Kenhuang Academy
  • English

      • Model Providers

      Every Dify app needs at least one model provider. Keys are workspace-scoped: Owner / Admin configures them once, every app can pick those models. Official entry: Integrations → Model Provider (the Marketplace installs more plugin providers).

      Add a cloud vendor

      Typical path:

      1. Find the vendor under “Install model providers” or in the Marketplace
      2. After install, click Setup on the card
      3. Paste the API key; add a custom endpoint / organization if you use a proxy or dedicated cloud
      4. Dify validates before enabling—do not select a failed provider in an app
      KindExamplesNotes
      Global cloudOpenAI, AnthropicWatch network and billing
      Regional cloudDeepSeek, Qwen / Tongyi, …Common default reasoners in CN
      LocalOllamaBase URL, no cloud key; see Ollama Integration

      Which cards you see depends on version and Marketplace—do not memorize a vendor list.

      Models are not only chat

      One provider often exposes several model types. Set Default Models (usually top-right) per job:

      DefaultUsed for
      System ReasoningChat / reasoning when an app or node picks nothing
      EmbeddingKnowledge indexing and retrieval
      RerankOptional re-score of retrieved chunks
      Speech-to-Text / TTSVoice in/out if you need it

      A High Quality knowledge index requires a working embedding model. Strong chat plus a missing embedder is the most common first-week stall.

      Multiple keys and custom models

      • Add more than one key per vendor to split dev/prod or quota (Configure / manage credentials on the card)
      • Missing or fine-tuned models: Add Model on vendors that support custom models; supply name and credentials
      • If name and type match an existing model, Dify attaches the new key instead of duplicating the card
      • Deleting a custom model’s last key removes the model; leftover credentials may remain under “manage credentials”

      Cloud may also spend platform AI credits on a popular subset. Bring your own account when you need higher limits. Self-host almost always means your keys or your Ollama.

      Choosing and comparing inside an app

      Chatbot / Agent / LLM nodes pick a model from a dropdown. Leave sampling defaults first. Compare with one frozen prompt:

      1. Fix a user question (include one that should be refused)
      2. Run two or three models
      3. Score instruction following, language quality, latency, price
      4. Agent preview can Debug as Multiple Models (docs: up to about four) side by side

      Tool-using agents prefer models with native Function Calling. Agent Settings labels Function Calling vs fallback ReAct. See Agent.

      Safety

      • Keep keys in the Dify workspace (or self-host secrets)—not in frontend code, shared DSL packs, or public Git
      • Use workspace environment variables for secrets so exported DSL does not leak them
      • Rotate by adding the new key before deleting the old one, or live apps get 401

      Versus LangChain: there you export OPENAI_API_KEY into a process; Dify is the same idea as a workspace UI.

      Troubleshooting

      Validation fails? Check key scope, quota, whether the endpoint needs /v1, and outbound HTTPS proxies.

      Chat works, knowledge ingest does not? Set a default embedding model or pick one in the knowledge-create flow.

      Ollama will not connect? Do not use localhost:11434 from Docker. See Ollama Integration.

      Next steps

      评论