Fava Extension for Natural Language Queries

[Patrick Ruckstuhl]

Hi all,

now that I have an ollama model that translates natural queries into bql
(https://ollama.com/tarioch/qwen2.5-coder-bql, did some more
improvements like dropping the requirement for a specific system prompt
and by default is a 3b model), I created a small fava extension that
uses it to allow you to ask questions in natural language and get the
results back


you can find the details here

https://github.com/tarioch/fava-nl2bql


It's released on pypi, so should be simple to add. You'll need an ollama
instance running that you can point towards for doing the translation.


Let me know if you encounter issues, have suggestions or general feedback.


Regards,

Patrick

[senorsmile]

This is great.  How well would this work if I used an existing locally hosted model.  For instance, I already have qwen 3.6 35b running.  I couldn't really run another model on that machine.  Is there a minimal template that could be used with a larger model without needing post-training (like your qwen2.5-coder-bql) that could provide similar results?  It would be great to simply reuse the model I already have.  

I as wrote this, I followed your links through to https://github.com/tarioch/bql-lora.  I suppose I could use this to train the qwen 3.6 I'm already using to create a bql specific version?

[Patrick Ruckstuhl]

Hi,

I'm not sure how well it will work. Potentially by training it for this, you might get worse results for other things, the advantage of this very specialized model is that it can really focus on only doing this. For sure you will want to add some trigger tokens in front of the question to make sure it's only picked up when you want it.

What I have done on my model is in the meantime to switch to the 3B base model (there are even smaller ones that I haven't tried yet), still works really well and uses less memory, so it might fit in. In my case I most of the time don't need to have multiple models loaded at the same time, so it just loads whichever model is needed at the time.

Let me know if you want to try to train your model. My hardware will not support training this large model but maybe I can provide you some guidance.

Regards,

Patrick

--
You received this message because you are subscribed to the Google Groups "Beancount" group.
To unsubscribe from this group and stop receiving emails from it, send an email to beancount+...@googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/beancount/a07947ea-4631-43db-9da1-925216007e15n%40googlegroups.com.