v0.1.14-update.2

Highlights

• Automatic Prompt Caching: One-line setup for Claude 4.x models — 40-70% cost reduction on cached tokens with zero message modifications
• Extended Thinking Support: Full reasoning capabilities for Claude, DeepSeek R1, OpenAI GPT OSS, and Moonshot models with signature preservation
• Ollama Cloud Support: Run 120B+ models without local GPU—simple API key configuration enables cloud deployment
• 15+ New Models: Claude 4.6, DeepSeek R1/V3.2, OpenAI GPT OSS, Moonshot Kimi, Qwen3-Next, Mistral Large3, GLM-4.7
• Production-Ready: Comprehensive documentation, flexible AWS authentication, 2,800+ lines of test coverage

Major Features & Improvements

Automatic Prompt Caching

Enable with single option—no message modifications needed:

llm, err := bedrock.New(
    bedrock.WithClient(client),
    bedrock.WithAutomaticCaching(),
)

• 40-70% cost reduction on cached tokens (charged at 10% rate)
• Works with both Legacy and Converse APIs
• Automatically caches conversation history before new user input
• Supports 5-minute (default) and 1-hour cache durations
• Real-world: 1,400+ tokens cached per turn in multi-turn agent conversations

Supported: All Claude 4.x models (Opus, Sonnet, Haiku)

Extended Thinking (Reasoning)

Full reasoning support with signature preservation across multi-turn conversations:

• Models: Claude 4.6/4.5/4.1/4.0, DeepSeek R1, OpenAI GPT OSS (120B/20B), Moonshot Kimi K2-Thinking
• Use Cases: Transparent decision-making for regulated industries, debugging agent behavior, higher quality responses
• Features: Streaming reasoning, tool call integration, automatic signature preservation
• Works with both Legacy API (InvokeModel) and Converse API

Ollama Cloud Support

Run powerful models without local GPU requirements—simply add an API key:

llm, err := ollama.New(
    ollama.WithServerURL(ollama.CloudURL),
    ollama.WithAPIKey(os.Getenv("OLLAMA_API_KEY")),
    ollama.WithModel("gpt-oss:120b"),
)

• Zero Infrastructure: No GPU, no model downloads, no server management
• Seamless Integration: Same API for local and cloud deployments—change one line to switch
• Backward Compatible: Existing local Ollama code works without modifications
• Smart Authentication: API key automatically applied from OLLAMA_API_KEY environment variable
• Production Features: Tool calling, JSON mode, streaming, reasoning support on cloud models

Implementation Highlights:

• Proxy authRoundTripper pattern preserves custom HTTP client settings (timeouts, TLS, proxies)
• Automatic credential detection—reads OLLAMA_API_KEY from environment
• Comprehensive test suite with httprr for deterministic testing without cloud credentials

Use Cases:

• Run 120B models on laptops without GPU
• Deploy AI features without infrastructure investment
• Quick prototyping with large models before optimizing for local deployment
• Hybrid approach: development on cloud, production on local servers

New Models

Added 15+ models with complete capability documentation:

• Claude 4.6 (Opus, Sonnet) - enhanced reasoning, automatic caching
• DeepSeek R1 - reasoning specialist, V3.2 - production model (164K context)
• OpenAI GPT OSS (120B/20B) - open-source reasoning models
• Moonshot Kimi - K2-Thinking (reasoning), K2.5 (multimodal)
• Qwen3 - Coder-Next, Next-80B (256K context), VL-235B (vision), 32B, Coder-30B
• GLM-4.7 / Flash - Z.AI models for front-end development
• Mistral - Large3, Large2402 (tool calling), Magistral Small

Documentation includes capability matrix showing which features work with each model

Other Improvements

AWS Authentication

• Support for long-lived credentials (access key/secret key) and bearer token authentication
• Enables deployment across EC2, ECS, temporary credentials, cross-account access scenarios

Documentation

• AWS Bedrock: Architecture guide with three-layer design and mermaid diagrams, capability matrix, integration guide for new models
• Ollama: Complete architecture guide covering local/cloud deployment modes, authentication patterns, testing strategies
• Debugging guides, performance considerations, and advanced usage examples for all major providers

Bug Fixes

• Fixed HTTP record/replay for AWS Signature V4 authentication
• Enhanced test infrastructure for both credential types

Provider Updates

Ollama

• Updated to v0.17.5: Latest Ollama SDK with cloud model support, improved API compatibility, and new features
• Cloud Deployment: Native support for Ollama Cloud—run large models without local GPU infrastructure
• Smart Authentication: Automatic API key detection from OLLAMA_API_KEY environment variable with Bearer token injection
• Proxy Pattern: Custom authRoundTripper preserves user-provided HTTP client settings while adding authentication headers
• Comprehensive Testing: Integration tests for both local and cloud deployments with httprr support
• Documentation: Complete architecture guide covering dual-mode deployment, authentication, and advanced features

Migration Path:

// Local deployment (existing code - no changes needed)
llm, _ := ollama.New(ollama.WithModel("llama3.2"))

// Cloud deployment (new capability - one-line change)
llm, _ := ollama.New(
    ollama.WithServerURL(ollama.CloudURL),
    ollama.WithModel("gpt-oss:120b"),
    // API key auto-detected from OLLAMA_API_KEY env var
)

Benefits:

• Access 120B+ models instantly without downloading or GPU hardware
• Preserve existing local deployment workflows—zero breaking changes
• Unified API across local/cloud modes—same code, different infrastructure
• Custom HTTP client support maintained (timeouts, proxies, TLS configuration)

Dependencies Updates

• Updated github.com/ollama/ollama to v0.17.5 with cloud model support and latest API features
• All dependencies verified for compatibility with Go 1.24.1

Technical Details

• Architecture: Unified interface for Legacy/Converse APIs, intelligent caching engine, extended provider detection (Qwen, Moonshot, Z.AI, OpenAI, Mistral, DeepSeek), dual-mode Ollama transport
• Testing: 2,800+ lines of integration tests with httprr recordings—no AWS/cloud credentials needed for CI
• Quality: Zero linter errors, fully backward compatible, comprehensive documentation

Contributors

This release incorporates improvements and testing from the broader langchaingo community:

• @asdek - Extended thinking implementation, automatic caching, Ollama Cloud support, comprehensive documentation, expanded model support, credential handling
• AWS Bedrock community - Feedback and testing across diverse deployment scenarios
• Ollama community - Testing and validation of cloud deployment features

Breaking Changes

None—this release is fully backward compatible. All new features are opt-in.

Full Changelog: v0.1.14-update.1...v0.1.14-update.2