issue: v0.6.33 RAG 'Focused Retrieval' mode is ignored, always sending the entire document context when searching collections #6681

New Issue

2025-11-11T17:03:10-06:00

GiteaMirror commented

2025-11-11 17:03:10 -06:00

Originally created by @dcastell1 on GitHub (Oct 15, 2025).

Check Existing Issues

I have searched for any existing and/or related issues.
I have searched for any existing and/or related discussions.
I am using the latest version of Open WebUI.

Installation Method

Docker

Open WebUI Version

v0.6.33

Ollama Version (if applicable)

No response

Operating System

Debian GNU/Linux 12 (bookworm)

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 included the browser console logs.
I have included the Docker container logs.
I have provided every relevant configuration, setting, and environment variable used in my setup.
I have clearly listed every relevant configuration, custom setting, environment variable, and command-line option that influences my setup (such as Docker Compose overrides, .env values, browser settings, authentication configurations, etc).
I have documented step-by-step reproduction instructions that are precise, sequential, and leave nothing to interpretation. My steps:
Start with the initial platform/version/OS and dependencies used,
Specify exact install/launch/configure commands,
List URLs visited, user input (incl. example values/emails/passwords if needed),
Describe all options and toggles enabled or changed,
Include any files or environmental changes,
Identify the expected and actual result at each stage,
Ensure any reasonably skilled user can follow and hit the same issue.

Expected Behavior

When a knowledge base collection is selected using the # command and the retrieval mode is set to "Using Focused Retrieval", the system is expected to perform a targeted, relevance-based search against the collection's content and retrieve only the top-k most relevant chunks, where k is the value configured in the Admin Settings.
It must inject only these selected chunks—not the entire file content—into the context sent to the model for generating a response.

Actual Behavior

Currently, the retrieval mode toggle is non-functional. Regardless of whether "Using Focused Retrieval" or "Using Entire Document" is selected, the system defaults to the "Entire Document" behavior.

The application consistently sends the full content of all files within the selected knowledge base collection to the model's context. This happens even when "Focused Retrieval" is explicitly enabled.

As a result:

There is no functional difference between the two retrieval modes.
Attempting to query large documents almost always results in a context overflow error, as the model's context window limit is exceeded.
The core RAG feature for querying specific information from large files is unusable. The top k setting in the admin panel has no effect, as the semantic search and chunking mechanism is never triggered.

This issue was non-existent in previous version v0.6.32

Steps to Reproduce

Start with an existing Kubernetes (AKS) deployment of Open WebUI using the ghcr.io/open-webui/open-webui:main image. Configure PostgreSQL as DB and PGVector as vector store.
Log in as an administrator and navigate to the Admin Settings -> Documents section.
Configure the RAG settings as follows: Content Extraction Engine: Set to a valid Azure AI Services endpoint. (Document Intelligence), Embedding Model Engine: Set to Azure OpenAI, using the text-embedding-3-small model and a valid Azure OpenAI endpoint. All other fields have default values.
Navigate to the "Workspace>knowledge" page from the main user menu.
Create a new collection named TEST
Upload approximately 10 multi-page PDF documents into the TEST collection and wait for the embedding process to complete
Start a new chat session with any configured model
In the chat input box, type # and select the TEST collection.
Before sending a message, click on the retrieval mode box that appears above the chat input and ensure that "Using Focused Retrieval" is selected.
Ask a specific question that requires knowledge from the uploaded documents.
Observe the actual result: The system either returns a context overflow error from the model API or, if it responds, the "sources" listed include all 10 documents. Clicking any citation shows that the entire file was used as context, not just the top 3 relevant chunks, confirming that the "Focused Retrieval" mode was ignored.

