Want to contribute performance optimizations

[misrasaurabh1]

Hi @pietermarsman, I am Saurabh and we are optimizing Unstructured.io who heavily uses pdfminer.six for pdf parsing. We've noticed that the primary bottleneck there is pdf parsing through this underlying library so we wanted to optimize this library for performance further.

We have created a code optimization agent codeflash.ai that automatically creates the most impactful optimizations for real workflows that are easy to review and merge.

I wanted to ask if we could collaborate with you or the contributors to discover optimizations for this library with Codeflash and create PRs for it?

[pietermarsman]

Hi @misrasaurabh1,

I'm open to optimization PRs as long as they:

1. Address a specific issue (create one first describing the performance problem)
2. Target actual bottlenecks (include profiling data)
3. Show measurable improvements (before/after benchmarks)

Please use the PDFs in samples/ for benchmarking so results are reproducible.

Review times may vary, but I'll get to it eventually.

[pietermarsman]

Do you have a first optimization in mind?

[misrasaurabh1]

Thank you @pietermarsman, I will follow these guidelines. Planning to collaborate with @CyMule soon to find the optimizations.

[pietermarsman]

👍 Op zo 11 jan 2026, 19:08 schreef Saurabh Misra ***@***.***>:

…

*misrasaurabh1* left a comment (pdfminer/pdfminer.six#1236) <#1236 (comment)> Thank you @pietermarsman <https://github.com/pietermarsman>, I will follow these guidelines. Planning to collaborate with @CyMule <https://github.com/CyMule> soon to find the optimizations. — Reply to this email directly, view it on GitHub <#1236 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AA5H44EJWCQZIJXRM2IJUX34GKGTDAVCNFSM6AAAAACQ7UZLDGVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZTOMZVGIZDQMJQGY> . You are receiving this because you were mentioned.Message ID: ***@***.***>

[misrasaurabh1]

Hi @pietermarsman , I found a lot of high quality and important optimizations on my fork https://github.com/codeflash-ai/pdfminer.six/pulls
Could you take a look and pick your top-5 most important optimizations? I will open them as PRs here. Feel free to leave comments there itself.

[dhdaines]

Hi @pietermarsman , I found a lot of high quality and important optimizations on my fork https://github.com/codeflash-ai/pdfminer.six/pulls Could you take a look and pick your top-5 most important optimizations? I will open them as PRs here. Feel free to leave comments there itself.

CodeFlash is pretty neat, but the issue I see here is that:

1. Everything is a micro-optimization to one specific method, without considering the possibility of higher-level algorithmic fixes
2. The PRs are very verbose as one might expect from LLM-generated text

Actually (1) is okay, because we probably don't want to risk changing the way pdfminer.six operates (if you don't care about compatibility you can use something else like PLAYA-PDF which is up to 10x faster ;-) )

It may be more useful to group them together at the class or module level to make them more digestible to maintainers with limited time to look at them.

[dhdaines]

I am taking a look at these now anyway, they are interesting!

[dhdaines]

I've done some annotation of certain specific pull requests. I did copy and merge one of them, as it was necessary to fix a (rather serious) bug revealed by fuzzing.

Most of them are good, common-sense changes, specifically things like not concatenating or using slices of bytes when you could use bytearray instead.

I assume that, because the output of an LLM cannot be copyrighted according to US law (and probably other countries) that these changes are in the public domain, but it might be good to verify this before merging any of them.

[misrasaurabh1]

Thanks for taking a look @dhdaines, there can be several micro-optimizations since we ask codeflash to go function-by-function when it is optimizing, without knowing the workloads. Codeflash also supports optimizing workloads end to end through tracing that gives much better results.

And yes we don't want to change the architecture, only try to improve the implementation.

Feel free to merge the changes, we are also happy to collaborate with you help you verify, but you can also feel free to verify or merge them yourselves. We just want to enable a faster pdfminer!