Plugin Custom LLM (OpenAI Compatible)

NocoBase plugin for integrating external LLM providers that support OpenAI-compatible /chat/completions API, with built-in response format normalization and response mapping for non-standard APIs.

Features

• OpenAI-compatible: Works with any LLM provider exposing /chat/completions endpoint
• Auto content detection: Handles both string and array content blocks ([{type: 'text', text: '...'}])
• Response mapping: Transform non-standard API responses to OpenAI format via JSON config (supports streaming SSE and JSON)
• Reasoning content: Display thinking/reasoning from DeepSeek-compatible providers (multi-path detection)
• Stream keepalive: Prevent proxy/gateway timeouts during long model thinking phases
• Tool calling support: Gemini-compatible tool schema fixing (Zod + JSON Schema)
• Configurable: JSON config editors for request and response customization
• Locale support: English, Vietnamese, Chinese

Installation

Upload plugin-custom-llm-x.x.x.tgz via NocoBase Plugin Manager UI, then enable.

Configuration

Provider Settings

| Field | Description | |---|---| | Base URL | LLM endpoint URL, e.g. https://your-llm-server.com/v1 | | API Key | Authentication key | | Disable Streaming | Disable streaming for models that return empty stream values | | Stream Keep Alive | Enable keepalive to prevent timeouts during long thinking phases | | Keep Alive Interval | Interval in ms between keepalive signals (default: 5000) | | Keep Alive Content | Visual indicator text during keepalive (default: ...) | | Timeout | Custom timeout in ms for slow-responding models | | Request config (JSON) | Optional. Extra request configuration | | Response config (JSON) | Optional. Response parsing and mapping configuration |

Request Config

{
  "extraHeaders": { "X-Custom-Header": "value" },
  "extraBody": { "custom_field": "value" },
  "modelKwargs": { "stop": ["\n"] }
}

• extraHeaders — Custom HTTP headers sent with every request
• extraBody — Additional fields merged into the request body
• modelKwargs — Extra LangChain model parameters (stop sequences, etc.)

Response Config

{
  "contentPath": "auto",
  "reasoningKey": "reasoning_content",
  "responseMapping": {
    "content": "message.response"
  }
}

• contentPath — How to extract text from LangChain chunks. "auto" (default) detects string, array, and object formats. Or use a dot-path like "0.text"
• reasoningKey — Key name for reasoning/thinking content in additional_kwargs (default: "reasoning_content")
• responseMapping — Maps non-standard LLM responses to OpenAI format before LangChain processes them:
  • content — Dot-path to the content field in the raw response (e.g. "message.response", "data.text")
  • role — Dot-path to role field (optional, defaults to "assistant")
  • id — Dot-path to response ID (optional)

Response Mapping Examples

| Raw LLM Response | responseMapping.content | |---|---| | {"message": {"response": "..."}} | message.response | | {"data": {"text": "..."}} | data.text | | {"result": "..."} | result | | {"output": {"content": {"text": "..."}}} | output.content.text |

Model Settings

Standard OpenAI-compatible parameters: temperature, max tokens, top P, frequency/presence penalty, response format, timeout, max retries.

Changelog

v1.2.0

• Fix: Keepalive no longer interferes with tool call sequences (prevents tool call corruption)
• Fix: Gemini-compatible tool schema fixing — handles Zod schemas via dual-phase approach (pre/post conversion)
• Fix: Keepalive content no longer contaminates saved messages in DB
• Fix: Response metadata extraction with long ID sanitization (>128 chars truncated)
• Fix: Multi-path reasoning content detection (additional_kwargs + kwargs.additional_kwargs)
• Fix: Improved error recovery in keepalive consumer (immediate error propagation)

v1.1.1

• Stream keepalive proxy for long thinking phases
• Response mapping for non-standard LLM APIs

v1.0.0

• Initial release with OpenAI-compatible LLM provider support

License

Apache-2.0