Making a watered down BERT like application

[amoynahan]

I'm making a a BERT like add-in using C#. I think a handy way to build a C++ add-in in excel is to do it first in a language like R, Julia, or Python, or just doing algorithm in VBA. At some point I will put them on github. It looks like my R, and Julia bridges are working. I also want to have python before I put them on line.

[amoynahan]

I just posted it.

https://github.com/amoynahan/ExcelBridgeSuite

Current versions include:

R
Julia
Python

I’ll be adding some examples soon.

You can either build it yourself in Visual Studio or use the files in the Releases.

[HanMuYa]

About https://github.com/amoynahan/ExcelBridgeSuite, this is a very innovative plugin, thank you to the author.
Many thanks to the author for providing this plugin. A few months ago, I also created a similar plugin, but it used Python in the sidebar. Today, I happened to come across this plugin, and it is a very innovative plugin, similar to the PY functions provided officially, but using the local Python environment, which I think is a better choice. I wonder if the author has considered creating a sidebar toolbar? Also, I hope the author can write a more detailed user guide in their spare time!

[amoynahan]

Next step is to build out tests and worked examples, and add screenshots to the usage page. The goal is to show complete workflows (Excel inputs → computation → results/plots back in Excel).

I also plan to add a macro to update plots. Right now, plots have to be imported from the Add-ins menu. A simple “refresh plots” macro would make this much smoother.

Another area to expand is calling C++ directly from Excel through add-ins, exposing routines from libraries like Eigen, Armadillo, CGAL, and Boost.

For example:

=CPP_Cholesky(A1:D4) → returns the Cholesky factor
=CPP_SparseColMultiply(A1:A1000, B1:B1000, C1:C1000, D1:D1000) → sparse column multiplication
=CPP_WeibullSurvival(t, shape, scale) → survival function in C++

I’ve already implemented flexsurv-style functions in C++ using Boost, so this approach works and should be easy to extend.

I may also try implementing some numerical algorithms in VBA and validating them against the R, Julia, and Python bridges.

The goal is to make Excel a live front end for computation, rather than just a place to paste results.

[amoynahan]

I added some examples for the R version and would appreciate any feedback, especially on testing.

I was able to work through the examples myself, but I am not sure what issues or rough edges someone else might run into.

[HanMuYa]

I haven't tested it in R, but when I was using Python, I found several aspects of your architecture that are very unfriendly: First, you can't write long code (Excel function variables or strings exceeding 255 bytes will have problems); second, to accommodate various returns and variables, you provided too many functions (which I feel is not user-friendly); third, for two-way interactions, the current exchange method is only suitable for small datasets (it seems you are using inter-process communication (IPC), but once the data is slightly larger, it can easily freeze, making it even more difficult for large data). Similarly, I think these issues might also appear in R/Julia language versions of the plugin.

[Moynahan]

I’ll look into the issues you raised.

I wouldn’t pass large Python scripts as strings. For complex code, I use an RStudio project or VS Code directory alongside the add-in. Debugging through Excel would be painful.

I agree performance is too slow. My original idea was to use C++ add-ins for speed, with R/Julia/Python for validation.

Running
=REval("matrix(rnorm(10000 * 20), nrow = 10000, ncol = 20)")
takes ~8 seconds for a 10,000 × 20 matrix, which is too slow. It should be nearly instant, so I’ll investigate.

It may be similar to VBA: cell-by-cell writes are slow, but writing arrays is fast. I’ll check if that’s the bottleneck.

Thanks for the feedback.

[amoynahan]

Data transfer in RExcelBridge is too slow for larger datasets.

It works fine for small objects, but performance drops quickly as size increases, which limits practical use.

Things I’m considering:

• Reduce data copying between R and Excel
• Use a binary/columnar format instead of text/JSON
• Ensure bulk transfer (no cell-by-cell movement)
• Push heavy lifting into C/C++ under Excel-DNA
• For large data, write/read via files (feather/parquet)
• Consider a shared layer like DuckDB
• Use async execution to avoid blocking Excel

The current design works for function calls, but large data likely needs a different pattern (shared memory, file-based, or database-backed).

[amoynahan]

I made an update. It is pretty fast now. I need to update my usage page. The R bridge is the most tested, but I put the new functions in Julia and Python.

RExcelBridge Smoke Tests

=RPing()

=REval("x <- matrix(rnorm(10000*20), nrow=10000, ncol=20)")
=RGetNumeric("x")
=RLastTransfer()

Excel → R numeric

=RANDARRAY(10000,20)
=RSet("y", D10:W10009)
=RCall("dim", RObj("y"))

RCall auto numeric return

=REval("make_matrix <- function(n,p) matrix(rnorm(n*p), nrow=n, ncol=p)")
=RCall("make_matrix", 10000, 20)

Data frame (R → Excel)

=REval("df <- data.frame(id=c('A','B','C'), x=c(1.1,2.2,3.3), group=c('low','mid','high'), stringsAsFactors=FALSE)")
=RGetTable("df")

Excel → R table

=RSetTable("df2", D10:F13, TRUE)
=RCall("names", RObj("df2"))
=RGetTable("df2")

=RLastTransfer()

[malcook]

I used to love BERT and use it all the time. Since it stopped being supported I have moved entirely away from Windows and Excel and instead use LibreOffice Calc for spreadsheet work. Any chance of making it work with both Excel and Calc? That I could develop on linux and deliver to Windows/Excel users.

[amoynahan]

I don't know how to do it with LibreOffice. Microsoft made it so you could connect c# to excel. I didn't invent anything, I just used the bridge they made. I will be adding more examples to the repository. It's all free. If you don't want to use the add-in's that I compiled, you can install visual studio 2022 and build your own add-in.

If you make it work with Linux, I would be interested in how you did it.