n8n-nodes-llm-observability

Developed and maintained by Ascenzia

Based on the original work by Ruby Lo (Wistron DXLab) - n8n-nodes-openai-langfuse

npm package: https://www.npmjs.com/package/n8n-nodes-llm-observability

Overview

Standalone AI Agent nodes for n8n with multi-provider LLM support and integrated observability (Langfuse, Arize Phoenix, or Helicone).

Why This Package?

The native n8n AI Agent requires connecting separate LLM nodes and doesn't have built-in observability. This package provides:

• All-in-one AI Agent: Select your LLM provider directly in the node (no separate LLM node needed)
• Multi-provider support: OpenAI, Anthropic, Azure OpenAI, Google Gemini, AWS Bedrock, Groq, Mistral, Ollama, xAI Grok, vLLM, and OpenAI-compatible APIs
• Integrated observability: Choose between Langfuse, Arize Phoenix, or Helicone for tracing
• Token usage tracking: Token consumption displayed in n8n logs and available in workflow output
• Clean architecture: Nodes that don't interfere with n8n's native AI components

Nodes

| Node | Description | |------|-------------| | AI Agent Langfuse | AI Agent with integrated Langfuse tracing | | AI Agent Phoenix | AI Agent with integrated Arize Phoenix tracing (OpenTelemetry) | | AI Agent Helicone | AI Agent with integrated Helicone observability (proxy-based) | | Qdrant Search Tool (Langfuse) | Vector search tool with Langfuse observability | | Qdrant Search Tool (Phoenix) | Vector search tool with Phoenix observability | | Qdrant Search Tool (Helicone) | Vector search tool with Helicone observability |

Architecture

[AI Agent Langfuse / Phoenix / Helicone] --> Output (with token usage)
         |
         ├── Provider: OpenAI / Anthropic / Azure / Gemini / Bedrock / Groq / Mistral / Ollama / Grok / vLLM / OpenAI Compatible
         ├── Memory (optional)
         ├── Tools (optional)
         ├── Output Parser (optional)
         └── Observability: Langfuse, Phoenix, or Helicone

Installation

Community Nodes (Recommended)

For n8n v0.187+, install directly from the UI:

1. Go to Settings > Community Nodes
2. Click Install
3. Enter n8n-nodes-llm-observability
4. Agree to the risks of using community nodes
5. Click Install

Manual Installation

cd ~/.n8n
npm install n8n-nodes-llm-observability
n8n start

Supported Providers

| Provider | Description | |----------|-------------| | OpenAI | GPT-4, GPT-4o, GPT-3.5-turbo, o1, etc. | | Anthropic | Claude 3.5, Claude 3, etc. | | Azure OpenAI | Azure-hosted OpenAI models | | Google Gemini | Gemini Pro, Gemini Flash, etc. | | AWS Bedrock | Amazon Bedrock models (Claude, Llama, etc.) | | Groq | Ultra-fast inference (Llama, Mixtral) | | Mistral | Mistral AI models | | Ollama | Local models via Ollama | | xAI Grok | Grok models | | vLLM | High-throughput self-hosted LLM serving | | OpenAI Compatible | Any OpenAI-compatible API |

Credentials

Langfuse Credentials

Each provider has a combined credential that includes both the LLM provider settings and Langfuse configuration:

• OpenAI + Langfuse API
• Anthropic + Langfuse API
• Azure OpenAI + Langfuse API
• Google Gemini + Langfuse API
• AWS Bedrock + Langfuse API
• Groq + Langfuse API
• Mistral + Langfuse API
• Ollama + Langfuse API
• xAI Grok + Langfuse API
• vLLM + Langfuse API
• OpenAI Compatible + Langfuse API
• Qdrant + OpenAI + Langfuse API (for Qdrant Search Tool)

Phoenix Credentials

Each provider has a combined credential that includes both the LLM provider settings and Phoenix configuration:

• OpenAI + Phoenix API
• Anthropic + Phoenix API
• Azure OpenAI + Phoenix API
• Google Gemini + Phoenix API
• AWS Bedrock + Phoenix API
• Groq + Phoenix API
• Mistral + Phoenix API
• Ollama + Phoenix API
• xAI Grok + Phoenix API
• vLLM + Phoenix API
• OpenAI Compatible + Phoenix API
• Qdrant + Phoenix API (for Qdrant Search Tool)

Helicone Credentials

Each provider has a combined credential that includes both the LLM provider settings and Helicone configuration:

• OpenAI + Helicone API
• Anthropic + Helicone API
• Azure OpenAI + Helicone API
• Google Gemini + Helicone API
• Groq + Helicone API
• Mistral + Helicone API
• OpenAI Compatible + Helicone API
• Qdrant + Helicone API (for Qdrant Search Tool)

Note: Helicone works as a proxy - requests are routed through Helicone's servers for observability without requiring SDK changes.

