Skip to content

amazon-bedrock

@ai-sdk/amazon-bedrock

AIProviders.amazonBedrock

  • Language
  • Embedding
  • Image
  • Reranking

Explicit disabled reasoning removes unused effort and budget fields. Explicit provider effort/budget overrides portable defaults; type/display-only settings still derive missing values. Portable max remains max for supported families and maps to high with a warning for Nova 2.

Titan, Cohere and Nova embeddings accept top-level dimensions unless the provider-specific dimensions override them. Haiku 5.5 uses the JSON tool fallback where Bedrock native structured output is unsupported; explicit strict: false does not warn. SigV4 excludes non-ASCII header values from signed headers while retaining outgoing values.

Bedrock only sends strict tool definitions and native structured-output fields to recognized supported Claude generations. Unknown future Claude IDs and unrecognized minor versions use JSON instructions in automatic structured-output mode, retaining ordinary application tools without adding an incompatible synthetic response tool. Explicit jsonTool selection keeps its existing model-specific behavior.

Claude Sonnet 4.6 and Haiku 4.5 retain strict tools but use the JSON-tool fallback for unreliable native output. Non-Claude models and opaque inference-profile IDs keep their existing capability behavior. The Anthropic Messages adapter also inherits normalized citations, web-search source deduplication, manual-thinking tool-choice safeguards, and replayable file-reference/tool-search-error transcripts.

Amazon Bedrock Converse, embeddings, images, and reranking mirror upstream 5.0.124: SigV4 and bearer auth, versioned user-agent suffix, Converse request/stream parsing, ordered generated content with signature, legacy redacted-data, and current redacted-content replay metadata, event-stream decode/process failure propagation and incomplete-frame EOF rejection, empty tool-call ID normalization, provider options under amazonBedrock/bedrock, service tiers, guardrails, top-level reasoning mapping, cache points, Converse inline and S3 video sources across ten supported MIME types, document citations, rich tool-result content, strict-tool warnings and JSON response fallback that preserves ordinary tools, Cohere batch embeddings up to 96 inputs, Nova/Titan/Cohere embedding response shapes, and the current bedrockRerankingConfiguration rerank request key. The latest pass omits assistant turns left empty after unsigned-reasoning filtering, and also preserves exact case-sensitive s3:// image sources in user and tool-result content, sanitizes replayed tool names, percent-encodes ARN model paths, and omits strict/native structured-output fields for Claude families that Bedrock rejects. The 5.0.68 pass also adds inference-profile reasoning, cache-only assistant omission, ARN embedding-family detection, citations, parallel-tool disabling, guardrail block handling, complete raw usage, model-specific reasoning shapes, and typed stream errors. ProviderV4 type names, ESM-only packaging, Node 22 engines, optional JS headers config, and workflow serialization helpers are JS-only upstream concerns. The 5.0.76 delta adds structuredOutputMode, separate native-output and strict-tool model gates including Claude Sonnet 4.6 and Haiku 4.5, sibling-safe output_config filtering, citation-nested text extraction, sanitized document names, and alternating user turns for replayed provider-executed tool results. The 2026-09-20 pass adds opaque Anthropic-family settings, recursive strict-schema validation, current web-tool warnings/bindings, and non-2xx stream error parity. The 2026-09-27 pass adds current model capability gates, forced-tool and sampling rules, reasoning normalization, Claude Opus 5.5 and Mistral ID handling, and Mantle routing; Swift still cannot distinguish omitted from explicitly empty stopSequences. The 2026-10-04 pass passes string requestMetadata, omits Nova 2 Lite maxTokens for high thinking, validates generated region hosts, and reports Nova Canvas image-input capabilities. The 2026-10-11 pass: Claude Haiku 5.5 capability gates, explicit disabled thinking, max reasoning and embedding dimensions follow the published baseline. SigV4 omits non-ASCII header values from the signed set while retaining outgoing headers. The 2026-10-18 pass: Recognizable supported Claude generations receive strict/native-output fields; unknown Claude IDs use conservative prompt JSON instructions in automatic mode. Non-Claude and opaque inference profiles retain their prior behavior.

Factory argument requirements are defined by the public Swift factory signatures. Use Public symbols when you need the exact initializer or factory declaration.

Return to the provider matrix.