[GH-ISSUE #11387] issue: RAG broken on latest versions #54875

Closed
opened 2026-05-05 16:50:03 -05:00 by GiteaMirror · 0 comments
Owner

Originally created by @galvanoid on GitHub (Mar 7, 2025).
Original GitHub issue: https://github.com/open-webui/open-webui/issues/11387

Check Existing Issues

  • I have searched the existing issues and discussions.

Installation Method

Docker

Open WebUI Version

0.5.20

Ollama Version (if applicable)

latest

Operating System

Ubuntu 24.04

Browser (if applicable)

No response

Confirmation

  • I have read and followed all instructions in README.md.
  • I am using the latest version of both Open WebUI and Ollama.
  • I have checked the browser console logs.
  • I have checked the Docker container logs.
  • I have listed steps to reproduce the bug in detail.

Expected Behavior

Accurate RAG or at least correct

Actual Behavior

I've been using open webui since almost the first versions. My main use is RAG in the legal sector.

I've always had very good, even surprising, results in reasoning in my field, using legal documents.

But for some time now I haven't been getting those results. I've spent many days doing endless tests, different embedding models, all possible parameters, I've installed tika, increased the context to 128k, even manually deleted the vector_db (which by the way, sometimes doesn't get deleted from configuration/administrator/documents).

In desperation, I started up an old VM, before I switched to SSD, where I have version 0.5.4 installed.

To my surprise, with exactly the same parameters, same embedding model, same chunk size, top k, same model, same context size, etc., putting one version against another (0.5.4 against the latest 0.5.15-20) the results were completely disparate.

While version 0.5.4 nails the results of the questions about documents, the newer versions give very bad results, sometimes almost completely fabricated and, when it gets it right, it does so only partially.

So I have reinstalled version 0.5.4 on my SSD machine and I confirm that the results are excellent, as they were before. With exact answers in relation to the documents provided.

There seems to be something wrong with later versions, which works fine in 0.5.

Steps to Reproduce

  1. Open side by side open webui versions 0.5.4 vs 0.5.20 (or earlier up to 0.5.4)
  2. Configure same RAG options in both
  3. Configure same model (same params)
  4. Upload a document or collecion
  5. Ask same questions

Logs & Screenshots

No logs.

Additional Information

No response

Originally created by @galvanoid on GitHub (Mar 7, 2025). Original GitHub issue: https://github.com/open-webui/open-webui/issues/11387 ### Check Existing Issues - [x] I have searched the existing issues and discussions. ### Installation Method Docker ### Open WebUI Version 0.5.20 ### Ollama Version (if applicable) latest ### Operating System Ubuntu 24.04 ### Browser (if applicable) _No response_ ### Confirmation - [x] I have read and followed all instructions in `README.md`. - [x] I am using the latest version of **both** Open WebUI and Ollama. - [x] I have checked the browser console logs. - [x] I have checked the Docker container logs. - [x] I have listed steps to reproduce the bug in detail. ### Expected Behavior Accurate RAG or at least correct ### Actual Behavior I've been using open webui since almost the first versions. My main use is RAG in the legal sector. I've always had very good, even surprising, results in reasoning in my field, using legal documents. But for some time now I haven't been getting those results. I've spent many days doing endless tests, different embedding models, all possible parameters, I've installed tika, increased the context to 128k, even manually deleted the vector_db (which by the way, sometimes doesn't get deleted from configuration/administrator/documents). In desperation, I started up an old VM, before I switched to SSD, where I have version 0.5.4 installed. To my surprise, with exactly the same parameters, same embedding model, same chunk size, top k, same model, same context size, etc., putting one version against another (0.5.4 against the latest 0.5.15-20) the results were completely disparate. While version 0.5.4 nails the results of the questions about documents, the newer versions give very bad results, sometimes almost completely fabricated and, when it gets it right, it does so only partially. So I have reinstalled version 0.5.4 on my SSD machine and I confirm that the results are excellent, as they were before. With exact answers in relation to the documents provided. There seems to be something wrong with later versions, which works fine in 0.5. ### Steps to Reproduce 1. Open side by side open webui versions 0.5.4 vs 0.5.20 (or earlier up to 0.5.4) 2. Configure same RAG options in both 3. Configure same model (same params) 4. Upload a document or collecion 5. Ask same questions ### Logs & Screenshots No logs. ### Additional Information _No response_
GiteaMirror added the bug label 2026-05-05 16:50:03 -05:00
Sign in to join this conversation.
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: github-starred/open-webui#54875