🪨 Bedrock Wrapper

Bedrock Wrapper is an npm package that simplifies the integration of existing OpenAI-compatible API objects with AWS Bedrock's serverless inference LLMs. Follow the steps below to integrate into your own application, or alternativly use the 🔀 Bedrock Proxy Endpoint project to spin up your own custom OpenAI server endpoint for even easier inference (using the standard baseUrl, and apiKey params).

---

Maintained by

---

Install

• install package: npm install bedrock-wrapper

---

Usage

1. import bedrockWrapper

   import { bedrockWrapper } from "bedrock-wrapper";

2. create an awsCreds object and fill in your AWS credentials

   const awsCreds = {
       region: AWS_REGION,
       accessKeyId: AWS_ACCESS_KEY_ID,
       secretAccessKey: AWS_SECRET_ACCESS_KEY,
   };

3. clone your openai chat completions object into openaiChatCompletionsCreateObject or create a new one and edit the values

   const openaiChatCompletionsCreateObject = {
       "messages": messages,
       "model": "Claude-4-5-Sonnet",
       "max_tokens": LLM_MAX_GEN_TOKENS,
       "stream": true,
       "temperature": LLM_TEMPERATURE,
       "stop_sequences": ["STOP", "END"], // Optional: sequences that will stop generation
   };

   the messages variable should be in openai's role/content format (not all models support system prompts)

   messages = [
       {
           role: "system",
           content: "You are a helpful AI assistant that follows instructions extremely well. Answer the user questions accurately. Think step by step before answering the question. You will get a $100 tip if you provide the correct answer.",
       },
       {
           role: "user",
           content: "Describe why openai api standard used by lots of serverless LLM api providers is better than aws bedrock invoke api offered by aws bedrock. Limit your response to five sentences.",
       },
       {
           role: "assistant",
           content: "",
       },
   ]

   the model value should be the corresponding modelName value in the bedrock_models section below (see Supported Models below)

4. call the bedrockWrapper function and pass in the previously defined awsCreds and openaiChatCompletionsCreateObject objects

   // create a variable to hold the complete response
   let completeResponse = "";
   // invoke the streamed bedrock api response
   for await (const chunk of bedrockWrapper(awsCreds, openaiChatCompletionsCreateObject)) {
       completeResponse += chunk;
       // ---------------------------------------------------
       // -- each chunk is streamed as it is received here --
       // ---------------------------------------------------
       process.stdout.write(chunk); // ⇠ do stuff with the streamed chunk
   }
   // console.log(`\n\completeResponse:\n${completeResponse}\n`); // ⇠ optional do stuff with the complete response returned from the API reguardless of stream or not

   if calling the unstreamed version you can call bedrockWrapper like this

   // create a variable to hold the complete response
   let completeResponse = "";
   if (!openaiChatCompletionsCreateObject.stream){ // invoke the unstreamed bedrock api response
       const response = await bedrockWrapper(awsCreds, openaiChatCompletionsCreateObject);
       for await (const data of response) {
           completeResponse += data;
       }
       // ----------------------------------------------------
       // -- unstreamed complete response is available here --
       // ----------------------------------------------------
       console.log(`\n\completeResponse:\n${completeResponse}\n`); // ⇠ do stuff with the complete response
   }

5. NEW: Using the Converse API (optional)

   You can now optionally use AWS Bedrock's Converse API instead of the Invoke API by passing useConverseAPI: true in the options parameter:

   // Use the Converse API for unified request/response format across all models
   for await (const chunk of bedrockWrapper(awsCreds, openaiChatCompletionsCreateObject, { useConverseAPI: true })) {
       completeResponse += chunk;
       process.stdout.write(chunk);
   }

   The Converse API provides:

   • Consistent API: Single request/response format across all models
   • Simplified conversation management: Better handling of multi-turn conversations
   • System prompts: Cleaner separation of system instructions
   • Tool use support: Native support for function calling (where supported)
   • Unified multimodal: Consistent handling of text and image inputs

   Note: Some models only support the Converse API and will automatically use it regardless of the useConverseAPI flag:

