Merge pull request #15 from jparkerweb/develop

2.7.0

Author: jparkerweb

.example.env

@@ -9,5 +9,4 @@ AWS_SECRET_ACCESS_KEY = "xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx"
 # == LLM PARAMS ==
 # ================
 LLM_MAX_GEN_TOKENS = 800
-LLM_TEMPERATURE = 0.1
-LLM_TOP_P = 0.9
+LLM_TEMPERATURE = 0.2

CHANGELOG.md

@@ -1,8 +1,38 @@
 # Changelog
 All notable changes to this project will be documented in this file.
 
+## [2.7.0] - 2025-11-18 (DeepSeek & Qwen 3)
+### ✨ Added
+- Support for DeepSeek foundation models
+  - DeepSeek-R1 (reasoning model with chain-of-thought capabilities, 8K max output tokens)
+  - DeepSeek-V3.1 (hybrid thinking mode for complex reasoning, 8K max output tokens, **Converse API only**)
+- Support for Qwen 3 foundation models
+  - Qwen3-32B (dense architecture, 32K max output tokens)
+  - Qwen3-Coder-30B-A3B (MoE architecture for code generation, 32K max output tokens)
+  - Qwen3-235B-A22B-2507 (MoE architecture for general reasoning, 32K max output tokens)
+  - Qwen3-Coder-480B-A35B (MoE architecture for advanced software engineering, 32K max output tokens)
+- Reasoning content extraction for DeepSeek-R1 via `reasoningContent.reasoningText`
+- Stop sequences support (max 10 items) for DeepSeek and Qwen models
+- Text-to-text completion with streaming support
+- MIT-licensed open weight models for commercial use (DeepSeek)
+- `converse_api_only` flag for models that only support Converse API (automatically forces `useConverseAPI = true`)
+- Long-context handling support for Qwen 3 (up to 256K tokens natively, 1M with extrapolation)
+- Hybrid thinking modes for complex problem-solving vs. fast responses
+- Repository-scale code analysis capabilities for Qwen Coder models
+
+### 🤬 Breaking Changes
+- Removed `top_p` parameter from all models as it is not fully supported by AWS Bedrock
+  - `temperature` should always be used instead
+
+### ⚙️ Technical Details
+- **Model Configuration**: All new models use messages API format (OpenAI-compatible)
+- **API Compatibility**:
+  - Qwen 3 models: Support both Invoke API and Converse API
+  - DeepSeek-R1: Supports both Invoke API and Converse API
+  - DeepSeek-V3.1: Converse API only (automatically enforced)
+
 ## [2.6.2] - 2025-10-16 (Claude Haiku 4.5)
-### Added
+### ✨ Added
 - Support for Claude Haiku 4.5 models
   - Claude-4-5-Haiku
   - Claude-4-5-Haiku-Thinking
@@ -12,24 +42,24 @@ All notable changes to this project will be documented in this file.
 - Temperature/Top-P mutual exclusion parameter handling for Haiku 4.5 models
 
 ## [2.6.1] - 2025-09-30 (Claude Sonnet 4.5)
-### Added
+### ✨ Added
 - Support for Claude Sonnet 4.5 models
   - Claude-4-5-Sonnet
   - Claude-4-5-Sonnet-Thinking
 
 ## [2.5.0] - 2025-08-12 (Converse API)
-### Added
+### ✨ Added
 - Support for Converse API (streaming and non-streaming)
 
-### Technical Details
+### ⚙️ Technical Details
 - **Model Configuration**: All models use standard messages API format
 - **API Compatibility**: Supports OpenAI-style requests
 - **Response Processing**: Automatic reasoning tag handling based on model variant
 - **Streaming Fallback**: Automatic detection and fallback to non-streaming for unsupported models
 - **Testing Coverage**: Full integration with existing test suites and interactive example
 
 ## [2.4.5] - 2025-08-06 (GPT-OSS Models)
-### Added
+### ✨ Added
 - Support for OpenAI GPT-OSS models on AWS Bedrock
   - GPT-OSS-120B (120B parameter open weight model)
   - GPT-OSS-20B (20B parameter open weight model)
@@ -41,21 +71,21 @@ All notable changes to this project will be documented in this file.
 - Non-streaming support for GPT-OSS models (streaming not supported by AWS Bedrock)
 - OpenAI-compatible API format with `max_completion_tokens` parameter
 
