AWS Bedrock Plugin for Genkit Go

Genkit Go plugin for Amazon Bedrock. It supports Converse text generation, streaming, tool calling, multimodal and document inputs, image generation, embeddings, prompt caching, and Cohere reranking through Bedrock.

Features

• Text generation through the Bedrock Converse API.
• Streaming with text, tool-use, and reasoning block reassembly.
• Tool calling with schema conversion and typed tool input coercion.
• Multimodal inputs for supported vision and document models.
• Image generation for Titan Image, Nova Canvas, Stable Diffusion XL, and modern Stability models.
• Embeddings for Titan text, Titan multimodal, Cohere text/image, and Nova text models.
• Reranking with Cohere Rerank models through Bedrock InvokeModel.
• Prompt caching with Bedrock cache point parts.
• Inference profiles for regional and global Bedrock profile IDs.

Installation

go get github.com/xavidop/genkit-aws-bedrock-go

Quick Start

package main

import (
	"context"
	"log"

	"github.com/firebase/genkit/go/ai"
	"github.com/firebase/genkit/go/genkit"
	bedrock "github.com/xavidop/genkit-aws-bedrock-go"
)

func main() {
	ctx := context.Background()

	bedrockPlugin := &bedrock.Bedrock{
		Region: "us-east-1", // Optional if the AWS SDK region chain is configured.
	}
	g := genkit.Init(ctx, genkit.WithPlugins(bedrockPlugin))

	model := bedrockPlugin.DefineModel(g, bedrock.ModelDefinition{
		Name: "amazon.nova-lite-v1:0",
		Type: "chat",
	}, nil)

	resp, err := genkit.Generate(ctx, g,
		ai.WithModel(model),
		ai.WithPrompt("What are the key benefits of using AWS Bedrock?"),
		ai.WithConfig(&bedrock.Config{MaxTokens: 512}),
	)
	if err != nil {
		log.Fatal(err)
	}
	log.Println(resp.Text())
}

You can also register common aliases:

models := bedrock.DefineCommonModels(bedrockPlugin, g)
embedders := bedrock.DefineCommonEmbedders(bedrockPlugin, g)

resp, err := genkit.Generate(ctx, g,
	ai.WithModel(models["nova-lite"]),
	ai.WithPrompt("Summarize what Bedrock does."),
)
_ = embedders
_ = resp
_ = err

AWS Configuration

The plugin uses the AWS SDK for Go v2 configuration chain. Set Bedrock.Region for an explicit region, or leave it empty and configure one with AWS_REGION, AWS_DEFAULT_REGION, shared config, SSO, or the runtime environment. If no region is resolved, initialization fails with a clear error.

bedrockPlugin := &bedrock.Bedrock{
	Region:         "us-west-2",
	MaxRetries:     3,
	RequestTimeout: 30 * time.Second,
	AWSConfig:      customAWSConfig,
}

Option	Default	Description
Region	AWS SDK region chain	Optional explicit region override.
MaxRetries	3	AWS SDK retry attempts when loading default config.
RequestTimeout	30s	Per-call timeout for generation, embedding, image, and rerank calls.
AWSConfig	nil	Full AWS SDK config override for credentials, endpoint, HTTP client, or tests.

Required permissions usually include:

{
  "Effect": "Allow",
  "Action": [
    "bedrock:InvokeModel",
    "bedrock:InvokeModelWithResponseStream"
  ],
  "Resource": "*"
}

Model access is managed per AWS account and region in the Bedrock console. Request access for the model families you plan to use before running examples or live tests.

Models and Inference Profiles

Pass direct Bedrock model IDs or inference profile IDs to DefineModel, DefineEmbedder, genkit.WithDefaultModel, and Rerank.

// Direct model ID.
"amazon.nova-lite-v1:0"

// Regional or global inference profile ID.
"us.anthropic.claude-3-7-sonnet-20250219-v1:0"
"global.anthropic.claude-haiku-4-5-20251001-v1:0"

The plugin preserves the full model ID when calling Bedrock. For local Genkit metadata only, it strips known profile prefixes (us., eu., apac., jp., au., global., us-gov.) before looking up capability metadata. Unknown chat models remain callable and are marked unstable in metadata.