   • DeepSeek-V3.1

---

Supported Models

modelName	AWS Model Id	Image
Claude-3-5-Haiku	us.anthropic.claude-3-5-haiku-20241022-v1:0	❌
Claude-3-5-Sonnet	us.anthropic.claude-3-5-sonnet-20240620-v1:0	✅
Claude-3-5-Sonnet-v2	us.anthropic.claude-3-5-sonnet-20241022-v2:0	✅
Claude-3-7-Sonnet	us.anthropic.claude-3-7-sonnet-20250219-v1:0	✅
Claude-3-7-Sonnet-Thinking	us.anthropic.claude-3-7-sonnet-20250219-v1:0	✅
Claude-3-Haiku	us.anthropic.claude-3-haiku-20240307-v1:0	✅
Claude-4-Opus	us.anthropic.claude-opus-4-20250514-v1:0	✅
Claude-4-Opus-Thinking	us.anthropic.claude-opus-4-20250514-v1:0	✅
Claude-4-Sonnet	us.anthropic.claude-sonnet-4-20250514-v1:0	✅
Claude-4-Sonnet-Thinking	us.anthropic.claude-sonnet-4-20250514-v1:0	✅
Claude-4-1-Opus	us.anthropic.claude-opus-4-1-20250805-v1:0	✅
Claude-4-1-Opus-Thinking	us.anthropic.claude-opus-4-1-20250805-v1:0	✅
Claude-4-5-Haiku	global.anthropic.claude-haiku-4-5-20251001-v1:0	✅
Claude-4-5-Haiku-Thinking	global.anthropic.claude-haiku-4-5-20251001-v1:0	✅
Claude-4-5-Opus	global.anthropic.claude-opus-4-5-20251101-v1:0	✅
Claude-4-5-Opus-Thinking	global.anthropic.claude-opus-4-5-20251101-v1:0	✅
Claude-4-5-Sonnet	us.anthropic.claude-sonnet-4-5-20250929-v1:0	✅
Claude-4-5-Sonnet-Thinking	us.anthropic.claude-sonnet-4-5-20250929-v1:0	✅
DeepSeek-R1	us.deepseek.r1-v1:0	❌
DeepSeek-V3.1	deepseek.v3-v1:0	❌
Gemma-3-4b	google.gemma-3-4b-it	✅
Gemma-3-12b	google.gemma-3-12b-it	✅
Gemma-3-27b	google.gemma-3-27b-it	✅
GPT-OSS-120B	openai.gpt-oss-120b-1:0	❌
GPT-OSS-120B-Thinking	openai.gpt-oss-120b-1:0	❌
GPT-OSS-20B	openai.gpt-oss-20b-1:0	❌
GPT-OSS-20B-Thinking	openai.gpt-oss-20b-1:0	❌
Kimi-K2	moonshot.kimi-k2-thinking	❌
Kimi-K2-Thinking	moonshot.kimi-k2-thinking	❌
Llama-3-8b	meta.llama3-8b-instruct-v1:0	❌
Llama-3-70b	meta.llama3-70b-instruct-v1:0	❌
Llama-3-1-8b	us.meta.llama3-1-8b-instruct-v1:0	❌
Llama-3-1-70b	us.meta.llama3-1-70b-instruct-v1:0	❌
Llama-3-1-405b	meta.llama3-1-405b-instruct-v1:0	❌
Llama-3-2-1b	us.meta.llama3-2-1b-instruct-v1:0	❌
Llama-3-2-3b	us.meta.llama3-2-3b-instruct-v1:0	❌
Llama-3-2-11b	us.meta.llama3-2-11b-instruct-v1:0	❌
Llama-3-2-90b	us.meta.llama3-2-90b-instruct-v1:0	❌
Llama-3-3-70b	us.meta.llama3-3-70b-instruct-v1:0	❌
Magistral-Small-2509	mistral.magistral-small-2509	❌
MiniMax-M2	minimax.minimax-m2	❌
Ministral-3-3b	mistral.ministral-3-3b-instruct	✅
Ministral-3-8b	mistral.ministral-3-8b-instruct	✅
Ministral-3-14b	mistral.ministral-3-14b-instruct	✅
Mistral-7b	mistral.mistral-7b-instruct-v0:2	❌
Mistral-Large	mistral.mistral-large-2402-v1:0	❌
Mistral-Large-3	mistral.mistral-large-3-675b-instruct	✅
Mixtral-8x7b	mistral.mixtral-8x7b-instruct-v0:1	❌
Nova-2-Lite	us.amazon.nova-2-lite-v1:0	✅
Nova-Micro	us.amazon.nova-micro-v1:0	❌
Nova-Lite	us.amazon.nova-lite-v1:0	✅
Nova-Pro	us.amazon.nova-pro-v1:0	✅
Qwen3-32B	qwen.qwen3-32b-v1:0	❌
Qwen3-235B-A22B-2507	qwen.qwen3-235b-a22b-2507-v1:0	❌
Qwen3-Coder-30B-A3B	qwen.qwen3-coder-30b-a3b-v1:0	❌
Qwen3-Coder-480B-A35B	qwen.qwen3-coder-480b-a35b-v1:0	❌
Qwen3-Next-80B-A3B	qwen.qwen3-next-80b-a3b	❌

