
                distance: this.DISTANCE_METRIC,
            },
        });
        created = true;
    } catch (recreationError) {
        const errorMessage = recreationError instanceof Error ? recreationError.message : String(recreationError);
        console.error(`[QdrantVectorStore] CRITICAL: Failed to recreate collection ${this.collectionName} for new vector size. Error: ${errorMessage}`);
        throw new Error(`Failed to update vector index for new model. Please try clearing the index and starting again. Details: ${errorMessage}`);
    }
}

This change would provide much clearer feedback if something goes wrong during the model switch and prevent the indexer from getting into a broken state.