Universal Reranker (n8n Community Node)

Universal Reranker provides document reranking capabilities for n8n workflows. It supports OpenAI-compatible rerank endpoints (vLLM, LocalAI, Infinity, custom) and Cohere Rerank API. The package includes two specialized nodes designed for different use cases.

Nodes

Universal Reranker Provider

AI provider node that connects to vector stores (like PGVector) to rerank retrieved documents. The vector store calls this node automatically during retrieval operations.

Universal Reranker (Flow)

Regular workflow node for reranking document arrays in your n8n flows. You provide the query and document field, and it outputs reranked results.

Installation

Option A - Community Nodes (Recommended)

1. In n8n: Settings → Community Nodes → Install
2. Search for "universal reranker" or use package name n8n-nodes-universal-reranker
3. Restart n8n when prompted

Option B - Self-Hosted Installation

Local Installation:

cd ~/.n8n/custom
npm install n8n-nodes-universal-reranker

Docker Installation:

# Mount custom folder and install inside container
docker exec -it n8n npm install n8n-nodes-universal-reranker

Restart n8n after installation.

Configuration

Service Options

• OpenAI-Compatible: Works with vLLM, LocalAI, Infinity, and custom endpoints
  • Set Endpoint URL (e.g., http://localhost:7997/rerank)
  • Set Model name (e.g., BAAI/bge-reranker-v2-m3)
• Cohere: Uses Cohere's rerank API
  • Select from predefined models or choose "Custom" for specific models
  • Requires Cohere API credentials

Parameters

• Top K: Maximum number of documents to return after reranking
• Threshold: Minimum relevance score (0-1) for returned documents
• Include Original Scores: Whether to preserve original document scores
• Enable Caching: Cache reranking results to improve performance for repeated queries
• Cache TTL: Time to live for cached results in minutes (1-60, default: 5)
• Enable Custom Templates: For special models like Qwen3 Reranker (see TEMPLATES.md)

Cohere Models

• rerank-v3.5 (default)
• rerank-english-v3.0
• rerank-multilingual-v3.0
• Custom: Enter any specific Cohere model name

Caching

Both nodes support optional caching to reduce API calls and improve performance. Caching is disabled by default and can be enabled per node.

Documentation

• EXAMPLES.md - Usage examples and workflow configurations
• TEMPLATES.md - Custom template guide for special models (Qwen3, etc.)

Docker Networking

When n8n runs in Docker, use host.docker.internal to access services on your host:

• Infinity: http://host.docker.internal:7997/rerank
• vLLM: http://host.docker.internal:8000/v1/rerank
• LocalAI: http://host.docker.internal:8080/v1/rerank

Troubleshooting

Common Issues

"No documents found in field"

• Verify your documentsField parameter matches the actual field name
• Ensure the field contains an array of documents

"API Error (404)"

• Check that your endpoint URL is correct
• Verify your reranker service is running and accessible
• Test the endpoint directly with curl

"Request failed: connect ECONNREFUSED"

• Ensure your reranker service is accessible from n8n
• Check Docker network configuration if using containers
• Verify firewall/security group settings

Debug Mode

Enable debug logging in n8n to see detailed error messages:

export N8N_LOG_LEVEL=debug

Document Format

The nodes expect documents in one of these formats:

• pageContent field (LangChain format)
• text field
• content field
• document field
• If none found, the entire document object is stringified

Output includes:

• _rerankScore: Relevance score from reranking service
• _originalIndex: Original position in input array
• _originalScore: Original document score (if includeOriginalScores is true)

Development

Testing

The package includes a comprehensive test suite with 78+ tests covering:

• Core reranking functionality
• Caching behavior
• Error handling and edge cases
• Both node types (Provider and Flow)

# Run tests
pnpm test

# Run tests with coverage
pnpm run test:coverage

# Build and test
pnpm run prepublishOnly

Contributing

Contributions are welcome!

License

MIT License