Replies: 3 comments 2 replies
-
|
This issue is usually caused by the embedding model not being properly registered or linked in the dataset configuration, which makes Dify silently fall back to the default model—often a chat model—for embedding tasks. This fallback can happen if the dataset’s embedding_model_provider or embedding_model fields are missing or misconfigured, or if there are credential issues (especially in offline or migrated environments) [source]. Here’s what you can do:
Recent Dify updates have improved error handling and validation for embedding model configuration, so make sure you’re on a recent version and that your dataset’s embedding model settings are correct [source]. If you need to manually fix linkage in the database, you’ll want to ensure the provider, model, and credential tables are all correctly connected for your embedding model. Let me know if you need step-by-step SQL or shell commands for this. To reply, just mention @dosu. How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other
|
Beta Was this translation helpful? Give feedback.
-
|
When I delete the chat model Qwen3-14B, I only keep one embeding model in models list. I also make sure the knowledge make sure has been configured with Embeding-4B model. I did a test that I added a pdf doc and with same process like before. I clicked the 预览快 button and want to preview the split block result. I checked dify-api docker logs, I found below error. the dify still want to find Qwen3-14B, but I have delete it from models. It so strange that why dify still want to use correct Embeding-4B. Even the Embeding-4B model has some confiugration errors. I should found some error logs in Embeding-4B， but I can not find them. Below is the error logs of dify-api Here is the configration of Knowledge and modes setting: |
Beta Was this translation helpful? Give feedback.
-
|
The error happens because Dify's QA document generation (like the "预览快" split block preview) always uses the default LLM (chat) model, not the embedding model—even if only an embedding model is present or selected for the knowledge base. If the default LLM model (in your case, Qwen3-14B) is deleted but still referenced in the database (TenantDefaultModel table), Dify will keep trying to use it and fail with ProviderTokenNotInitError, as you saw in the logs. Deleting a model from the UI does not automatically update or clear this reference [source]. To fix this, you need to set a valid, existing LLM model as the default for your workspace. If you don't have any LLM models available, QA document generation features (like preview) will not work—they require an LLM, not just an embedding model [source]. You can update the default LLM model using the admin UI (Settings > Model Provider), or if that's not possible, by updating the TenantDefaultModel record in your database to point to an existing LLM model. After updating, restart the Dify backend to clear any cached references. If you only want to use embedding (and not QA preview), embedding will work as long as the embedding model is correctly configured in the knowledge base settings. But for any QA or split preview, an LLM is required. To reply, just mention @dosu. How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other
|
Beta Was this translation helpful? Give feedback.
-
|
I already test that I only have embeding model and no LLM Chat model. Whatever the embeding model set, the upload model will not call any embeding action, The document was not used embed mode to analyze at all. The knowledge not buid by pipeline seems only use normal LLM chat model and it don't care which embeding mode you set. 
The dify api logs: the embeding model logs only has |
Beta Was this translation helpful? Give feedback.
-
|
Dify only triggers embedding actions if the knowledge base is set up for embedding-only workflows. If you configure your knowledge base for QA segmentation (Q&A mode) or use features like split block preview, Dify always requires an LLM (chat) model—even if you only have an embedding model configured. Without an LLM, the pipeline aborts before any embedding is performed, which is why you see no embedding requests or logs, and only health checks on your embedding model server. This is a design limitation: QA segmentation and preview features always call the LLM, not the embedding model, and there is no fallback to embedding-only processing in these modes reference. To use only an embedding model, set your knowledge base segmentation mode to standard chunking (not Q&A) and avoid preview features that require LLM output. Embedding will then work as expected if your embedding model is valid and configured reference. If you need QA segmentation but have no LLM, you can register a dummy LLM model just for pipeline compatibility. Also, pipeline-based knowledge bases have stricter embedding model validation than traditional KBs, so some models may work in one but not the other reference. If embedding still doesn’t trigger, check that your worker is running, your embedding model credentials are valid, and your dataset settings match your intended workflow. To reply, just mention @dosu. How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other
|
Beta Was this translation helpful? Give feedback.
Uh oh!
There was an error while loading. Please reload this page.
-
Self Checks
1. Is this request related to a challenge you're experiencing? Tell me about your story.
I firstly added two models, one is Qwen3-14 for chat, Qwen3-4b-embeding for crate RAG embeding. I create a RAG database and set the Embeding model with Qwen3-4b-embeding. Then I upload a pdf file and start to process. But I found the embed model did not have any logs while dify process embeding data instead of the Qwen3-14 chat model was doing these embeding work. The embed model docker logs just show some health log. Its so strange about this.
From the model provider page, just the Qwen3-14B display in the first place.
2. Additional context or comments
No response
Beta Was this translation helpful? Give feedback.
All reactions