-### Technical Details
+### ⚙️ Technical Details
 - **Model Configuration**: All GPT-OSS models use standard messages API format
 - **API Compatibility**: Supports OpenAI-style requests with Apache 2.0 licensed models
 - **Response Processing**: Automatic reasoning tag handling based on model variant
 - **Streaming Fallback**: Automatic detection and fallback to non-streaming for unsupported models
 - **Testing Coverage**: Full integration with existing test suites and interactive example
 
 ## [2.4.4] - 2025-08-05 (Claude 4.1 Opus)
-### Added
+### ✨ Added
 - Support for Claude 4.1 Opus models
   - Claude-4-1-Opus
   - Claude-4-1-Opus-Thinking
 
 ## [2.4.3] - 2025-07-31 (Stop Sequences Fixes)
-### Fixed
+### 🛠️ Fixed
 - **Critical Discovery**: Removed stop sequences support from Llama models
   - AWS Bedrock does not support stop sequences for Llama models (confirmed via official AWS documentation)
   - Llama models only support: `prompt`, `temperature`, `top_p`, `max_gen_len`, `images`
@@ -64,7 +94,7 @@ All notable changes to this project will be documented in this file.
   - Removed conflicting empty `inferenceConfig: {}` from Nova model configurations
 - Improved error handling for empty responses when stop sequences trigger early
 
-### Updated
+### 📝 Updated
 - **Documentation corrections**
   - Corrected stop sequences support claims (removed "all models support" language)
   - Added accurate model-specific support matrix with sequence limits
@@ -75,30 +105,30 @@ All notable changes to this project will be documented in this file.
   - ✅ Mistral models: Full support (up to 10 sequences)
   - ❌ Llama models: Not supported (AWS Bedrock limitation)
 
-### Technical Details
+### ⚙️ Technical Details
 - Based on comprehensive research of official AWS Bedrock documentation
 - All changes maintain full backward compatibility
 - Test results show significant improvements in stop sequences reliability for supported models
 - Added detailed explanations to help users understand AWS Bedrock's actual capabilities
 
 ## [2.4.2] - 2025-07-31 (Stop Sequences Support)
-### Added
+### ✨ Added
 - Stop sequences support for compatible models
   - OpenAI-compatible `stop` and `stop_sequences` parameters
   - Automatic string-to-array conversion for compatibility
   - Model-specific parameter mapping (stop_sequences for Claude, stopSequences for Nova, stop for Mistral)
 - Enhanced request building logic to include stop sequences in appropriate API formats
 - Comprehensive stop sequences testing and validation with `npm run test-stop`
 
-### Fixed
+### 🛠️ Fixed
 - **Critical Discovery**: Removed stop sequences support from Llama models
   - AWS Bedrock does not support stop sequences for Llama models (confirmed via official documentation)
   - Llama models only support: `prompt`, `temperature`, `top_p`, `max_gen_len`, `images`
   - This is an AWS Bedrock limitation, not a wrapper limitation
 - Fixed Nova model configuration conflicts that were causing stop sequence inconsistencies
 - Improved error handling for empty responses when stop sequences trigger early
 
-### Technical Details
+### ⚙️ Technical Details
 - **Model Support Matrix**:
   - ✅ Claude models: Full support (up to 8,191 sequences)
   - ✅ Nova models: Full support (up to 4 sequences)
@@ -110,7 +140,7 @@ All notable changes to this project will be documented in this file.
 - Added comprehensive documentation in README.md and CLAUDE.md explaining support limitations
 
 ## [2.4.0] - 2025-07-24 (AWS Nova Models)
-### Added
+### ✨ Added
 - Support for AWS Nova models
   - Nova-Pro (300K context, multimodal, 5K output tokens)
   - Nova-Lite (300K context, multimodal, optimized for speed)
@@ -120,15 +150,15 @@ All notable changes to this project will be documented in this file.
 - Automatic content array formatting for Nova message compatibility
 
 ## [2.3.1] - 2025-05-22 (Claude 4 Opus / Sonnet)
-### Added
+### ✨ Added
 - Support for Claude 4 Opus & Claude 4 Sonnet models
   - Claude-4-Opus
   - Claude-4-Opus-Thinking
   - Claude-4-Sonnet
   - Claude-4-Sonnet-Thinking
 
