Bedrock

AWS Bedrock provider for GoAI - multi-model access via the native Converse API with SigV4 signing. No AWS SDK dependency.

Setup

go get github.com/zendev-sh/goai@latest

Set AWS credentials via environment variables, or pass them explicitly:

import "github.com/zendev-sh/goai/provider/bedrock"

model := bedrock.Chat("anthropic.claude-sonnet-4-20250514-v1:0",
    bedrock.WithAccessKey("AKIA..."),
    bedrock.WithSecretKey("..."),
    bedrock.WithRegion("us-east-1"),
)

The provider self-resolves credentials from standard AWS environment variables when not explicitly configured.

Models

Bedrock supports models from multiple providers. Use the Bedrock model ID format.

Provider	Model ID Example	Notes
Anthropic	anthropic.claude-sonnet-4-20250514-v1:0	Claude models
Amazon	amazon.nova-pro-v1:0	Nova models
Meta	meta.llama4-scout-17b-instruct-v1:0	Llama models
Mistral	mistral.mistral-large-2407-v1:0	Mistral models
DeepSeek	deepseek.deepseek-v3-0324-v1:0	DeepSeek models
AI21	ai21.jamba-1-5-mini-v1:0	Jamba models
Cohere	cohere.command-r-plus-v1:0	Command models
Others	Various	Writer, ZAI, NVIDIA, Qwen, etc.

Cross-Region Inference

Bedrock supports cross-region inference profiles. Prefix the model ID with a geographic identifier:

// US cross-region inference
model := bedrock.Chat("us.anthropic.claude-sonnet-4-6")

// EU cross-region inference
model := bedrock.Chat("eu.anthropic.claude-sonnet-4-6")

// Global cross-region inference
model := bedrock.Chat("global.anthropic.claude-sonnet-4-6")

The provider auto-detects the geographic prefix and routes to the correct regional endpoint. If a bare model ID fails, the provider automatically retries with a us. prefix.

Tested Models

E2E tested with real AWS credentials. Last run: 2026-03-15 -- 61 models tested across 15 vendors.

Vendor	Models	Generate	Stream	Status
Meta (11)	llama4-scout, llama4-maverick, llama3-3-70b, llama3-2-90b, llama3-2-11b, llama3-2-3b, llama3-2-1b, llama3-1-70b, llama3-1-8b, llama3-70b, llama3-8b	PASS	PASS	Stable
Anthropic (10)	claude-sonnet-4-6, claude-sonnet-4-5, claude-sonnet-4, claude-opus-4-6, claude-opus-4-5, claude-opus-4-1, claude-haiku-4-5, claude-3-5-sonnet, claude-3-5-haiku, claude-3-haiku	PASS	PASS	Stable
Mistral (7)	mistral-large, mixtral-8x7b, mistral-7b, ministral-3-14b, ministral-3-8b, voxtral-mini, voxtral-small	PASS	PASS	Stable
Amazon (5)	nova-micro, nova-lite, nova-pro, nova-premier, nova-2-lite	PASS	PASS	Stable
Qwen (5)	qwen3-32b, qwen3-235b, qwen3-coder-30b, qwen3-coder-480b, qwen3-next-80b	PASS	PASS	Stable
OpenAI (4)	gpt-oss-120b, gpt-oss-20b, gpt-oss-safeguard-120b, gpt-oss-safeguard-20b	PASS	PASS	Stable
Google Gemma (3)	gemma-3-4b, gemma-3-12b, gemma-3-27b	PASS	PASS	Stable
AI21 (2)	jamba-1-5-mini, jamba-1-5-large	PASS	PASS	Stable
Cohere (2)	command-r-plus, command-r	PASS	PASS	Stable
DeepSeek (2)	deepseek-v3, deepseek-r1	PASS	PASS	Stable
MiniMax (2)	minimax-m2, minimax-m2.1	PASS	PASS	Stable
Moonshot (2)	kimi-k2-thinking, kimi-k2.5	PASS	PASS	Stable
NVIDIA (2)	nemotron-nano-12b, nemotron-nano-9b	PASS	PASS	Stable
Writer (2)	palmyra-x4, palmyra-x5	PASS	PASS	Stable
ZhipuAI (2)	glm-4.7, glm-4.7-flash	PASS	PASS	Stable

Usage

Chat

import (
    "context"
    "fmt"

    "github.com/zendev-sh/goai"
    "github.com/zendev-sh/goai/provider/bedrock"
)

func main() {
    model := bedrock.Chat("anthropic.claude-sonnet-4-20250514-v1:0")
    result, err := goai.GenerateText(context.Background(), model,
        goai.WithPrompt("Explain Go interfaces in one paragraph."),
    )
    if err != nil {
        panic(err)
    }
    fmt.Println(result.Text)
}