Configuration

Provider Settings

| Field | Description | |-------|-------------| | Provider | Select your LLM provider | | Model | Model name (e.g., gpt-4o, claude-3-5-sonnet-20241022) | | Temperature | Creativity (0-2, default: 0.7) | | Max Tokens | Maximum response length |

Prompt Settings

| Field | Description | |-------|-------------| | Prompt Type | "Define Below" or "Take From Previous Node" | | System Message | Instructions for the AI agent | | User Message | The user's input/question |

Langfuse Options

| Field | Type | Description | |-------|------|-------------| | Session ID | string | Group related traces together | | User ID | string | Identify the end user | | Trace Name | string | Custom name for the trace | | Tags | string | Comma-separated tags for filtering | | Custom Metadata | JSON | Additional context as JSON |

Phoenix Options

| Field | Type | Description | |-------|------|-------------| | Session ID | string | Group related traces together | | User ID | string | Identify the end user | | Tags | string | Comma-separated tags for filtering |

Helicone Options

| Field | Type | Description | |-------|------|-------------| | Session ID | string | Group related requests into a session | | User ID | string | Identify the end user | | Custom Properties | key-value | Add custom properties to requests (e.g., environment, app version) |

Agent Options

| Field | Description | |-------|-------------| | Max Iterations | Maximum tool calling iterations (default: 10) | | Return Intermediate Steps | Include tool call details in output |

Inputs

| Input | Required | Description | |-------|----------|-------------| | Memory | No | Standard n8n memory nodes for conversation history | | Tool | No | Standard n8n tool nodes for agent capabilities | | Output Parser | No | Standard n8n output parsers |

Output

The node outputs:

{
  "output": "The AI response...",
  "tokenUsage": {
    "promptTokens": 150,
    "completionTokens": 250,
    "totalTokens": 400
  },
  "intermediateSteps": [...]
}

• output: The AI agent's response
• tokenUsage: Token consumption (accumulated across all LLM calls when using tools)
• intermediateSteps: Tool call details (when enabled)

Note: Token usage availability depends on the LLM provider. Some providers (like Ollama) may not return token information.

Usage Examples

Basic AI Agent with Langfuse

1. Add AI Agent Langfuse node
2. Select your Provider (e.g., OpenAI)
3. Configure credentials (includes Langfuse settings)
4. Enter the Model name (e.g., gpt-4o)
5. Configure the prompt

AI Agent with Phoenix Tracing

1. Add AI Agent Phoenix node
2. Select your Provider
3. Configure credentials (includes Phoenix collector URL)
4. Enter the Model name
5. Configure the prompt

AI Agent with Helicone

1. Add AI Agent Helicone node
2. Select your Provider (OpenAI, Anthropic, Azure, Gemini, Groq, Mistral)
3. Configure credentials (includes Helicone API key)
4. Enter the Model name
5. Configure the prompt
6. Optionally add Session ID, User ID, or Custom Properties

AI Agent with Tools

1. Set up the AI Agent (Langfuse, Phoenix, or Helicone)
2. Add n8n tool nodes (Calculator, HTTP Request, Code, Qdrant Search, etc.)
3. Connect tools to the Tool input

Using Qdrant Vector Search

1. Add Qdrant Search Tool (Langfuse), Qdrant Search Tool (Phoenix), or Qdrant Search Tool (Helicone)
2. Configure Qdrant credentials
3. Connect to the AI Agent's Tool input

Using Self-hosted vLLM

1. Select vLLM as provider
2. Configure vLLM credentials (Base URL of your vLLM server)
3. Enter the model name (e.g., meta-llama/Llama-3.1-8B-Instruct)

Compatibility

• n8n: Version 1.117.0 or later (tested on 1.117+)
• Node.js: Version 20.15 or later
• Langfuse: Cloud and self-hosted instances
• Arize Phoenix: Cloud and self-hosted instances
• Helicone: Cloud instance (helicone.ai)

Resources

• n8n Community Node Docs
• Langfuse Documentation
• Arize Phoenix Documentation
• Helicone Documentation
• vLLM Documentation
• n8n Community Forum
• Ascenzia

Version History

• v0.10.40 - Add Qdrant Search Tool for Helicone
• v0.10.39 - Add Helicone observability support (AI Agent Helicone node)
• v0.10.38 - Fix: accumulate token usage across all LLM calls (when using tools)
• v0.10.37 - Add token usage to output and n8n AI logs panel
• v0.10.36 - Update Phoenix logos to official Arize branding
• v0.10.35 - Add all LLM providers to Phoenix node (Gemini, Bedrock, Groq, Mistral, Ollama, Grok, vLLM, OpenAI Compatible)
• v0.10.34 - Show provider name in Phoenix span (agent_openai, agent_anthropic, etc.)
• v0.10.33 - Implement OpenInference semantic conventions for Phoenix tracing
• v0.10.32 - Fix Azure OpenAI max_tokens error (use max_completion_tokens)
• v0.10.30 - Add Azure OpenAI support for Phoenix nodes
• v0.10.29 - Add Arize Phoenix support: AI Agent Phoenix node with OpenTelemetry tracing
• v0.10.x - Bug fixes: trace tool calls as separate spans, handle nested JSON from vector stores
• v0.9.x - Enhanced Langfuse tracing and stability improvements
• v0.8.0 - Standalone AI Agent with integrated multi-provider LLM support
• v0.7.0 - Simplified package: AI Agent with Langfuse + vLLM only
• v0.6.0 - AI Agent with integrated observability
• v0.5.0 - Added Grok (xAI) support
• v0.4.0 - Simplified architecture
• v0.3.0 - Modular architecture
• v0.2.x - Multi-provider support with integrated Langfuse
• v0.1.x - Initial release (OpenAI + Langfuse)

License

MIT

Original project: n8n-nodes-openai-langfuse by Ruby Lo (Wistron DXLab)

Extended and maintained by: Ascenzia