 ## [2.3.0] - 2025-02-15 (Claude 3.7 & Image Support)
-### Added
+### ✨ Added
 - Support for Claude 3.7 models
   - Claude-3-7-Sonnet
   - Claude-3-7-Sonnet-Thinking
@@ -140,20 +170,20 @@ All notable changes to this project will be documented in this file.
 - Enhanced message handling for multimodal content
 - Documentation for image support usage
 
-### Changed
+### 🔄 Changed
 - Updated model configuration for image-capable models
 - Improved response handling for multimodal inputs
 
 ## [2.2.0] - 2025-01-01 (Llama 3.3 70b)
-### Added
+### ✨ Added
 - Support for Llama 3.3 70b
 
 ## [2.1.0] - 2024-11-21 (Claude 3.5 Haiku)
-### Added
+### ✨ Added
 - Support for Claude 3.5 Haiku
 
 ## [2.0.0] - 2024-10-31 (Claude Sonnet & Haiku)
-### Added
+### ✨ Added
 - Support for Anthropic Sonnet & Haiku models
   - Claude-3-5-Sonnet-v2
   - Claude-3-5-Sonnet
@@ -163,37 +193,37 @@ All notable changes to this project will be documented in this file.
 - Stardardize output to be a string via Streamed and non-Streamed responses  
   > **NOTE:** This is a breaking change for previous non-streaming responses. Existing streaming responses will remain unchanged.
 
-### Changed
+### 🔄 Changed
 - Complete architecture overhaul for better model support
 - Improved message handling with role-based formatting
 - Enhanced error handling and response processing
 - Standardized model configuration format
 - Updated AWS SDK integration
 
-### Technical Details
+### ⚙️ Technical Details
 - Implemented messages API support for compatible models
 - Added system message handling as separate field where supported
 - Configurable token limits per model
 - Flexible response parsing with chunk/non-chunk handling
 - Cross-region profile support for certain models
 
 ## [1.3.0] - 2024-07-24 (Llama3.2)
-### Added
+### ✨ Added
 - Support for Llama 3.2 series models
   - Llama-3-2-1b
   - Llama-3-2-3b
   - Llama-3-2-11b
   - Llama-3-2-90b
 
 ## [1.1.0] - 2024-07-24 (Llama3.1)
-### Added
+### ✨ Added
 - Support for Llama 3.1 series models
   - Llama-3-1-8b
   - Llama-3-1-70b
 
 
 ## [1.0.14] - 2024-05-06 (Initial Stable Release)
-### Added
+### ✨ Added
 - Initial stablerelease of Bedrock Wrapper
 - Basic AWS Bedrock integration
 - OpenAI-compatible API object support

CLAUDE.md

@@ -4,7 +4,7 @@ This file provides guidance to Claude Code (claude.ai/code) when working with co
 
 ## Project Overview
 
-Bedrock Wrapper (v2.5.0) is an npm package that translates OpenAI-compatible API objects to AWS Bedrock's serverless inference LLMs. It supports 32+ models including Claude, Nova, GPT-OSS, Llama, and Mistral families with features like vision support, thinking modes, and stop sequences.
+Bedrock Wrapper (v2.6.2) is an npm package that translates OpenAI-compatible API objects to AWS Bedrock's serverless inference LLMs. It supports 40 models including Claude, Nova, GPT-OSS, Llama, Mistral, and Qwen families with features like vision support, thinking modes, and stop sequences.
 
