Embedding Model Selection for Production RAG

Dimension, cost, multilingual capability, deployment architecture — four evaluation criteria. Traditional Chinese and mixed-language contexts carry specific considerations

In production RAG (Retrieval-Augmented Generation) systems, the embedding model selection sets the ceiling for semantic search quality. Embedding models convert text into high-dimensional vectors; the system uses cosine similarity in vector space to determine which content is most relevant to a query. Selection involves four main dimensions: dimension size, cost, multilingual capability, and deployment architecture.

Comparison of Major Embedding Models

OpenAI text-embedding-3 Series

text-embedding-3-small (1536 dimensions) and text-embedding-3-large (3072 dimensions) are the most widely used commercial embedding models. English semantic understanding is stable; integration cost is low. Chinese and multilingual performance is less consistent, and every query requires an API call — no local deployment. Pricing: $0.02/M tokens (small) and $0.13/M tokens (large).

Google Gemini text-embedding Series

text-embedding-004 (768 dimensions) is one of the mainstream choices for multilingual RAG. It outperforms OpenAI's small model on Chinese and Traditional Chinese semantic understanding, and at 768 dimensions, storage and compute costs are lower. Suitable for mixed-language systems handling both English and Chinese content.

Open-Source Local Models

multilingual-e5-large, BGE (BAAI/General Embedding), and others can be deployed locally or on private cloud, with no text sent to external APIs. Full data control, suitable for scenarios with strict data sovereignty requirements. The cost is managing your own inference infrastructure — GPU or high-spec CPU running costs must be accounted for separately.

Traditional Chinese and Mixed-Language Considerations

Embedding quality for content mixing Traditional Chinese, English, and written Cantonese is a key concern for Hong Kong applications. Testing shows that for text combining Traditional Chinese, English, and written Cantonese, Gemini text-embedding-004 produces more stable semantic relevance ranking than OpenAI's small model. Before final selection, run a small-scale recall rate test on actual business data rather than relying solely on public benchmarks.

For the complete bilingual RAG architecture analysis for Hong Kong, see Your Document Is Half Chinese, Half English — This Is Where Most AI Systems Fall.

Dimension vs Storage Cost Trade-off

Higher-dimensional vectors carry stronger semantic representation but increase storage and compute costs linearly. With 1M text records: 768 dimensions (float32) requires approximately 3GB storage, 3072 dimensions requires 12GB. Index build time and query speed in pgvector are also affected by dimension count. For production pgvector configuration details, see Neon Postgres and pgvector: RAG Database Design for Production.

An Overlooked Design Principle

Summary

Embedding model selection is not a one-time decision — as business corpus language distribution changes, the right model may change too. Systems with primarily Chinese content should test the Gemini series first. Data sovereignty requirements point toward open-source local deployment. Always validate recall rate on a small dataset before building at scale.