A practical RAG setup question for OpenWebUI and PDF reasoning
The setup uses OpenWebUI s to handle several PDFs, such as plans, , case studies, and standards, and answer questions by connecting ideas across those files. Kreuzberg handles content , jina--v5-text-small is used as the , and jina-reranker-v3 is used as the .
The answer model is either the or a local model. The system works fairly well, but some questions need to be asked several ways before the answer catches details from the source documents.
The likely weak point may be chunking, meaning how the PDFs are split into smaller pieces before search. Chonkie is being considered as a semantic chunking option, where the split is based more on meaning than fixed size.
Key points
- The setup is for OpenWebUI s with multiple PDFs that need cross-document reasoning.
- It uses Kreuzberg for , Jina models for embedding and reranking, and or local Qwen for answers.
- The main problem is missed details from source documents, which forces repeated questioning.
- Chunking may be the cause because the right source details may not be grouped well enough.
- Semantic chunking with Chonkie is being considered as a possible improvement.