Open WebUI 0.11.4 shrinks slim image to about 175 MB; RAG and voice require external services

Open WebUI released version 0.11.4 on September 21. According to the project, the rebuilt slim container image is approximately 175 MB, about 89% smaller than the previous version. The release removes the local model runtime and related packages, reducing the deployment burden for a chat frontend that only needs to connect to model services. Existing users, however, must recheck their feature dependencies. [Release notes](https://github.com/open-webui/open-webui/releases/tag/v0.11.4)

The new slim image no longer includes PyTorch, Transformers, local Whisper, or certain document processing tools. It can still start with the default SQLite database and local file storage, and chat functionality continues to connect to model providers as usual. Deployments using MySQL or MariaDB for the application database, or object storage such as S3 for files, must continue using the standard image. [Image documentation](https://docs.openwebui.com/getting-started/quick-start/)

RAG limitations are more likely to surface only after an upgrade: the slim image ships only with the pgvector client for vector retrieval, and embeddings must come from Ollama, OpenAI, or Azure OpenAI. Without the corresponding services configured, the container may start successfully but fail on the first retrieval request. PDF and Office documents also require an external extractor, while voice input and output require separate engines. These services can run on an enterprise's internal network. [Feature dependencies](https://docs.openwebui.com/getting-started/quick-start/#what-slim-leaves-out)

The release also removes local reranking models. When no external reranker is configured, results fall back to scoring based on cosine similarity between embedding vectors. As a result, even if queries still return content, ranking quality may differ from before the upgrade. [Reranking changes](https://github.com/open-webui/open-webui/releases/tag/v0.11.4)

Migration also involves data compatibility. If the change requires switching embedding models, existing knowledge bases must be reindexed because vector spaces from different models cannot be mixed. The official documentation also states that reindexing uses previously extracted text and does not parse the original files again. Attachments uploaded directly to chats without being added to a knowledge base are outside the scope of reindexing and must be uploaded again after switching models. [RAG migration guidance](https://docs.openwebui.com/features/chat-conversations/rag/)

From a deployment perspective, a smaller image primarily reduces transfer and storage requirements; it does not directly imply faster model inference. Engineering teams should first inventory their document formats, embedding models, and storage backends, then use existing Chinese question-and-answer cases to test retrieval quality and error reporting. Validation should cover the initial document import, restarts, and temporary loss of connectivity to the embedding service, ensuring that health checks go beyond checking the chat landing page. Deployment teams also need to maintain a list of service versions and dependencies to prevent an update to one component from breaking the entire processing pipeline. The cost of the complete workflow still includes external services, network round trips, and reindexing. These should all be measured to determine whether switching is worthwhile.