To return the list progrmatically you can import and call listBedrockWrapperSupportedModels:

import { listBedrockWrapperSupportedModels } from 'bedrock-wrapper';
console.log(`\nsupported models:\n${JSON.stringify(await listBedrockWrapperSupportedModels())}\n`);

Additional Bedrock model support can be added.
Please modify the bedrock_models.js file and submit a PR 🏆 or create an Issue.

---

Thinking Models

Some models support extended reasoning capabilities through "thinking mode". These models include:

• Claude models: Claude-4-5-Opus-Thinking, Claude-4-1-Opus-Thinking, Claude-4-Opus-Thinking, Claude-4-5-Sonnet-Thinking, Claude-4-5-Haiku-Thinking, Claude-4-Sonnet-Thinking, Claude-3-7-Sonnet-Thinking
• GPT-OSS models: GPT-OSS-120B-Thinking, GPT-OSS-20B-Thinking
• Kimi models: Kimi-K2-Thinking (preserves reasoning tags in output)

To use thinking mode and see the model's reasoning process, set include_thinking_data: true in your request:

const openaiChatCompletionsCreateObject = {
    "messages": messages,
    "model": "Claude-4-5-Sonnet-Thinking",
    "max_tokens": 4000,
    "stream": true,
    "temperature": 1.0, // Thinking models require temperature of 1.0
    "include_thinking_data": true // Enable thinking output
};

let completeResponse = "";
for await (const chunk of bedrockWrapper(awsCreds, openaiChatCompletionsCreateObject)) {
    completeResponse += chunk;
    process.stdout.write(chunk); // Shows both thinking and response
}

Features:

• Thinking content appears in <think>...</think> tags for Claude models
• Thinking content appears in <reasoning>...</reasoning> tags for GPT-OSS models
• Temperature is automatically set to 1.0 for optimal thinking performance
• Budget tokens are automatically calculated based on max_tokens

---

Image Support

For models with image support (Claude 4+ series including Claude 4.5 Opus, Claude 4.5 Sonnet, Claude 4.5 Haiku, Claude 3.7 Sonnet, Claude 3.5 Sonnet, Claude 3 Haiku, Nova Pro, Nova Lite, Nova 2 Lite, Mistral Large 3, Ministral 3 series, and Gemma 3 series), you can include images in your messages using the following format (not all models support system prompts):

messages = [
    {
        role: "system",
        content: "You are a helpful AI assistant that can analyze images.",
    },
    {
        role: "user",
        content: [
            { type: "text", text: "What's in this image?" },
            { 
                type: "image_url", 
                image_url: {
                    url: "data:image/jpeg;base64,/9j/4AAQSkZJRgABAQEA..." // base64 encoded image
                }
            }
        ]
    }
]

You can also use a direct URL to an image instead of base64 encoding:

messages = [
    {
        role: "user",
        content: [
            { type: "text", text: "Describe this image in detail." },
            { 
                type: "image_url", 
                image_url: {
                    url: "https://example.com/path/to/image.jpg" // direct URL to image
                }
            }
        ]
    }
]

You can include multiple images in a single message by adding more image_url objects to the content array.

