RAGFlow

RAGFlow is a RAG engine. It parses and chunks your documents, stores vectors per chunk, and answers questions with chat over the retrieved chunks. You set models in the web UI.

Upstream: source, docs.

Plugboard options

Plugboard serves the LLM role. Every dataset needs an embedding model, and plugboard serves none, so RAGFlow needs a second provider for that.

Role	Served by
LLM	Plugboard, through the OpenAI-API-Compatible provider
Embedding (required)	Another provider
Rerank (optional)	Another provider

Give RAGFlow its own key and a chat alias. Docker needs a reachable address (Serving other machines).

{
  "options": {
    "listen": { "host": "0.0.0.0", "port": 4010 },
    "keys": { "ragflow": "{env:PLUGBOARD_KEY_RAGFLOW}" },
    "models": ["anthropic/claude-sonnet-4-6"],
    "aliases": { "chat": "anthropic/claude-sonnet-4-6" },
    "apps": { "ragflow": { "allowedModels": ["chat"] } }
  }
}

Connect

The LLM

1. Open User settings, Model providers, and choose OpenAI-API-Compatible.
2. Create an instance. Enter the ragflow key as the API Key and http://host.docker.internal:4010/v1 as the Base URL. The /v1 is required. Save and verify.
3. Add a custom model named chat, the plugboard alias. Choose the LLM type and set Max tokens. Turn on Tool call only if the model behind the alias supports tools. The alias may not appear in the auto-listed models, so add it by hand.
4. Under Set default models, set chat as the default LLM.

The embedding model

Add a second provider instance for an embeddings endpoint, such as a local server or a hosted one. The same OpenAI-API-Compatible card works if that server speaks /v1/embeddings. Add the model with the Embedding type and set it as the default embedding model.

Pick the embedding model before you index. Changing it after parsing affects existing indexes.

Reranking (optional)

Retrieval reranks only when you set a rerank model. Configure another provider for it. Plugboard has no rerank route.

What works

• Chat and answers through the chat alias.
• Streaming with usage.

Limits

• Uploading, indexing and retrieval need the embedding provider.
• The Thinking toggle does nothing through plugboard. Leave it off.
• Keep dropUnsupportedParams on. RAGFlow can send temperature and top_p together. A Claude 4.6 or older model takes only one of them, so plugboard drops top_p and lists it in x-plugboard-dropped-params.
• Speech models (ASR, TTS) are not served.