n8n-nodes-ollama-reranker

Advanced Reranker Provider for n8n - Supporting Ollama-compatible APIs, Custom Rerank servers, and VL Classifiers.

⚠️ Important Note: Sorry folks, Ollama doesn't natively support reranker models! We're developing our own solution to bring powerful reranking capabilities to n8n. This package works with Ollama-compatible APIs that implement reranking through prompt-based scoring, custom rerank endpoints, and now Vision-Language classification servers.

Features

• 🎯 Integrates seamlessly with n8n Vector Store nodes
• 🚀 Multiple API types: Ollama Generate, Custom Rerank, VL Classifier
• 🤖 Auto-detection of server capabilities
• 🔧 Multiple models supported (BGE Reranker, Qwen3 family)
• 🎨 VL Classification for document complexity analysis (v1.4.0+)
• ⚡ Concurrent processing with configurable batch sizes
• 🔄 Automatic retries with exponential backoff
• 📊 Flexible scoring with threshold and topK parameters

Installation

Via npm (Recommended)

npm install n8n-nodes-ollama-reranker

Via n8n Community Nodes UI

1. Go to Settings → Community Nodes
2. Select Install
3. Enter n8n-nodes-ollama-reranker
4. Click Install

From Docker

Add to your n8n Dockerfile:

FROM n8nio/n8n:latest
USER root
RUN cd /usr/local/lib/node_modules/n8n && \
    npm install n8n-nodes-ollama-reranker
USER node

Prerequisites

Choose your server type:

Option 1: Ollama (Prompt-based reranking)

1. Ollama must be running and accessible
2. Pull a reranker model:

# Recommended - BGE Reranker v2-M3
ollama pull bge-reranker-v2-m3

# Or Qwen3 models
ollama pull dengcao/Qwen3-Reranker-4B:Q5_K_M

Option 2: Custom Rerank API

Use any service implementing /api/rerank endpoint (like deposium-embeddings-turbov2)

Option 3: VL Classifier Server (NEW in v1.4.0)

Deploy a Vision-Language classifier server with:

• /api/status - Server health and capabilities
• /api/classify - Document complexity classification
• Optional /api/rerank - Direct reranking support

Example: deposium_embeddings-turbov2 with ResNet18 ONNX INT8 model

Usage

Basic Setup

1. Add an Ollama Reranker node to your workflow
2. Connect it to a Vector Store node (e.g., Pinecone, Qdrant, Supabase)
3. Configure:
   • API Type: Choose between:
     • Ollama Generate API - Standard Ollama prompt-based
     • Custom Rerank API - Direct reranking endpoint
     • VL Classifier + Reranker - Vision-Language classification
     • Auto-Detect - Automatically detect server type
   • Model: Select a reranker model
   • Top K: Number of documents to return
   • Threshold: Minimum relevance score (0-1)
   • Base URL: URL to your server

VL Classifier Options (v1.4.0+)

When using VL Classifier:

• Enable VL Classification: Use complexity analysis
• Classification Strategy:
  • Metadata - Add complexity as document metadata
  • Filter - Filter by complexity before reranking
  • Both - Combine filtering and metadata
• Filter Complexity: Keep LOW, HIGH, or both complexity documents

Example Workflow

User Query → Vector Store (retrieve 50 docs)
           → Ollama Reranker (rerank to top 10)
           → Continue with top-ranked documents

Supported Configurations

Reranker Models (Ollama/Custom API)

Model	Size	Speed	Accuracy	Best For
bge-reranker-v2-m3	~600MB	⚡⚡⚡	⭐⭐⭐⭐	General purpose (Recommended)
Qwen3-Reranker-0.6B	~400MB	⚡⚡⚡⚡	⭐⭐⭐	Low resource environments
Qwen3-Reranker-4B	~2.5GB	⚡⚡	⭐⭐⭐⭐	Balanced performance
Qwen3-Reranker-8B	~5GB	⚡	⭐⭐⭐⭐⭐	Maximum accuracy

VL Classifier Models

Model	Size	Speed	Use Case
ResNet18-ONNX-INT8	11MB	⚡⚡⚡⚡	Document complexity classification
Custom VL models	Varies	Varies	Vision-Language tasks

Development

Setup

cd custom-nodes
npm install

Build

npm run build

Lint & Format

npm run lint        # Check linting
npm run lint:fix    # Auto-fix issues
npm run format      # Format code

Test

npm test

Local Testing

Use the included docker-compose.yml:

docker-compose up -d

Access n8n at http://localhost:5678 (admin/admin)

Git Hooks

Pre-commit hooks are configured with Husky to:

• Run lint-staged (ESLint + Prettier)
• Run qlty quality checks

Publishing

Publishing to npm is automated via GitHub Actions:

1. Update version in custom-nodes/package.json
2. Commit changes
3. Create and push a tag:

git tag v1.0.1
git push origin v1.0.1

GitHub Actions will automatically:

• Build the package
• Run tests
• Publish to npm

Project Structure

n8n_reranker/
├── custom-nodes/              # Main npm package
│   ├── src/
│   │   └── nodes/
│   │       └── OllamaReranker/
│   │           ├── OllamaReranker.node.ts
│   │           ├── OllamaReranker.node.test.ts
│   │           └── ollama.svg
│   ├── .eslintrc.js
│   ├── .husky/                # Git hooks
│   ├── .prettierrc.json
│   ├── package.json           # Main package.json
│   ├── tsconfig.json
│   └── jest.config.js
├── Dockerfile                 # For local development
├── docker-compose.yml         # Complete dev environment
└── .github/
    └── workflows/
        └── npm-publish.yml    # Automated publishing

How It Works

The Reranking Challenge

Ollama doesn't natively support reranker models that output relevance scores. Instead, we implement three approaches:

1. Prompt-based Scoring: Use Ollama's /api/generate with specially formatted prompts
2. Custom Rerank API: Connect to servers with dedicated /api/rerank endpoints
3. VL Classification: Pre-process with Vision-Language models for intelligent filtering

API Type Detection

The node automatically detects your server type by checking:

1. /api/status → VL Classifier server
2. /api/tags → Ollama server
3. /api/rerank → Custom rerank server
4. Fallback → Ollama (default)

Architecture

This node implements two n8n patterns:

Provider Node (OllamaReranker)

1. No inputs - Provider nodes don't receive workflow data
2. AiReranker output - Connects to Vector Store nodes
3. supplyData() - Returns a reranker provider object
4. Standard interfaces:
   • rerank() - Main reranking method
   • compressDocuments() - LangChain compatibility

Workflow Node (OllamaRerankerWorkflow)

1. Main inputs/outputs - Processes workflow items
2. execute() - Transforms documents in the workflow
3. usableAsTool - Can be used as AI Agent tool

Contributing

1. Fork the repository
2. Create a feature branch
3. Make your changes
4. Run tests and linting
5. Submit a pull request

Security

• ✅ No vulnerabilities in dependencies
• ✅ form-data updated to 4.0.4 (fixes CVE-2025-7783)
• ✅ Code quality validated with qlty

License

MIT © Gabriel BRUMENT

Links

• GitHub Repository
• npm Package
• VL Classifier Integration Guide
• n8n Documentation
• Ollama Documentation

Support

For issues and feature requests, please use the GitHub Issues.