Skip to content

AI Provider Setup

Jump to bottom
Will Miao edited this page Jul 13, 2026 · 1 revision

Overview

LoRA Manager can use an LLM (Large Language Model) provider to read a HuggingFace model card (README.md) and extract structured metadata automatically.

This page explains how to configure a provider and what features make use of it.

Configuration

Open Settings → AI Provider and set the following:

Field Description
Provider Select your LLM provider.
API Base URL Populated automatically for preset providers. Set manually for Custom.
API Key Required for most providers. Not needed for local Ollama.
Model The model to use for extraction. Select from the catalog or type a custom name.

Available providers:

  • OpenAI
  • Ollama (local)
  • DeepSeek
  • Groq
  • OpenRouter
  • OpenCode Go
  • Custom

All providers use the OpenAI-compatible API format, so any compatible endpoint (vLLM, LM Studio, etc.) works with the Custom option.

What Uses the AI Provider

Enrich HF Metadata (AI)

Right-click a model that was downloaded from (or linked to) HuggingFace, and select Enrich HF Metadata (AI). LoRA Manager fetches the model's HuggingFace README and sends it to the configured LLM. The LLM extracts the following fields, which are then written to the model's .metadata.json:

  • base_model -- only written if currently empty
  • Trigger words (trainedWords) -- only written if none exist
  • Short description (modelDescription) -- only written if currently empty
  • Tags -- merged with existing tags, deduplicated
  • Preview image -- downloaded and optimized when found in the README
  • Notes -- plain-text usage summary from the model card
  • Usage tips -- structured JSON with recommended strength ranges, CLIP skip, etc.

This feature requires the model to have an hf_url set. Use Link to HuggingFace in the right-click context menu to associate a model with a HuggingFace repo first.

Recommended Models

  • Online: deepseek-v4-flash -- extremely fast and cheap.
  • Local: qwen3.5:9b Q4_K_M -- good balance of quality and resource usage.

More LLM-powered features are planned and will be added gradually.

Clone this wiki locally