Logs & Screenshots

aks_openwebui_logs.txt

Additional Information

No response

Originally created by @dcastell1 on GitHub (Oct 15, 2025). ### Check Existing Issues - [x] I have searched for any existing and/or related issues. - [x] I have searched for any existing and/or related discussions. - [x] I am using the latest version of Open WebUI. ### Installation Method Docker ### Open WebUI Version v0.6.33 ### Ollama Version (if applicable) _No response_ ### Operating System Debian GNU/Linux 12 (bookworm) ### 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 included the browser console logs. - [x] I have included the Docker container logs. - [x] I have **provided every relevant configuration, setting, and environment variable used in my setup.** - [x] I have clearly **listed every relevant configuration, custom setting, environment variable, and command-line option that influences my setup** (such as Docker Compose overrides, .env values, browser settings, authentication configurations, etc). - [x] I have documented **step-by-step reproduction instructions that are precise, sequential, and leave nothing to interpretation**. My steps: - Start with the initial platform/version/OS and dependencies used, - Specify exact install/launch/configure commands, - List URLs visited, user input (incl. example values/emails/passwords if needed), - Describe all options and toggles enabled or changed, - Include any files or environmental changes, - Identify the expected and actual result at each stage, - Ensure any reasonably skilled user can follow and hit the same issue. ### Expected Behavior When a knowledge base collection is selected using the # command and the retrieval mode is set to "Using Focused Retrieval", the system is expected to perform a targeted, relevance-based search against the collection's content and retrieve only the top-k most relevant chunks, where k is the value configured in the Admin Settings. It must inject only these selected chunks—not the entire file content—into the context sent to the model for generating a response. ### Actual Behavior Currently, the retrieval mode toggle is non-functional. Regardless of whether "Using Focused Retrieval" or "Using Entire Document" is selected, the system defaults to the "Entire Document" behavior. The application consistently sends the full content of all files within the selected knowledge base collection to the model's context. This happens even when "Focused Retrieval" is explicitly enabled. As a result: 1. There is no functional difference between the two retrieval modes. 2. Attempting to query large documents almost always results in a context overflow error, as the model's context window limit is exceeded. 3. The core RAG feature for querying specific information from large files is unusable. The top k setting in the admin panel has no effect, as the semantic search and chunking mechanism is never triggered. This issue was non-existent in previous version v0.6.32 ### Steps to Reproduce 1. Start with an existing Kubernetes (AKS) deployment of Open WebUI using the ghcr.io/open-webui/open-webui:main image. Configure PostgreSQL as DB and PGVector as vector store. 2. Log in as an administrator and navigate to the Admin Settings -> Documents section. 3. Configure the RAG settings as follows: Content Extraction Engine: Set to a valid Azure AI Services endpoint. (Document Intelligence), Embedding Model Engine: Set to Azure OpenAI, using the text-embedding-3-small model and a valid Azure OpenAI endpoint. All other fields have default values. 4. Navigate to the "Workspace>knowledge" page from the main user menu. 5. Create a new collection named TEST 6. Upload approximately 10 multi-page PDF documents into the TEST collection and wait for the embedding process to complete 7. Start a new chat session with any configured model 8. In the chat input box, type # and select the TEST collection. 9. Before sending a message, click on the retrieval mode box that appears above the chat input and ensure that "Using Focused Retrieval" is selected. 10. Ask a specific question that requires knowledge from the uploaded documents. 11. Observe the actual result: The system either returns a context overflow error from the model API or, if it responds, the "sources" listed include all 10 documents. Clicking any citation shows that the entire file was used as context, not just the top 3 relevant chunks, confirming that the "Focused Retrieval" mode was ignored. ### Logs & Screenshots [aks_openwebui_logs.txt](https://github.com/user-attachments/files/22928502/aks_openwebui_logs.txt) <img width="1309" height="424" alt="Image" src="https://github.com/user-attachments/assets/7cf04642-a69a-40a7-bc42-01ef6c048310" /> ### Additional Information _No response_

GiteaMirror added the bug label 2025-11-11 17:03:10 -06:00

GiteaMirror closed this issue

2025-11-11 17:03:10 -06:00

GiteaMirror commented

2025-11-11 17:03:12 -06:00