If the Coze knowledge base uploads a document with 'code=105000004' and prompts such as inconsistent vector dimensions and 'dense_text_content', it is usually not that the document itself is broken, but that the embedding configuration is not correct. In public issues, this type of error appears repeatedly, and community feedback basically points in the same direction: the dimension of generating vectors is not the same as the already established collection dimension.
If you're looking at the open source version of Coze Studio, the official repository is https://github.com/coze-dev/coze-studio. The official README emphasizes that the model and basic component configuration should be completed first, and then the agent, workflow, and knowledge base should be done, otherwise many problems will be concentrated later.
What exactly is this error reporting?
Simply put, the system expects you to pass in 2048 dimensions, but you actually send in 1024 dimensions, or vice versa. For vector libraries, this is not a "small deviation", but a format mismatch that cannot be written directly. The problem is that one layer of the embedding model, request parameters, old collection, and environment variables is not unified.
The three most common causes
- You changed the embedding model, but the underlying collection of the knowledge base is still the old dimension.
- You changed '.env', but the container that actually ran didn't get the new configuration.
- The actual dimension returned by the model is different from what you think, causing the library to fail later.
How it is usually handled in the community
The most direct way is to unify the embedding dimension first, and then redo the document processing. If the knowledge base has already been built, many times it is not only necessary to change the configuration, but also to rebuild the collection or re-import the documentation. Someone in the open issue has changed the model of 'bge-m3', 'm3e-base', etc., but if the underlying dimensions are not synchronized, the problem will continue to occur.
When troubleshooting, it is recommended to look at three things first: the embedding model in the model configuration page, the dimension parameter in '.env', and the definition of the current collection in the knowledge base. As long as one of these three places is not aligned, it will continue to report similar errors.
One sentence conclusion
The vector dimension of the Coze knowledge base is errored, which is essentially a misalignment between the "model output" and the "library definition". Unifying the embedding configuration before considering rebuilding the knowledge base is usually the most time-saving.
