Concurrent model chat, pollutes context between the sessions. #3323

New Issue

GiteaMirror · 2025-11-11T15:29:06-06:00

GiteaMirror commented

2025-11-11 15:29:06 -06:00

Originally created by @marked23 on GitHub (Jan 19, 2025).

Bug Report

I noticed that when I use the concurrent model feature, the context from one session, pollutes the context of the other sessions. I'm not sure if this is a UI thing or something in ollama.

Installation Method

I used git clone, and I do the same with ollama.
I run both open-webui and ollama as systemd services.

Environment

Open WebUI Version: 0.5.2 (Is this exposed in the UI somewhere?)
Ollama (if applicable): 0.4.5
Operating System: Kubuntu 24.10
Browser (if applicable): Chrome 132 on Windows 10.

Confirmation:

I have read and followed all the instructions provided in the README.md.
I am on the latest version of both Open WebUI and Ollama.
I have included the browser console logs.
I have included the Docker container logs.
I have provided the exact steps to reproduce the bug in the "Steps to Reproduce" section below.

Expected Behavior:

I expected that the session context for each model would remain independent.

Actual Behavior:

I observed that all three model sessions repeated the same unique word.

Description

I asked all three models to produce a random word.
I asked all three models to repeat their random word.
All three models repeated the random word chosen by the third model.

Bug Summary:

Reproduction Details

Steps to Reproduce:

Enable three models to concurrently respond to prompts.
Prompt the models to generate a random word.
Prompt the models to repeat their random word.

Note that the models all repeat the same word.

Logs and Screenshots

Screenshots/Screen Recordings (if applicable):

Originally created by @marked23 on GitHub (Jan 19, 2025). # Bug Report I noticed that when I use the concurrent model feature, the context from one session, pollutes the context of the other sessions. I'm not sure if this is a UI thing or something in ollama. --- ## Installation Method I used git clone, and I do the same with ollama. I run both open-webui and ollama as systemd services. ## Environment - **Open WebUI Version:** 0.5.2 (Is this exposed in the UI somewhere?) - **Ollama (if applicable):** 0.4.5 - **Operating System:** Kubuntu 24.10 - **Browser (if applicable):** Chrome 132 on Windows 10. **Confirmation:** - [x] I have read and followed all the instructions provided in the README.md. - [ ] I am on the latest version of both Open WebUI and Ollama. - [ ] I have included the browser console logs. - [ ] I have included the Docker container logs. - [x] I have provided the exact steps to reproduce the bug in the "Steps to Reproduce" section below. ## Expected Behavior: I expected that the session context for each model would remain independent. ## Actual Behavior: I observed that all three model sessions repeated the same unique word. ## Description I asked all three models to produce a random word. I asked all three models to repeat their random word. All three models repeated the random word chosen by the third model. **Bug Summary:** ## Reproduction Details **Steps to Reproduce:** Enable three models to concurrently respond to prompts. Prompt the models to generate a random word. Prompt the models to repeat their random word. Note that the models all repeat the same word. ## Logs and Screenshots **Screenshots/Screen Recordings (if applicable):** ![Image](https://github.com/user-attachments/assets/d085b1e8-f0dd-4667-ac63-1643ddf5a4ef)

GiteaMirror closed this issue

2025-11-11 15:29:06 -06:00

GiteaMirror commented

2025-11-11 15:29:06 -06:00

@tjbck commented on GitHub (Jan 19, 2025):

Intended behaviour here, you choose the response from the model and that's what gets fed into as an answer in the next turn!

@tjbck commented on GitHub (Jan 19, 2025): Intended behaviour here, you *choose* the response from the model and that's what gets fed into as an answer in the next turn!

GiteaMirror commented

2025-11-11 15:29:06 -06:00

@marked23 commented on GitHub (Jan 20, 2025):

I misunderstood the feature. Although the way I thought it worked still sounds interesting.
Thank you.

@marked23 commented on GitHub (Jan 20, 2025): I misunderstood the feature. Although the way I _thought_ it worked still sounds interesting. Thank you.