Generation Configuration

Use bedrock.Config for typed Converse configuration:

temperature := float32(0.2)
topP := float32(0.9)

resp, err := genkit.Generate(ctx, g,
	ai.WithModel(model),
	ai.WithPrompt("Answer briefly."),
	ai.WithConfig(&bedrock.Config{
		MaxTokens:    512,
		Temperature:  &temperature,
		TopP:         &topP,
		ToolChoice:   bedrock.ToolChoiceAuto,
		StopSequences: []string{"END"},
	}),
)

ai.GenerationCommonConfig and legacy map[string]any configs are still accepted for compatibility. Use AdditionalModelRequestFields for model-specific Converse fields such as Claude extended thinking:

resp, err := genkit.Generate(ctx, g,
	ai.WithModel(model),
	ai.WithPrompt("Think carefully, then answer."),
	ai.WithConfig(&bedrock.Config{
		MaxTokens: 4096,
		AdditionalModelRequestFields: map[string]any{
			"thinking": map[string]any{
				"type":          "enabled",
				"budget_tokens": 1024,
			},
		},
	}),
)

Tool Calling

Define tools with Genkit and pass them to genkit.Generate. ToolChoice may be auto, any, required, none, or the name of a declared tool.

weatherTool := genkit.DefineTool(g, "get_weather", "Get weather for a city.",
	func(ctx *ai.ToolContext, input struct {
		Location string `json:"location"`
	}) (string, error) {
		return "Sunny, 22C", nil
	})

resp, err := genkit.Generate(ctx, g,
	ai.WithModel(model),
	ai.WithPrompt("Use the weather tool for London."),
	ai.WithTools(weatherTool),
	ai.WithConfig(&bedrock.Config{
		MaxTokens:  512,
		ToolChoice: "get_weather",
	}),
)

Image Generation

Define image models with Type: "image". Generated images are returned as image/png media parts with data:image/png;base64,... URLs.

Supported families:

• amazon.titan-image-generator-*
• amazon.nova-canvas-*
• stability.stable-diffusion-xl-*
• stability.sd3-*
• stability.stable-image-*

Titan Image and Nova Canvas accept nested imageGenerationConfig overrides. Stable Diffusion XL accepts flat Stability fields. Modern Stability models currently use fixed plugin defaults and ignore extra config fields.

Embeddings

Define embedders with the Bedrock model ID:

titan := bedrockPlugin.DefineEmbedder(g, "amazon.titan-embed-text-v2:0")

resp, err := genkit.Embed(ctx, g,
	ai.WithEmbedder(titan),
	ai.WithTextDocs("Bedrock provides managed foundation models."),
)

Supported families:

• Titan text: amazon.titan-embed-text-v1, amazon.titan-embed-text-v2:0
• Titan multimodal: amazon.titan-embed-image-v1
• Cohere text and image: cohere.embed-english-v3, cohere.embed-multilingual-v3
• Nova text: amazon.nova-embed-text-v1:0

Reranking

Genkit Go does not yet expose a first-class reranker action, so this plugin provides Rerank. It reuses the initialized Bedrock plugin client and returns documents ordered by score with ai.RankedDocumentMetadata.

resp, err := bedrock.Rerank(ctx, g, "cohere.rerank-v3-5:0", &ai.RerankerRequest{
	Query: ai.DocumentFromText("Which document explains Bedrock authentication?", nil),
	Documents: []*ai.Document{
		ai.DocumentFromText("Configure AWS credentials with environment variables or AWS SSO.", nil),
		ai.DocumentFromText("Titan Image Generator returns base64-encoded PNG data.", nil),
	},
	Options: &bedrock.RerankOptions{TopN: 1},
})
if err != nil {
	log.Fatal(err)
}
for _, doc := range resp.Documents {
	log.Printf("score=%.3f text=%s", doc.Metadata.Score, doc.Content[0].Text)
}

Prompt Caching

Use bedrock.NewCachePointPart() in system or message content where Bedrock supports prompt caching. Cache points are passed through to Converse as Bedrock cache point blocks.

resp, err := genkit.Generate(ctx, g,
	ai.WithModel(model),
	ai.WithMessages(
		ai.NewSystemMessage(
			ai.NewTextPart(largeReusablePrompt),
			bedrock.NewCachePointPart(),
		),
		ai.NewUserTextMessage("Summarize the policy."),
	),
)
if err == nil {
	log.Println("cache read tokens:", resp.Usage.CachedContentTokens)
}

Media and Document Inputs

Media inputs must use a supported MIME type and a base64 data URL or bare base64 payload. Remote URLs, raw non-base64 content, missing content types, and unknown MIME types are rejected before calling Bedrock.

Supported document MIME types include PDF, CSV, DOC, DOCX, XLS, XLSX, HTML, plain text, and Markdown. Supported image inputs include common Bedrock image formats such as PNG, JPEG, WebP, and GIF, depending on the target model.

