perf: optimze figure parser

[liuzhenghua]

What problem does this PR solve?

When parsing documents containing images, the current code uses a single-threaded approach to call the VL model, resulting in extremely slow parsing speed (e.g., parsing a Word document with dozens of images takes over 20 minutes).

By switching to a multithreaded approach to call the VL model, the parsing speed can be improved to an acceptable level.

Type of change

• Performance Improvement

[asiroliu]

@liuzhenghua
Could you provide a test Word document?

[liuzhenghua]

@liuzhenghua Could you provide a test Word document?

@asiroliu It’s a simple Word document serving as an operation manual, containing around 90 screenshots. Sorry, I can’t provide the file as it contains sensitive company data.

[asiroliu]

@liuzhenghua
Thank you for your response. Based on your suggestion, i can use the following search terms on Google to find the relevant documentation:

filetype:doc OR filetype:docx "User Manual" OR "Instruction Manual" OR "Operation Guide"

[liuzhenghua]

@liuzhenghua Thank you for your response. Based on your suggestion, i can use the following search terms on Google to find the relevant documentation:

filetype:doc OR filetype:docx "User Manual" OR "Instruction Manual" OR "Operation Guide"

@asiroliu Sorry for my previous response — the document I mentioned belongs to the company and can't be shared. You'll need to create a Microsoft Word document yourself, including some text and around 90+ images.

[asiroliu]

@liuzhenghua
I've compared the multi-image document parsing performance between the nightly build and your latest commit. There doesn't appear to be any noticeable efficiency improvement.

• Test word docment: https://irtfweb.ifa.hawaii.edu/~tcs3/oldstuff/osprey/userman.doc

• nightly(78b00d61fd59)(2025_04_29): 207 secs

• you lastest commit(3281d47): 215 secs

[liuzhenghua]

• https://irtfweb.ifa.hawaii.edu/~tcs3/oldstuff/osprey/userman.doc

@asiroliu
The test document you used didn't contain any images. The one I tested had around 90+ images. Before the optimization, it took 20 minutes to parse the images and another 20 minutes to upload them to MinIO. After the changes, both steps now only take 2 minutes each.

[liuzhenghua]

@asiroliu My local version is 0.17.2. When the log message "Visual model detected. Attempting to enhance figure extraction" appears, I debugged and found that it processes the 90 images in the document by calling the VL model one by one in a single queue, which leads to a long processing time.

You observed a similar processing time in your test, but that might be due to one or more of the following reasons:

• Your document contains very few images.
• You haven't configured a VL model.
• Version 0.18.0 doesn't have this issue.

[asiroliu]

Got it, I'll verify this later per your suggestions.