Releases · vxcontrol/langchaingo

23 Jun 19:41

asdek

v0.1.13-update.5

0e58efc

v0.1.13-update.5 Latest

Latest

Highlights

HTTP Record and Replay Testing: Added httprr package for deterministic HTTP testing that dramatically speeds up test execution
Enhanced Documentation: Completely restructured documentation site with improved search, architecture guides, and tutorials
Comprehensive Test Coverage: Added extensive test coverage across all components using the new httprr system

Breaking Changes

Google AI Model Update: Deprecated textembedding-gecko model replaced with text-embedding-005
Test Context Usage: Standardized context usage in tests from t.Context() to context.Background()

Major Features & Improvements

HTTP Record and Replay Testing (httprr)

Added comprehensive httprr package for deterministic HTTP testing by @tmc in #1296
Enables recording real HTTP interactions during development and replaying during CI/testing
Built-in scrubbing automatically removes API keys, tokens, and sensitive headers
Supports compressed .httprr.gz files for repository storage efficiency
Test flags: -httprecord=<regexp> for re-recording and -httpdebug for HTTP logging
Supports env variables HTTPRR_RECORD, HTTPRR_DEBUG, HTTPRR_HTTPDEBUG, HTTPRR_RECORD_DELAY in tests
Dramatically improves test execution speed and reliability
Enables analysis of LLM response formats through recorded .httprr files

Documentation Overhaul

Complete documentation site restructure by @tmc in #1311
Added comprehensive architecture documentation and how-to guides
New tutorials section with basic chat application examples
Enhanced search functionality with Go-based indexing by @tmc in #1313
Switched from npm to pnpm for package management
Added Vale linting configuration for documentation consistency
Fixed documentation typos and improved API consistency by @BowieHe in #1314

CI and Development Infrastructure

Enhanced CI workflows with matrix testing and coverage reporting by @tmc in #1288
Added Git hooks for pre-push quality checks
Separated example builds into dedicated workflows
Added race condition testing and automated PR coverage comments
Added comprehensive linting for test patterns and architectural rules by @tmc in #1312

Provider Updates

Huggingface

Updated embeddings to use /models/ endpoint instead of deprecated /pipeline/ by @tmc in #1299
Added inference provider support for router-based requests
Improved token configuration with HF_TOKEN as primary environment variable
Added support for HF_TOKEN_PATH, HF_HOME, and XDG_CACHE_HOME
Updated default embedding model to BAAI/bge-small-en-v1.5

Google AI

Replaced deprecated textembedding-gecko with text-embedding-005 by @alanrichman in #1292
Updated to latest text embedding model version for improved performance

Azure AI Search

Added WithEndpoint option for endpoint configuration by @thecoons in #1308
Provides alternative to environment variable AZURE_AI_SEARCH_ENDPOINT

OpenAI

Fixed tool call handling in OpenAI functions agent by @tmc in #1288
Addressed ID passing issue in tool calls
Added support for both legacy and new tool calling APIs

Testing Infrastructure Improvements

Added extensive httprr test coverage across all components by @tmc in #1299 and #1312
Updated all HTTP-based tests to use httprr for consistent test execution
Reorganized transport layer with structured logging via slog
Added testutil/testctr helper to work around testcontainers-go+colima bug
Enhanced test execution speed and reliability

Dependencies Updates

Restructured and organized go.mod dependency declarations
Updated examples with clean module declarations
Updated testcontainers to new API where supported
Switched from npm to pnpm for documentation package management

Bug Fixes

Fixed OpenAI functions agent tool call ID passing issue by @tmc in #1288
Removed redundant error checks in LLM chain tests by @tmc in #1299
Fixed various linting issues across the codebase
Fixed deprecated Google AI embedding model usage

Contributors

A huge thank you to all the contributors who made this release possible:

@tmc - HTTP record/replay system, documentation overhaul, comprehensive test coverage
@alanrichman - Google AI embedding model update
@thecoons - Azure AI Search endpoint configuration
@BowieHe - Documentation fixes and API consistency
@jacobcode91 - Documentation contributions

Full Changelog: v0.1.13-update.4...v0.1.13-update.5

Contributors

tmc, thecoons, and 3 other contributors

Assets 2

28 May 23:51

asdek

v0.1.13-update.4

c4bffe6

v0.1.13-update.4

Highlights

Upstream Sync: Updated to the latest upstream version with critical bug fixes and new features
Google AI Tool Handling: Fixed issue where Gemini models would ignore all but the first tool call
AWS Bedrock Knowledge Bases: Added comprehensive vector store support for AWS Bedrock Knowledge Bases
Anthropic Reasoning Enhancement: Added reasoning content support for function calling choices to maintain consistency with Claude Sonnet 4

Major Improvements

Upstream Integration

This release incorporates the latest changes from the upstream repository, bringing several important fixes and enhancements.

