Picking the cheapest embedding model without checking whether your use case is multilingual is one of the most common ways to silently degrade a RAG system's quality on Bedrock.
Titan Embeddings V2 costs $0.00002 per 1,000 tokens and produces 1024-dimensional vectors -- but it supports flexible embedding sizes (1024, 512, 256) while retaining precision even at smaller dimensions: reducing from 1024 to 512 dimensions keeps roughly 99% retrieval precision, and reducing to 256 dimensions keeps 97%. It's an attractive option when vector storage space matters (fewer dimensions = less storage and faster searches).
Cohere Embed costs $0.0001 per 1,000 tokens -- five times more than Titan for English-language workloads -- but supports multiple output formats (float, int8, uint8, binary, ubinary) with configurable dimensions from 256 to 1536, and is specifically designed for cross-language retrieval, mapping text in more than 100 languages into the same vector space.
Titan puts more emphasis on sentence-level semantics, which makes it less sensitive to input noise -- it handles minor typos or spelling variations well. Cohere Embed English v3 focuses on finer text differentiation, at the cost of lower typo tolerance. For a RAG system indexing user-generated content (with real typing errors), this difference can matter more than price.
For English-only workloads with no need for cross-language retrieval, Titan is the cheaper option with no significant quality sacrifice. For any RAG system that needs to search across multiple languages or retrieve content in a language different from the query's, Cohere is the only one of the two that actually solves that problem -- there, the 5x higher cost isn't the relevant variable, multilingual capability is.
Carlos Montiel is an enterprise AI solutions architect. He implements LLMs, Agents, RAG, and orchestrators for companies across Guatemala and Latin America. Reach out for a consultation.
Contact Carlos Montiel