 ## Core Architecture
 
@@ -82,8 +82,7 @@ AWS_REGION=us-west-2
 AWS_ACCESS_KEY_ID=your_access_key
 AWS_SECRET_ACCESS_KEY=your_secret_key
 LLM_MAX_GEN_TOKENS=1024
-LLM_TEMPERATURE=0.1
-LLM_TOP_P=0.9
+LLM_TEMPERATURE=0.2
 ```
 
 ## Adding New Models
@@ -92,19 +91,25 @@ Required fields in bedrock-models.js:
 - `modelName`: Consumer-facing name
 - `modelId`: AWS Bedrock identifier
 - `vision`: Boolean for image support
-- `messages_api`: Boolean (true for Claude/Nova/GPT-OSS, false for prompt-based)
+- `messages_api`: Boolean (true for Claude/Nova/GPT-OSS/Qwen, false for prompt-based)
 - `response_chunk_element`: JSON path for streaming responses
 - `response_nonchunk_element`: JSON path for non-streaming responses
 - `special_request_schema`: Model-specific requirements
 - `stop_sequences_param_name`: Parameter name for stop sequences
 
 ## Critical Implementation Details
 
+### Converse API Only Models
+Some models only support the Converse API and will automatically use it regardless of the `useConverseAPI` flag:
+- DeepSeek-V3.1
+
+These models have `converse_api_only: true` in their configuration and the wrapper automatically forces `useConverseAPI = true` for them.
+
 ### Converse API Thinking Support
 - Thinking configuration added via `additionalModelRequestFields`
 - Response thinking data extracted from `reasoningContent.reasoningText.text`
 - Budget tokens calculated with constraints: 1024 <= budget_tokens <= (maxTokens * 0.8)
-- Temperature forced to 1.0, top_p removed for thinking models
+- Temperature forced to 1.0 for thinking models
 
 ### Nova Models Special Handling
 - Detect via `special_request_schema.schemaVersion === "messages-v1"`
@@ -124,6 +129,7 @@ Required fields in bedrock-models.js:
 | Nova        | ✅      | stopSequences | 4 |
 | GPT-OSS     | ✅      | stop_sequences | TBD |
 | Mistral     | ✅      | stop | 10 |
+| Qwen        | ✅      | stop | TBD |
 | Llama       | ❌      | N/A | N/A |
 
 ### Test Files Output

README.md

@@ -43,7 +43,6 @@ Bedrock Wrapper is an npm package that simplifies the integration of existing Op
         "max_tokens": LLM_MAX_GEN_TOKENS,
         "stream": true,
         "temperature": LLM_TEMPERATURE,
-        "top_p": LLM_TOP_P,
         "stop_sequences": ["STOP", "END"], // Optional: sequences that will stop generation
     };
     ```
@@ -158,7 +157,11 @@ Bedrock Wrapper is an npm package that simplifies the integration of existing Op
 | Mistral-7b                 | mistral.mistral-7b-instruct-v0:2             |  ❌  |
 | Mixtral-8x7b               | mistral.mixtral-8x7b-instruct-v0:1           |  ❌  |
 | Mistral-Large              | mistral.mistral-large-2402-v1:0              |  ❌  |
-   
+| Qwen3-32B                  | alibaba.qwen3-32b-instruct-v1:0              |  ❌  |
+| Qwen3-Coder-30B-A3B        | alibaba.qwen3-coder-30b-a3b-instruct-v1:0    |  ❌  |
+| Qwen3-235B-A22B-2507       | alibaba.qwen3-235b-a22b-instruct-2507-v1:0   |  ❌  |
+| Qwen3-Coder-480B-A35B      | alibaba.qwen3-coder-480b-a35b-instruct-v1:0  |  ❌  |
+
 To return the list progrmatically you can import and call `listBedrockWrapperSupportedModels`:  
 ```javascript
 import { listBedrockWrapperSupportedModels } from 'bedrock-wrapper';
@@ -235,9 +238,10 @@ const openaiChatCompletionsCreateObject = {
 
 **Model Support:**
 - ✅ **Claude models**: Fully supported (up to 8,191 sequences)
-- ✅ **Nova models**: Fully supported (up to 4 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
 - ❌ **Llama models**: Not supported (AWS Bedrock limitation)
 
 **Features:**
@@ -251,7 +255,7 @@ const openaiChatCompletionsCreateObject = {
 // 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, or Mistral models
+    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")
@@ -274,27 +278,6 @@ Some AWS Bedrock models have specific parameter restrictions that are automatica
 - Claude-4-Opus & Claude-4-Opus-Thinking
 - Claude-4-1-Opus & Claude-4-1-Opus-Thinking
 
-**Restriction:** These models cannot accept both `temperature` and `top_p` parameters simultaneously.
-
-**Automatic Handling:** When both parameters are provided, the wrapper automatically:
-1. **Keeps `temperature`** (prioritized as more commonly used)
-2. **Removes `top_p`** to prevent validation errors
-3. **Works with both APIs** (Invoke API and Converse API)
-
-```javascript
-const request = {
-    messages: [{ role: "user", content: "Hello" }],
-    model: "Claude-4-5-Sonnet",
-    temperature: 0.7,  // ✅ Kept
-    top_p: 0.9         // ❌ Automatically removed
-};
-
-// No error thrown - wrapper handles the restriction automatically
-const response = await bedrockWrapper(awsCreds, request);
-```
-
-**Why This Happens:** AWS Bedrock enforces this restriction on newer Claude models to ensure optimal performance and prevent conflicting sampling parameters.
-
 ---
 
 ### 🧪 Testing