Google AI Provider Enhancements

Fixed Multi-Tool Processing: Google AI provider now correctly handles multiple tool calls in a single request
Improved Tool Consolidation: Changed from creating multiple tools with single function declarations to a single tool with multiple function declarations
Better API Compatibility: Aligned tool handling with Google AI API expectations for more reliable function calling

AWS Bedrock Knowledge Bases Vector Store

Complete Vector Store Implementation: Added full support for AWS Bedrock Knowledge Bases as a vector store backend
Document Ingestion: Comprehensive document ingestion capabilities with support for various file formats
S3 Integration: Native S3 bucket integration for document storage and retrieval
Metadata Filtering: Advanced filtering capabilities for precise document retrieval

Anthropic Provider Improvements

Reasoning Content for Function Calls: Added reasoning content support to function calling choices
Claude Sonnet 4 Compatibility: Enhanced compatibility with the latest Claude Sonnet 4 model behavior
Consistent API Behavior: Synchronized reasoning content handling with OpenAI provider for consistent cross-provider experience

Borrowed Upstream Pull Requests

This release incorporates code and improvements from the following upstream pull requests:

Google AI Multi-Tool Fix from PR #1244 by @jpoz - Fixed issue where Gemini would ignore all but the first tool
AWS Bedrock Knowledge Bases from PR #1200 by @lucaronca - Added comprehensive vector store support
Additional Enhancements from PR #1135 by @lucaronca - Various improvements and bug fixes
Provider Updates from PR #1260 by @laurence-nearform and @lmorg - Enhanced provider capabilities

Contributors

@jpoz - Google AI multi-tool fix (upstream PR tmc#1244)
@lucaronca - AWS Bedrock Knowledge Bases implementation (upstream PRs tmc#1200, tmc#1135)
@laurence-nearform - Provider enhancements (upstream PR tmc#1260)
@lmorg - Provider enhancements (upstream PR tmc#1260)
@asdek - Anthropic reasoning enhancements

Full Changelog: v0.1.13-update.3...v0.1.13-update.4

Contributors

jpoz, asdek, and 3 other contributors

Assets 2

13 May 18:42

asdek

v0.1.13-update.3

ae63c5a

v0.1.13-update.3

Highlights

Better Reasoning: Improved step-by-step thinking support in AI models, including models through OpenRouter
Stream Done Signal: Streaming API now clearly tells you when a response is finished

Major Improvements

Reasoning Capabilities

Access to AI Thinking Process: Access the reasoning content (step-by-step thinking) from models that support this feature
Cross-Provider Support: Works with multiple AI providers including OpenAI, Anthropic, and models available through OpenRouter
Flexible Configuration: Use either a simple effort level or specific token budget to control reasoning behavior
Multiple Format Support: Compatible with both modern and legacy reasoning formats across different providers

Streaming Enhancements

Reliable Resource Management: New ChunkTypeDone notification ensures your application knows exactly when a stream is complete, allowing for predictable resource release
Simplified Error Handling: No more need for timeout workarounds or complex stream completion detection logic

Technical Details

Added ReasoningOptions type with Effort and MaxTokens parameters
Extended necessary structures to support reasoning fields reasoning_content and reasoning
Implemented ChunkTypeDone for reliable stream completion notification in all LLM implementations

Contributors

@asdek

Full Changelog: v0.1.13-update.2...v0.1.13-update.3

Contributors

asdek

Assets 2

12 May 12:48

asdek

v0.1.13-update.2

18a7017

v0.1.13-update.2

Highlights

This release introduces significant enhancements to the streaming capabilities, reasoning features, and tool calling functionality across multiple LLM providers. It focuses on improving both API consistency and the robustness of interactions with various AI models.

Breaking Changes

Streaming API Interface: Changed the signature of StreamingFunc from func(ctx context.Context, chunk []byte) error to func(ctx context.Context, chunk streaming.Chunk) error
Removed StreamingReasoningFunc: This function has been consolidated into the new streaming architecture
Reasoning Configuration: Added explicit reasoning configuration options that may affect existing code using reasoning capabilities
Tool Call IDs: Changed tool call ID format for better cross-provider compatibility

Major Improvements

New Packages

Streaming Package: Added a dedicated package for handling streaming responses from language models
- Unified interface for processing text, reasoning, and tool call chunks
- Type-safe chunk structures with proper content type discrimination
- Helper functions for properly invoking streaming callbacks by chunk type
- Support for handling tool call data in streaming mode with proper state tracking
Reasoning Package: Added support for extracting and processing step-by-step thinking from models
- Pattern-matching extraction for <think> and <thinking> tags in model responses
- Stateful ChunkContentSplitter for handling streaming reasoning content
- Support for both incremental streaming and complete content processing
- Integration with all major LLM providers (OpenAI, Anthropic, Ollama)