Examples

git clone https://github.com/genkit-ai/aws-bedrock-go-plugin
cd aws-bedrock-go-plugin

go run ./examples/basic
go run ./examples/streaming
go run ./examples/tool_calling
go run ./examples/image_generation
go run ./examples/embeddings
go run ./examples/reranking
(cd examples/multimodal && go run .)
go run ./examples/document /path/to/document.pdf
go run ./examples/prompt_caching

The examples make live Bedrock calls. Configure AWS credentials, region, and model access before running them. Set BEDROCK_REGION to override the AWS SDK region chain for any example.

Model overrides:

• BEDROCK_MODEL: text, streaming, tool, advanced schema, document, and multimodal fallback.
• BEDROCK_VISION_MODEL: multimodal image-input model.
• BEDROCK_DOCUMENT_MODEL: document/PDF model.
• BEDROCK_PROMPT_CACHING_MODEL: prompt-caching-capable model. Required for examples/prompt_caching because support and marketplace access vary by model/account.
• BEDROCK_EMBED_MODEL: embedding model.
• BEDROCK_IMAGE_MODEL: image generation model. Required for examples/image_generation because image model IDs vary by region and model lifecycle.
• BEDROCK_RERANK_MODEL: reranking model.

Opt-in Live Tests

Live tests are skipped by default. Pass -test-bedrock-region and the model flag for each family you want to exercise.

go test . -run TestBedrockLive_ClaudeSync \
  -test-bedrock-region=us-east-1 \
  -test-bedrock-model-claude=us.anthropic.claude-haiku-4-5-20251001-v1:0

go test . -run TestBedrockLive_NovaSync \
  -test-bedrock-region=us-east-1 \
  -test-bedrock-model-nova=amazon.nova-lite-v1:0

go test . -run TestBedrockLive_ClaudeStreamingToolCall \
  -test-bedrock-region=us-east-1 \
  -test-bedrock-model-claude=us.anthropic.claude-haiku-4-5-20251001-v1:0

go test . -run 'TestBedrockLive_.*Image' \
  -test-bedrock-region=us-east-1 \
  -test-bedrock-image-titan=amazon.titan-image-generator-v2:0 \
  -test-bedrock-image-nova-canvas=amazon.nova-canvas-v1:0 \
  -test-bedrock-image-sdxl=stability.stable-diffusion-xl-v1:0 \
  -test-bedrock-image-modern-stability=stability.sd3-large-v1:0

go test . -run 'TestBedrockLive_.*Embed' \
  -test-bedrock-region=us-east-1 \
  -test-bedrock-embed-titan-text=amazon.titan-embed-text-v1 \
  -test-bedrock-embed-titan-text-v2=amazon.titan-embed-text-v2:0 \
  -test-bedrock-embed-titan-multimodal=amazon.titan-embed-image-v1 \
  -test-bedrock-embed-cohere-text=cohere.embed-english-v3 \
  -test-bedrock-embed-cohere-multilingual=cohere.embed-multilingual-v3 \
  -test-bedrock-embed-nova=amazon.nova-embed-text-v1:0

go test . -run TestBedrockLive_CohereRerank \
  -test-bedrock-region=us-east-1 \
  -test-bedrock-rerank-model=cohere.rerank-v3-5:0

Known Limitations

• Reranking is exposed as bedrock.Rerank because Genkit Go does not currently expose a first-class reranker primitive.
• Live model availability varies by AWS region, account, and model-access grants.
• Image generation config shapes differ by model family; use the family-specific Bedrock fields accepted by the target model.
• Nova embeddings are currently text-only in this plugin.
• Bedrock reasoning metadata is provider-specific and should be replayed only through the Bedrock conversation that produced it.

Troubleshooting

• No region resolved: set Bedrock.Region, AWS_REGION, AWS_DEFAULT_REGION, or a region in ~/.aws/config.
• Access denied: check AWS credentials, IAM permissions, and Bedrock model access.
• Model not found or invalid model identifier: verify the model ID, inference profile ID, account access, and region availability.
• ValidationException: check media MIME types, tool schemas, config shape, and model-specific Bedrock requirements.
• ThrottlingException: reduce concurrency, retry with backoff, or request higher Bedrock quotas.

Contributing

Use Conventional Commits for changes:

feat(models): add support for a new Bedrock model
fix(streaming): handle tool-use stream deltas
docs(readme): refresh live test commands

License

Apache 2.0. See LICENSE.

---

Built with ❤️ for the Genkit Go community