Pipe LLM response to files

[eddie]

I've been leveraging your llm, strip-tags, and ttok tools and this is the perfect addition! Thank you @simonw!

I have been toying with the reverse of this, where we can pipe typical LLM responses back to the files system with a quick confirmation step. Would this functionality belong in the files-to-prompt tool or in it's own utility, and I'm curious if anyone knows of a CLI tool that already does this?

Many thanks!

[irthomasthomas]

It is easy enough to send the output to a file with > and >>, but I guess you talking about something more intelligent, for instance, updating a function definition in a file? Or splitting one response into multiple files?

[fry69]

@simonw 's llm tool already stores every response inside a SQLite database by default, in case that is not known. Of course you can use his datasette tool to dig into that database.

If you want to go really nuts with the responses from the LLM, have a look at e.g. LangChain or Instructor.

LangChain offers a broad spectrum to persist responses, from in-memory, simple files to several databases. See langchain_community and langchain-postgres.

Instructor is more lightweight and focuses on forcing LLMs to return data in a structured form (it generates JSON schemas on the fly and adds them to the prompt, leveraging tool calling if available), so they can easily get processed further. For TypeScript aficionados like me, there is also an official port instructor-js. This is so brilliant: JSON (TypesScript) + Prompt -> LLM -> JSON back.

[eddie]

Thanks for your input @irthomasthomas and @fry69. I was thinking a unix-like tool for parsing out stdout responses e.g:

Some LLM response...

file.cpp:
```cpp
#include <iostream>
..

into corresponding files on the filesystem. What would be a typical approach for this? I was initially experimenting with some cobbled together regex's. It looks like instructor etc is close but unsure how would approach in this case.

[stoerr]

@eddie : there is a command line tool chatgptextractcodeblock in my chatgpt tool suite that extracts the code block, though that's done in Javascript.

BTW, you guys:
files-to-prompt is a wonderful light weight idea to give an LLM files input on the command line, especially if it's just a one shot action. Just mentioning - if you want to explore that way of generating files / doing changes systematically on many files or even in a build process, you might want to have a look at my AI based code generation pipeline (source on Github). That's a command line tool that started out from doing basically a files-to-prompt with writing the LLM response to a file, but has quite some more features now:

• only calls the AI if the input files or prompts have changed
• supports updating the output files instead of completely regenerating them
• can update just parts of the output file, mixed with existing manually written content
• you can ask the LLM questions why it did what it did / ask for suggestions

That also uses my "put it into the AI's mouth pattern" to encapsulate the
files into several chat messages, instead of just putting into the prompt. That might limit prompt injection risks
or other confusions.

Some of that might be implemented into LLM or plugins or partially into files-to-prompt by somebody, but that could
possibly too much for such general tools.