Visual overview
Bedrock models can create embeddings or generate content; S3 often holds source documents for downstream retrieval workflows.
Technical reference
Exact capacities are model-dependent. Use these mechanics to reason about behavior, then verify the chosen model's current documentation.
Token boundaries differ between models; token count affects context use, latency, and often price.
Prompts, messages, retrieved context, tool definitions, and output constraints must fit the model's supported rules.
Encodes semantic information so related inputs can be compared in vector space.
A common semantic-similarity measure; the vector database, not the embedding model, executes retrieval.
Chunk boundaries and size affect how much relevant context retrieval can return without excess noise.
Service limits and capabilities can change. Values shown here reflect the current AWS documentation; use the linked official sources below as the source of truth.
Foundation models are reusable pretrained models
A foundation model is trained on a broad, large-scale dataset and can support many downstream tasks rather than one narrowly defined prediction. Large language models (LLMs) are foundation models focused on language and related modalities, while other FMs can generate or understand images, audio, or video. Transformer architectures underpin many modern text FMs; other model families, including diffusion models, are common for generative media. Model size alone does not determine fitness for a task—latency, modality, context capacity, quality, language support, cost, and governance also matter.
Pretraining gives an FM broad statistical capabilities, but its internal knowledge is not a guaranteed current database of facts. A model generates from learned representations and the context supplied to the request. For application-specific information, a team can provide context in the prompt, retrieve relevant information with RAG, or customize a supported model. Those approaches differ in freshness, cost, operational effort, and whether the model's weights change.
Tokens are the units processed by language models
Text models do not operate on prose exactly as humans see it. A tokenizer converts text into tokens—units that may represent whole words, fragments, punctuation, whitespace patterns, or other encoded pieces depending on the model. Input and generated output therefore consume token budgets that are model-specific. Token counts matter for context limits, latency, and pricing when a service bills by input and output tokens; character count and word count are only rough proxies.
The context window is the amount of tokenized information a model can consider for a request under that model's rules. Instructions, conversation history, retrieved passages, tool descriptions, examples, and the current user input may all compete for this budget. More context is not automatically better: irrelevant material can increase cost and distract the model. Context engineering deliberately selects, structures, compresses, and refreshes the information presented at inference time.
Embeddings represent meaning for retrieval, not generation
An embedding model maps input such as text into a high-dimensional numerical vector. Items with similar meaning should be placed near one another in the embedding space. A vector store can compare an embedded query with stored document vectors and return semantically similar content even when the exact keywords differ. Similarity functions such as cosine similarity are common, although the retrieval engine and index determine the actual search mechanics.
Embedding generation and vector retrieval are separate responsibilities. For example, Amazon Titan Text Embeddings can create vectors, while a supported vector store performs similarity search. In a RAG system, documents are typically split into chunks, embedded and indexed; a user query is embedded, relevant chunks are retrieved, and those chunks become context for a generative model. Chunk size, overlap, metadata, embedding choice, and retrieval strategy can all change quality.
Key takeaways
- 01
Foundation models are broadly pretrained models that can support many tasks and modalities.
- 02
Tokens are model-specific processing units; context limits and token pricing cannot be inferred reliably from word count.
- 03
Context engineering manages the information supplied at inference time rather than changing model weights.
- 04
Embeddings are numeric representations used for semantic comparison and retrieval.
- 05
The embedding model creates vectors; the vector store performs similarity search and indexing.
Official AWS sources
Use these primary AWS resources for the source material behind this article and for deeper reference.