It's kind of antithetical to the tool's deterministic positioning, but have you considered making TERMy leverage an LLM for unseen or low-confidence queries, and then generate the config and update itself to make future similar queries deterministic?
This is such a nice idea! I could add a fallback towards LLMs, it was present but I removed it. Would you be interested to help me implement the auto-update? I must admit, the LLMs are very useful for this kind of work. I think that TERMY's design is now feasible BECAUSE OF the availability of LLMs. They make the dataset development feasible.
It would make sense to have this integrate with a self-learning routine for an agent: e.g. at night it looks through what it did and writes NPC-Forge recipes. Tomorrow it can answer queries (which he turned added to NPC-Forge) without an LLM. Of course this implies a branching where a query is either processed by NPC-Forge or an LLM, depending on some measure of confidence that NPC-Forge can answer it well.
I'll be the first to admit I burn too many tokens on the command line. Some form of LLM harness has been my new shell for the past 3+ years, they are much better at *sh then me, and I prefer not context switching from my terminal when I don't have to. In a way the rise of agent harnesses and their shell expertise has opened up apis and command line tools which have always been pain points for a terminal user.
Great, simple idea applied in a novel way. Peak engineering if I ever seen it and I don't even have to look at the code. Nice work!
I haven't evaluated it yet, but I love the fact that the output is (at least claimed to be) deterministic. I can't trust an LLM to do the right thing after I deploy it to production, because their output is non-deterministic by design.
TERMy (or is it the NPC-forge) seems to be worth a try.
You can get determinostic output (mostly) by setting the temperature to zero. Using couple of other tricks you can get close to 100% of determinism with LLMs.
That's reproducible, I wouldn't call it deterministic. Small, semantically meaningless changes in the input can still result in wildly different output.
That's the definition of a chaotic system (small change in initial conditions results in large, seemingly -- but not actually -- random changes in output), but it's still deterministic (same input results in same output).
> because their output is non-deterministic by design.
It isn't. At least not by design, even though in practice it often can be. If you do greedy decoding (or use a preset seed) and deterministically compute everything (e.g. only use integer math) then it will be 100% always deterministic.
That’s true, but not true-true. Sure, every time you prompt “what is the weather in kansas” you’ll get the same output, but if you prompt “what is the weather in kansas right now” you’ll get a different output, and then “what is the weather in kansas today” gets a different output. Language being language, there are infinite ways to say things, so there are infinite variations in what the llm can output in response to very similar prompts.
This tool has a finite amount of outputs for an infinite amount of inputs. Which is different from an llm based tool.
Yes i think I acknowledged that, but is that useful for making a tool that can be trusted to safely run shell commands when asked arbitrary questions? No. It’s not.
Cool, but system and user should probably stick to short, clear commands. E.g., I see you do some anaphora resolution (in particular: find what "it" refers to), but in a complex dialog, the human intention can differ from the machine's understanding. That will give problems when you end your dialog with "delete it".
Adding more sentences to your data set will slowly degrade performance. It's a delicate system.
Source: I have written software with similar functionality (NLP search) in SaaS form, a long time ago. It required quite a bit of work to configure.
> Models like ornith:9b, mistral:7b or cogito:14b can get the job done sometimes, but they are not fast and reliable enough for general use, specially if you have only 4GB of VRAM.
Have you considered/tried using a model that's, well, more appropriate size-wise for an use case like this? These are relatively big. Something like FunctionGemma [1] finetuned for a given set of tasks would be a lot more speedy.
FunctionGemma never worked well for me (without fine tuning). Liquid has released 230M and 350M models that work far, far better in my testing: https://huggingface.co/LiquidAI/LFM2.5-230M
I really look forward to a hypothetical LFM3-230M, because LFM2.5-230M is so close to being usable, while FunctionGemma is miles away from being usable.
I tried functiongemma, it is for sure faster than those models, the problem is that is not reliable enough for a terminal assistant. I would say that no LLM is good for a terminal assistant, if you take into account the operational cost and the risk of damage. Even if it fails only 1 time out of 10 becomes useless. That's why I developed FlintParser!
Warp uses LLMs so it is slow and prone to hallucination. Using very colloquial terms TERMy is more or less a calculator that knows english :) so it can run on your CPU and respond instantly! The difference is that it can only answer predetermined responses (with optional arguments) this makes it useless if you need to generate text, but makes it safe and predictable for a use case like a terminal assistant.
is it supported to have it propose a command for approval rather than running autmatically? in the YT video it looks likw it ran the cpu temp command on its own
Yep he runs on its own when the command is non-destructive, like checking the CPU temperature, it does ask for permission if the command is potentially destructive. I agree it is so cool, it looks sci-fi :)
Very interesting project, I like it. Just wanted to clarify though the sentiment analysis is just the count of stripped words and used to tag things with the emoji? I was initially expecting it to be a part of the actual command construction process (even though I couldn't figure out how that would be relevant) given how it was listed.
Ciao, yes for now the sentiment analysis is used only to provide an emoji related to the response. In the future I would like also to influence the choices of adjectives and interjections according to the sentiment.
I like the project. I see a lot of potential integrating it with LLM providers in an effort to lower token usage for repetitive tasks. Your solution becomes the "main model" and the LLM the fallback
what I would love to see along those lines is something that can answer "what packages do I have installed to do task $foo"; I keep installing things that I use for one thing and then forget about when I need to do the same task some months or years later.
Yes tldr, navi cheatsheets are essentially a great set of how to do X using Y for terminal utilities. Pretty well curated set of commands for majority of the tasks, so are great training material.
And great work! Love it
One significant advantage of not using a local LLM is the significantly simplified dependency stack.
paper: https://arxiv.org/abs/1802.08979
WOW! With that dataset the capabilities of TERMy could be vastly extended!
Thank you.
Great, simple idea applied in a novel way. Peak engineering if I ever seen it and I don't even have to look at the code. Nice work!
What do you think about it?
TERMy (or is it the NPC-forge) seems to be worth a try.
It isn't. At least not by design, even though in practice it often can be. If you do greedy decoding (or use a preset seed) and deterministically compute everything (e.g. only use integer math) then it will be 100% always deterministic.
This tool has a finite amount of outputs for an infinite amount of inputs. Which is different from an llm based tool.
Adding more sentences to your data set will slowly degrade performance. It's a delicate system.
Source: I have written software with similar functionality (NLP search) in SaaS form, a long time ago. It required quite a bit of work to configure.
Have you considered/tried using a model that's, well, more appropriate size-wise for an use case like this? These are relatively big. Something like FunctionGemma [1] finetuned for a given set of tasks would be a lot more speedy.
[1] https://blog.google/innovation-and-ai/technology/developers-...
I really look forward to a hypothetical LFM3-230M, because LFM2.5-230M is so close to being usable, while FunctionGemma is miles away from being usable.
But, yes, still tangential to TERMy.
I hope the community will help me to enhance it :) it is just a proof of concept for now
love the idea/simplicity of this tool!
I really like it, this flavor of specialization gives the user a win on privacy and speed. Seems like the right idea for such a tool.
For now it is a bit of a gimmick I agree :)
$ termy create file test.txt and write Hello
TERMy | template match | Confidence: 100.00%
Thinking: Ok, I am asked to create the file test.txt.
echo 'Hello' > 'test.txt' && termy_set_context 'active_file' 'test.txt'
Description: Writes Hello in file test.txt.
Response: Affirmative
Now that I think about it, I should let TERMy use tldr...