A new way to build with Large Language Models
blog.fixie.ai
A new way to build with Large Language Models
1–10 of 46 posts
Re: A new way to build with Large Language Models
#2if I’m understanding this correctly this is a way for LLMs to call external APIs by showing them special key words that call functions
I assume the output of the LLM is monitored by fixie and the result of the function inserted allow the LLM to respond again?
Re: A new way to build with Large Language Models
#3if I’m understanding this correctly this is a way for LLMs to call external APIs by showing them special key words that call functions I assume the output of the LLM is monitored by fixie and the result of the function inserted allow the LLM to respond again?
yes, the LLM sees the result of the function and processes it according to what it has learned from its few-shots (which may involve calling more functions, or returning a formatted response).
Re: A new way to build with Large Language Models
#4Neat. I was just thinking that one thing you can do that current large language models don't do is use a calculator to check your work.
So, the first thing I'd want to try is hooking up some basic arithmetic to see if its math gets more reliable. And if not, why not?
Seems like someone would have tried this already?
Re: A new way to build with Large Language Models
#5Neat, this is like a SaaS version of Langchain[1]. For those who haven't already seen it, Langchain is a framework that gives you the ability to do this as well as integrate out of the box solutions for other common problems like memory.
Re: A new way to build with Large Language Models
#6Neat, this is like a SaaS version of Langchain[1]. For those who haven't already seen it, Langchain is a framework that gives you the ability to do this as well as integrate out of the box solutions for other common problems like memory. [1] https://langchain.readthedocs.io/en/latest/
Langchain is awesome. Another great library is GPT Index. This field is apparently being called "prompt orchestration" (getting data from external sources, chaining going back and forth with an LLM, etc).
Re: A new way to build with Large Language Models
#7Neat. I was just thinking that one thing you can do that current large language models don't do is use a calculator to check your work. So, the first thing I'd want to try is hooking up some basic arithmetic to see if its math gets more reliable. And if not, why not? Seems like someone would have tried this already?
It's built-in :-)
Re: A new way to build with Large Language Models
#8Neat, this is like a SaaS version of Langchain[1]. For those who haven't already seen it, Langchain is a framework that gives you the ability to do this as well as integrate out of the box solutions for other common problems like memory. [1] https://langchain.readthedocs.io/en/latest/
You can use both Langchain and GPT Index with Fixie Agents. We're fans of what they've built, and they can be great tools.
Re: A new way to build with Large Language Models
#9What's your guidance concerning the risk of prompt injection with applications built on this kind of platform?
If developers on your platform are building code that accepts prompt input from untrusted users do you have any measures in place to avoid them tricking the system into executing functions in an inappropriate or unintended way?
One thing I'd find useful would be a limit on the number of functions that can be called during a single execution - and rate limits and budget calls on functions generally might be smart as well.
Re: A new way to build with Large Language Models
#10Could you use this to provide better support chat bots? That is, support bots obviously need to be scoped so that they only give information the user asking has rights to, e.g. "Show me the stocks in my portfolio". Assuming the user has logged into the interface showing the chat, is there a way to pass that user's unique access token through to one of these Fixie Agents to ensure that the data being accessed is scoped by that access token?