---

Stop Sequences

Stop sequences are custom text sequences that cause the model to stop generating text. This is useful for controlling where the model stops its response.

const openaiChatCompletionsCreateObject = {
    "messages": messages,
    "model": "Claude-3-5-Sonnet",
    "max_tokens": 100,
    "stop_sequences": ["STOP", "END", "\n\n"], // Array of stop sequences
    // OR use single string format:
    // "stop": "STOP"
};

Model Support:

• ✅ Claude models: Fully supported (up to 8,191 sequences)
• ✅ Nova models: Fully supported (up to 4 sequences)
• ✅ GPT-OSS models: Fully supported
• ✅ Mistral models: Fully supported (up to 10 sequences)
• ✅ Qwen models: Fully supported
• ✅ Gemma models: Fully supported
• ✅ Kimi models: Fully supported
• ✅ MiniMax models: Fully supported
• ❌ Llama models: Not supported (AWS Bedrock limitation)

Features:

• Compatible with OpenAI's stop parameter (single string or array)
• Also accepts stop_sequences parameter for explicit usage
• Automatic conversion between string and array formats
• Model-specific parameter mapping handled automatically

Example Usage:

// Stop generation when model tries to output "7"
const result = await bedrockWrapper(awsCreds, {
    messages: [{ role: "user", content: "Count from 1 to 10" }],
    model: "Claude-3-5-Sonnet",  // Use Claude, Nova, Mistral, or Qwen models
    stop_sequences: ["7"]
});
// Response: "1, 2, 3, 4, 5, 6," (stops before "7")

// Note: Llama models will ignore stop sequences due to AWS Bedrock limitations

---

Parameter Restrictions

Some AWS Bedrock models have specific parameter restrictions that are automatically handled by the wrapper:

Claude 4+ Models (Temperature/Top-P Mutual Exclusion)

Affected Models:

• Claude-4-5-Opus & Claude-4-5-Opus-Thinking
• Claude-4-5-Sonnet & Claude-4-5-Sonnet-Thinking
• Claude-4-5-Haiku & Claude-4-5-Haiku-Thinking
• Claude-4-Sonnet & Claude-4-Sonnet-Thinking
• Claude-4-Opus & Claude-4-Opus-Thinking
• Claude-4-1-Opus & Claude-4-1-Opus-Thinking

---

🧪 Testing

The package includes comprehensive test suites to verify functionality:

# Test all models with the Both APIs (Comparison)
npm run test

# Test all models with the Invoke API
npm run test:invoke

# Test all models with the Converse API
npm run test:converse

# Test vision/multimodal capabilities with Both APIs (Comparison)
npm run test-vision

# Test vision/multimodal capabilities with Invoke API
npm run test-vision:invoke

# Test vision/multimodal capabilities with Converse API
npm run test-vision:converse

# Test stop sequences functionality with Both APIs (Comparison)
npm run test-stop

# Test stop sequences functionality with Invoke API
npm run test-stop:invoke

# Test stop sequences functionality with Converse API
npm run test-stop:converse

# Test Converse API specifically
npm run test-converse

# Run all test suites
npm run test:all

# Interactive testing
npm run interactive

---

📢 P.S.

In case you missed it at the beginning of this doc, for an even easier setup, use the 🔀 Bedrock Proxy Endpoint project to spin up your own custom OpenAI server endpoint (using the standard baseUrl, and apiKey params).

---

📚 References

• AWS Meta Llama Models User Guide
• AWS Mistral Models User Guide
• OpenAI API
• AWS Bedrock
• AWS SDK for JavaScript

---

Please consider sending me a tip to support my work 😀

🍵 tip me here