Use open-source embeddings (BGE, Voyage, Nomic) for cost control and alignment with Scout; ensure embeddings and generation models come from compatible training.
Your embedding model converts documents and queries into vectors that Zilliz Cloud stores and retrieves. Misalignment between embeddings and Scout causes semantic drift: embed with generic web-trained models but generate with Scout (trained on math, code, reasoning), and retrieval may miss relevant content. BGE (BAAI General Embeddings) is popular for Scout deployments because it aligns well with Llama models and is freely available. Voyage and Nomic embeddings are also strong and compatible. For maximum cost control, use smaller embeddings (BGE-base 768 dims) for speed, larger embeddings (BGE-large 1024 dims) for accuracy—Zilliz Cloud indexes both efficiently.
With Zilliz Cloud, embedding selection affects retrieval quality, which directly impacts Scout's answer quality. Better embeddings = fewer false positives for Scout to filter. Run offline benchmarks: embed your domain docs with candidate embeddings, create test queries, measure recall in Zilliz Cloud, then choose the best accuracy/cost trade-off. Zilliz Cloud's analytics show retrieval metrics; use these to validate your embedding choice in production.
Related Resources
- Vector Embeddings — embedding selection guide
- Getting Started with LlamaIndex — embedding integration patterns
- Retrieval-Augmented Generation (RAG) — embeddings in RAG pipelines