Overview

Following the above,RagflowPlusAfter a round of emergency updates, 6 major issues were fixed and a new function for resetting user passwords was added.

Problem Solving Overview

1. Picture does not display problem

In some cases, the image cannot be displayed correctly and becomes a string as shown in the following figure.

After investigation, it was found that the problem was in adding access permissions for the image.

To fix this problem, modify the permission settings of the minio image file and set public access permissions for all images in the image storage folder.

"Resource" : [f "arn:aws:s3:::{kb_id}/images/*" ]

2. Embedded model connection error problem

This situation occurs when detecting the embedded model status in the background knowledge base management menu module.

There are three kinds of errors here:

1. Model name query error: The chat model information will be filled in during the query. The solution is to add a filter to the model type information in the talent_llm table.

2. Before the model configuration information is updated incorrectly, the earliest model configuration information in the talent_llm table will be directly queried. However, when the initial user changes the model configuration, the information does not change.

The solution is to query the latest embedding model configuration of the original user to ensure that it is updated synchronously with the front end.

3. Non-vllm model configuration information error due to the vllm mirror connection interface carrying the last
v1This is incompatible with other standard interfaces, resulting in the front-end configuration passing but the back-end configuration failing when using other models.

The solution is to robustly fill the model API address during background verification to adapt to the interfaces of different models.

Here is an explanation: Why do I disable user editing during background testing? The main reason is that I want to reuse the existing MySQL data table.

If the initial user embedding model of the backend and foreground is consistent, the same information of the data table can be read. If users are allowed to modify in the backend, especially when changing the embedding model, the foreground and background information will conflict, which will affect the query effect.

Therefore, users are only allowed to make changes in the foreground to avoid inconsistencies before and after the changes.

3. 413 error message when uploading files > 1M

When uploading a file > 1M, the following problem will occur.

The reason is that nginx has a default file size limit of 1M.

In version v0.2.1, the nginx file is processed as an external mount, which can be modifieddocker\nginx\management_nginx.confIn the fileclient_max_body_sizeThe parameter dynamically modifies the maximum file limit, the default is 500M.

4. Parsing block backtracking does not correspond

In the front-end knowledge base, you can see the original position of the specific parsing block in the PDF file, but in some cases, there is a misalignment problem.

In version v0.2.1, the matching logic has been readjusted to solve this problem, and the insertion of table parsing blocks is further supported. The effect is shown in the following figure:

5. Display the reason for parsing failure

In the background knowledge base, when parsing a file, if an error occurs, the error message shown in the figure below will be displayed.

This problem is caused by the need to install LibreOffice first when deploying source code, so the error message is fixed to this value. In the docker image, the LibreOffice environment is built-in, so this problem will not occur. The actual error cause is presented in the background log information.

Version v0.2.1 optimizes the display of error information and will feed back the actual error cause of parsing failure to the front end for display.

6. The parsing block shows 0 problem

In some cases, such as uploading txt or excel files, parsing will proceed normally, but the parsing block is 0.

The reason for this problem is that minerU currently only supports three types of files: word, ppt, and pdf. It does not perform abnormal judgment and timely error feedback for other types of files. Version v0.2.1 fixes this problem.

New feature: User password reset

Readers reported this problem: users forget their passwords after not using the app for a long time. In the database, passwords are stored in encrypted form, and administrators cannot see the user's plain text passwords.

Therefore, in version v0.2.1, the background user management module adds the function of resetting user passwords. The user password can be reset and updated through the button shown in the figure below.

Other improvements: The front-end system of RagflowPlus (v0.2.1) is based onv0.17.2-slimPackaging, remove the built-in model, so the image package size will be much smaller.

It is not recommended to use the built-in model because when using the GPU, the video memory usage and graphics card distribution of the built-in model cannot be effectively controlled.