Streaming

import (
    "context"
    "fmt"

    "github.com/zendev-sh/goai"
    "github.com/zendev-sh/goai/provider"
    "github.com/zendev-sh/goai/provider/bedrock"
)

model := bedrock.Chat("anthropic.claude-sonnet-4-20250514-v1:0")
stream, err := goai.StreamText(context.Background(), model,
    goai.WithPrompt("Write a haiku about Go."),
)
if err != nil {
    panic(err)
}
for chunk := range stream.Stream() {
    if chunk.Type == provider.ChunkText {
        fmt.Print(chunk.Text)
    }
}

Extended Thinking (Anthropic Models)

model := bedrock.Chat("anthropic.claude-sonnet-4-20250514-v1:0",
    bedrock.WithReasoningConfig(bedrock.ReasoningConfig{
        Type:         bedrock.ReasoningEnabled,
        BudgetTokens: 8192,
    }),
)
result, err := goai.GenerateText(ctx, model,
    goai.WithPrompt("Solve this step by step: What is 127 * 43?"),
)

Options

Option	Type	Description
WithRegion(r)	string	AWS region. Falls back to AWS_REGION, AWS_DEFAULT_REGION, then us-east-1.
WithAccessKey(k)	string	AWS access key ID. Falls back to AWS_ACCESS_KEY_ID.
WithSecretKey(k)	string	AWS secret access key. Falls back to AWS_SECRET_ACCESS_KEY.
WithSessionToken(t)	string	AWS session token for temporary credentials. Falls back to AWS_SESSION_TOKEN.
WithBearerToken(t)	string	Bearer token auth, bypasses SigV4. Falls back to AWS_BEARER_TOKEN_BEDROCK.
WithBaseURL(url)	string	Override endpoint URL. Falls back to AWS_BEDROCK_BASE_URL.
WithHeaders(h)	map[string]string	Additional HTTP headers.
WithHTTPClient(c)	*http.Client	Custom HTTP client.
WithReasoningConfig(cfg)	ReasoningConfig	Extended thinking configuration.
WithAdditionalModelRequestFields(f)	map[string]any	Extra fields merged into the request body.
WithAnthropicBeta(betas)	[]string	Anthropic beta features to enable.

ReasoningConfig

type ReasoningConfig struct {
    Type               ReasoningType   // "enabled", "disabled", "adaptive"
    BudgetTokens       int             // Token budget (for Type "enabled")
    MaxReasoningEffort ReasoningEffort // "low", "medium", "high", "max"
}

For Anthropic models, Type: "enabled" maps to thinking: {type: "enabled", budget_tokens: N}. For non-Anthropic models, it maps to a reasoningConfig field.

ReasoningConfig can also be passed per-request via ProviderOptions:

result, err := goai.GenerateText(ctx, model,
    goai.WithPrompt("..."),
    goai.WithProviderOptions(map[string]any{
        "reasoningConfig": map[string]any{
            "type":               "enabled",
            "budgetTokens":       4096,
            "maxReasoningEffort": "high",
        },
    }),
)

Notes

• For direct Anthropic API access without AWS, see the Anthropic provider.
• SigV4 signing: Pure Go implementation with no AWS SDK dependency. Signs requests using HMAC-SHA256 with the standard AWS4-HMAC-SHA256 algorithm.
• EventStream protocol: Streaming uses AWS EventStream binary framing (not SSE). The decoder validates CRC32 checksums on both the prelude and the full message.
• Auto-fallback chain: The provider automatically retries on several error conditions:
  1. Bare model ID fails with "invalid model identifier" or "on-demand throughput isn't supported" - retries with us. prefix
  2. us. prefix also fails - retries with the bare model ID from us-east-1
  3. maxTokens exceeds model limit - retries without maxTokens
  4. Tool calling not supported in streaming mode - retries without tools
• Anthropic tools on Bedrock: Bedrock's Converse API converts Anthropic server-side tools (web_search, web_fetch, code_execution) into standard function-call tools. The model generates tool_call requests, but these are NOT executed server-side. Provide Execute handlers to run them. Computer, bash, and text editor tools behave the same on both APIs - they always require client-side execution.
• Prompt caching: When enabled, a cachePoint block is appended to the system prompt.
• Document handling: PDF and other file attachments are sent as Bedrock document blocks. Filenames are sanitized to match AWS requirements (alphanumeric, hyphens, parentheses, brackets only; max 200 chars).
• Thinking + maxTokens: When thinking is enabled for Anthropic models, the provider automatically adds budgetTokens to maxTokens (Bedrock requires maxTokens >= budget_tokens) and removes temperature, topP, and topK from the inference config.