What’s New in Higress 2.0? 30 Updates Including RAG MCP Server and Performance Fixes

Release Overview

This release bundles 30 updates (including 2 critical ones) for the Higress cloud‑native gateway, covering new functionality, bug fixes, refactoring, documentation, and testing improvements.

Key New Features

1. RAG MCP Server

Background : Modern applications need efficient knowledge‑management and retrieval. The RAG (Retrieval‑Augmented Generation) approach combines vector search with LLM generation. This PR adds a Model Context Protocol (MCP) server that provides knowledge‑block creation, block management, keyword search, and chat capabilities.

Core Functions :

Knowledge management – create knowledge blocks from text.

Block management – list and delete blocks.

Search – keyword‑based retrieval.

Chat – send messages and receive LLM responses.

The implementation uses external libraries such as github.com/dlclark/regexp2, github.com/milvus-io/milvus-sdk-go/v2, and github.com/pkoukk/tiktoken-go for regex, vector DB, and token encoding.

Configuration Steps :

Enable the MCP server in higress-config and set the path and options.

Configure RAG parameters (splitter type, chunk size, overlap, top_k, threshold).

Set LLM provider credentials (API key, model name).

Set embedding provider credentials.

Configure the vector‑database connection.

rag:
  splitter:
    type: "recursive"
    chunk_size: 500
    chunk_overlap: 50
  top_k: 5
  threshold: 0.5
llm:
  provider: "openai"
  api_key: "your-llm-api-key"
  model: "gpt-3.5-turbo"
embedding:
  provider: "openai"
  api_key: "your-embedding-api-key"
  model: "text-embedding-ada-002"
vectordb:
  provider: "milvus"
  host: "localhost"
  port: 19530
  collection: "test_collection"

Value : Provides a complete knowledge‑management pipeline, improving efficiency and accuracy for large‑scale text processing.

2. ECDS Configuration Refactor

Problem : Directly embedding Golang filter configuration in HTTP_FILTER caused map ordering issues and connection‑drain when the configuration changed.

Solution : Split the configuration into two parts— HTTP_FILTER now only references a config_discovery filter, while the actual Golang filter settings are moved to EXTENSION_CONFIG. The refactor introduces constructMcpSessionStruct and constructMcpServerStruct to produce an EXTENSION_CONFIG -compatible structure and updates unit tests accordingly.

Benefit : Eliminates connection‑drain during config updates, improves stability, and simplifies future configuration changes.

Bug Fixes

Fix Authorization header overwrite issue (#2904) – ensures the original Authorization header is preserved.

Support AttackLevel field in MultiModalGuard (#2938) – restores security checks for high‑risk prompts.

Correct JSON unmarshalling for Claude API array‑content responses (#2898) – prevents parsing errors.

Resolve SSE chunk handling for Claude streaming (#2889) – prevents connection blockage.

Fix HTTP request handling that could drop the original Authorization header (#2904).

Refactoring and Optimizations

Removed duplicate think tags in Bedrock and Vertex plugins (#2933) – reduces code redundancy.

Optimized host matching by replacing regex with simple string operations and port stripping (#2890) – improves performance.

Updated matchDomain to use HostMatcher structure for faster matching.

General code cleanup and removal of unused fields in MCP server implementation.

Documentation Updates

Fixed broken link in README_JP.md and added missing sections in README.md (#2915).

Improved English and Chinese docs for hmac-auth-apisix plugin (#2912).

Corrected syntax errors across README files (#2880).

Added detailed usage instructions for the RAG MCP server and ECDS refactor.

Testing Improvements

Extended test coverage for ai-security-guard component, adding new cases and adjusting existing logic (#2928).

Release Statistics

🚀 New Features: 13 items

🐛 Bug Fixes: 7 items

♻️ Refactoring/Optimizations: 5 items

📚 Documentation Updates: 4 items

🧪 Test Improvements: 1 item

Overall, this release enhances Higress’s knowledge‑management capabilities, stabilizes configuration handling, and provides clearer documentation, delivering a more flexible and reliable cloud‑native gateway experience.