2.4.3. Embedding Models for Text and Images
💡 First Principle: An embedding model maps content to a point in a vector space where distance means similarity. Retrieval quality therefore depends on two things you configure: which space (the model, which fixes the modalities and languages it understands) and how detailed the coordinates are (the dimension count and numeric type, which trade accuracy against storage and speed).
In 2.2.3, encoding turned categories into numbers by rule. An embedding model learns the numbers, so semantically similar inputs land close together even when they share no words. Three configuration decisions dominate:
1. Modality and model. Use a text embedding model for text-only corpora and a multimodal model when text and images (or audio and video) must be compared with each other.
| Model (Amazon Bedrock) | Inputs | Output dimensions | Notable limits and options |
|---|---|---|---|
| Amazon Titan Text Embeddings V2 | Text | 1,024 (default), 512, 256 | 8,192 tokens; normalize defaults to true; embeddingTypes float and/or binary; optimized for English with multilingual support |
| Amazon Titan Multimodal Embeddings G1 | Text, image, or both | 1,024 (default), 384, 256 | 256 text tokens; images up to 25 MB and 2,048 × 2,048 px; shared text-image space |
| Cohere Embed v3 | Text or one image per call | 1,024 | 512 tokens per text, up to 96 texts per call; input_type distinguishes documents from queries |
| Amazon Nova Multimodal Embeddings | Text, documents, images, video, audio | 3,072, 1,024, 384, 256 | Up to 8K tokens; audio and video segments up to 30 seconds; synchronous and asynchronous APIs |
2. Asymmetric document and query embeddings. Some models embed a stored passage differently from a search query. Cohere Embed requires input_type: embed the corpus with search_document and queries with search_query (other values are classification, clustering and image). Nova Multimodal Embeddings uses an embedding purpose such as GENERIC_INDEX for stored content and GENERIC_RETRIEVAL for queries. Using the document setting for both sides silently lowers retrieval relevance.
3. Dimensions and numeric type. Larger dimensions capture more nuance; smaller ones reduce storage, memory and query latency. Binary embeddings (Titan Text V2 embeddingTypes: ["binary"]) shrink storage further but must be stored in a vector store that supports binary vectors — among Bedrock Knowledge Bases vector stores, only Amazon OpenSearch Serverless and OpenSearch Service managed clusters do.
The consistency rule. The query must be embedded by the same model, with the same dimensions, as the stored vectors, and the vector index dimension must match the model output. Changing the embedding model or dimension count means re-embedding the whole corpus and rebuilding the index — vectors from different models are not comparable, even at identical length.
⚠️ Exam Trap: "Improve accuracy by switching to a better embedding model" is incomplete unless the answer also re-embeds the existing corpus. A new model's query vectors compared against an old model's document vectors return nonsense rankings. Watch for distractors that swap only the query-side model.