LLM Client Enhancements

Ollama Client

Enhanced streaming support with proper chunk processing for text and tool calls
Added reasoning content extraction for models with thinking capabilities
Improved tool call handling with robust ID generation and argument parsing
Better error handling during streaming operations with detailed error messages
Support for concurrent tool calls with proper state tracking

Anthropic Client

Added support for parallel tool calls with proper tool ID tracking
Improved streaming API with sophisticated delta handling for various content types
Enhanced support for the thinking capability through dedicated API parameters
Better tool call state management during streaming with improved payload handling
Improved error handling with retries for rate limit errors

OpenAI Client

Added comprehensive test coverage for DeepSeek API and OpenRouter API, ensuring compatibility
Added support for multiple models through the same client interface
Improved handling of multi-tool call responses with better state management
Enhanced streaming response processing with proper content type discrimination
Added support for controlling reasoning effort with token budgeting
Fixed streaming tool call processing to properly accumulate arguments

Improved API Consistency

Standardized streaming interfaces across all LLM providers
Unified reasoning content extraction and processing
Consistent tool call format and handling across providers
Better type safety for message processing and tool interactions
Normalized effort levels for reasoning capabilities across providers

Comprehensive Testing

Added extensive test matrix covering 20+ models across multiple providers
Specific test cases for parallel tool calls functionality
Tests for reasoning capabilities with various model configurations
Cross-provider compatibility tests ensuring API consistency
Tests for streaming functionality with all content types

Documentation Updates

Added detailed package-level documentation for streaming and reasoning packages
Updated examples to demonstrate the new streaming API
Added examples for using the reasoning capabilities
Updated tool call examples to show parallel tool usage
Improved error messages and logging

Borrowed Upstream Pull Requests

This release incorporates code and ideas from the following upstream pull requests:

Support for reasoning in Anthropic and OpenAI requests from #1228 by @ppipada
Improved streaming API with proper content type discrimination from #950 by @diemus

Dependencies Updates

Same as in v0.1.13-update.1.

Contributors

Full Changelog: v0.1.13-update.1...v0.1.13-update.2

Contributors

asdek, ppipada, and diemus

Assets 2

08 May 20:26

asdek

v0.1.13-update.1

9c8174e

v0.1.13-update.1

Major Announcement

Important: This fork can now be used without replace directives in go.mod. Simply run:

go get github.com/vxcontrol/langchaingo@v0.1.13-update.1

Highlights

This release builds on all improvements from v0.1.13-update.0, adding several enhancements to key components and improving overall code quality.

Major Improvements

Ollama Integration

Refactored message processing and model handling for better maintainability
Improved method organization with dedicated functions for conversions and request creation
Enhanced error handling during chat request creation and response processing
Optimized message conversion logic for text content and tool calls

OpenAI Functions Agent

Added tool name support to ToolChatMessage for better tracking
Enhanced tool call parsing with dedicated methods
Improved message structure in the agent's scratchpad
Added support for handling multiple choices and tool calls in responses

Tool Enhancements

Web Scraper: Replaced traditional map and mutex with sync.Map for improved concurrency
Implemented safe concurrent writes with a dedicated writeToBuffer function
Streamlined the logic for checking and storing scraped links
Enhanced error handling for streaming output

Anthropic Client

Improved delta event handling in message processing
Added dedicated functions for handling text and JSON deltas
Enhanced error handling for streaming output
Improved response content management

Dependencies Updates

github.com/testcontainers/testcontainers-go v0.37.0
github.com/testcontainers/testcontainers-go/modules/* v0.37.0
github.com/stretchr/testify v1.10.0
github.com/cpuguy83/dockercfg v0.3.2
github.com/docker/docker v28.0.1+incompatible
go.opentelemetry.io/otel/metric v1.32.0
go.opentelemetry.io/otel/trace v1.32.0
github.com/weaviate/weaviate v1.29.0
github.com/weaviate/weaviate-go-client/v5 v5.0.2

Contributors

@asdek

Full Changelog: v0.1.13-update.0...v0.1.13-update.1

Contributors

asdek

Assets 2

06 May 20:02

asdek

v0.1.13-update.0

30fb496

v0.1.13-update.0 Pre-release

Pre-release

Highlights

Official Ollama Client Integration: Added native tool support for Ollama using the official client
In-Memory Vector Store: Added a new vector store implementation for quick prototyping and testing
Advanced Retrievers: Added Merger Retriever and Multi Query Retriever for improved document retrieval
MS Office Document Loaders: Added loaders for Word, Excel, and PowerPoint files
OpenAI Text-to-Speech: Integrated the OpenAI TTS API
Audio Transcription: Added support for transcribing audio files using the Whisper model

