Differencence in performances: Python vs Go codebase

[JPFrancoia]

Hi all,

I just have a quick question. I'm running Vectorcode and ChromaDB locally (chroma DB in docker). My CPU is a AMD Ryzen 5 5600X 6-Core Processor. I have 48GB of RAM. I'm using neovim + codecompanion. First, is my CPU good enough to run vectorcode well enough?

I have vectorized a Python codebase, and I'm asking questions to codecompanion using the vectorcode tool (sonnet-4 as the LLM). While not super speedy fast, it generally responds within 10s.
I have tried the same thing with a Go codebase, with much less success. I tried asking the same questions for the Go codebase, and codecompanion was hanging forever (I stopped waiting after a minute).

Any idea why such a discrepancy?

[Davidyz]

My current dev machine runs a 5600 (non-x) with 32GB RAM, so the hardware are certainly not the issue. Could you try making some queries from the command line with logging enabled, so that we can get some more insight of what's happening? Also are the sizes of the python and go repos comparable (in terms of LoC)?

[JPFrancoia]

Ah 🤦‍♂️ I hadn't thought about that.

For the Go codebase:

-----------------------------------------------------------------------------------------------------------------------------------
File                                                                                            blank        comment           code
-----------------------------------------------------------------------------------------------------------------------------------
-----------------------------------------------------------------------------------------------------------------------------------
SUM:                                                                                             1725           1007           8156
-----------------------------------------------------------------------------------------------------------------------------------

And for the Python codebase:

----------------------------------------------------------------------------------
File                                blank        comment           code
----------------------------------------------------------------------------------
----------------------------------------------------------------------------------
SUM:                                 1665           1534           3506
----------------------------------------------------------------------------------

But in my Go codebase I have a lot of auto-generated code (output of protobufs). If I exclude the directory that contains the generated Go code for the protos:

-------------------------------------------------------------------------------------------------
File                                                          blank        comment           code
-------------------------------------------------------------------------------------------------
-------------------------------------------------------------------------------------------------
SUM:                                                           1020            548           3729
-------------------------------------------------------------------------------------------------

The size is comparable to the Python code base. And vectorcode is not hanging anymore 🎉

Thank you very much for your help!

[Davidyz]

While selectively vectorising your codebase is definitely the best option there, there are some config options that might improve the query speed. I wrote the initial versions of this project on a potato, so I spent quite some time playing around with those. With 0.7.0 introducing the vectorisation ability to the codecompanion extension (and the MCP server, too), you could try letting the LLM decide what files to vectorise (There's a demo in the readme showing how I let the LLM handle the vectorisation process before asking questions about the codebase). If the auto-generated code follows a conventional filesystem hierarchy, you might be able to get the LLM to exclude those from the vectorisation tool call (I haven't tried this myself, but if this worked for you, please lmk).