Breaking Changes

Go 1.24 Requirement: Updated minimum supported Go version to 1.24
Official Ollama Go Client: Replaced custom implementation with the official Ollama Go client
Changed RecursiveCharacter Default Settings: Reduced chunk size from 4000 to 512 and overlap from 200 to 100
Google AI Model Change: Updated default model from "gemini-pro" to "gemini-2.0-flash"
Response Status Code Check: Changed DuckDuckGo API check from status != 200 to status > 299
XML Formatting for Documents: Added XML tag formatting in map-reduce chains and document formatting

Major Features & Improvements

Tool Integrations

Added arXiv tool for scientific paper searches by @longkeyy in #1042
Fixed OpenAI agents by migrating from functions to tools by @nonchan7720 in #1240
Added multi-tool call responses by @lonelycode in #1030
Fixed status code check in DuckDuckGo tool by @kidandcat in #968

Retrievers & Vector Stores

Added in-memory vector store implementation by @HatsuneMiku3939 in #613
Added Merger Retriever to combine results from multiple retrievers by @Abirdcfly in #747
Added Multi Query Retriever for improved search capabilities by @Abirdcfly in #650
Improved vector store tests for Pinecone, Qdrant, Redis, and Weaviate
Added HNSW indexing support for in-memory vector store

Document Loaders

Added Microsoft Office document loaders for Word, Excel, and PowerPoint by @Struki84 in #1068
Added Whisper OpenAI loader for audio transcription by @devalexandre in #696

Document Formatting

Added XML tagging for document content in map-reduce chains
Updated StuffDocuments chain with improved document formatting
Added metadata filtering functionality in StuffDocuments
Added template-based document joining with consistent formatting

Examples

Added short-term memory to chatbots example by @ju4nv1e1r4 in #1194
Added Ollama Retrieval Example by @hcmlinj in #1051
Added document loading example by @AnthonyDasse in #954
Added in-memory vector store example

Provider Updates

OpenAI

Added text-to-speech API integration by @mazyaryousefinia in #1126
Updated openaillm.go by @Arrowarcher in #1182
Fixed migration from functions to tools by @nonchan7720 in #1240
Added proper handling of ToolCallResponse in GenerateContent
Added ReasoningContent support to ContentChoice

Anthropic

Fixed streaming for tool calls by @rainu in #1193
Added unit tests for anthropic tools by @lwlee2608 in #952
Added improved handling for tool streaming by @rainu in #1193
Added PartialJSONContent support for streaming tool responses

Google AI

Added support for items and recursion in schema by @sami-sweng in #1234
Changed default model from "gemini-pro" to "gemini-2.0-flash"
Improved schema conversion and parameter handling

Ollama

Added native tool support using the official Ollama client by @rainu in #1197
Removed custom client implementation in favor of official API
Added tool call functionality in Ollama integration
Updated client options and parameters for better performance

DeepSeek

Added base URL for DeepSeek examples by @ValeryVerkhoturov in #1207

Perplexity

Updated model references from "llama-3.1-sonar-large-128k-online" to "sonar"
Updated default model to ModelSonar
Added support for new model variants with different context windows

Bug Fixes

Fixed toolCall unmarshalling by @wangjiancn in #932
Fixed incorrect comments about default values in NewRecursiveCharacter by @teddysupercuts in #1189
Fixed Text Splitters conceptual link by @elkcityhazard in #1206
Fixed case sensitivity in vector store metadata filters
Fixed exact-match search in Serpapi tool to prevent errors

Output Parsers

Added Markdown parser for extracting content from markdown code blocks

Dependencies Updates

Bumped http-proxy-middleware from 2.0.7 to 2.0.9 in /docs by @dependabot in #1230
Bumped @babel/helpers from 7.26.7 to 7.27.0 in /docs by @dependabot in #1224
Bumped @babel/runtime from 7.26.7 to 7.27.0 in /docs by @dependabot in #1223
Bumped @babel/runtime-corejs3 from 7.26.7 to 7.27.0 in /docs by @dependabot in #1222
Bumped image-size from 1.2.0 to 1.2.1 in /docs by @dependabot in #1203
Bumped prismjs from 1.29.0 to 1.30.0 in /docs by @dependabot in #1180
Added new indirect dependencies: github.com/ollama/ollama, github.com/coder/hnsw, github.com/chewxy/math32, and more