Top LLM frameworks may not be as reliable as you think
Nearly a month ago, I decided to add Gemini support to Feeds Fun and did some research on top LLM frameworks — I didn't want to write my own bicycle.
As a result, I found an embarrassing bug (in my opinion, of course) in the integration with Gemini in LLamaIndex. Judging by the code, it is also present in Haystack and in the plugin for LangChain. And the root of the problem is in the Google SDK for Python.
When initializing a new client for Gemini, the framework code overwrites/replaces API keys in all clients created before. Because the API key, by default, is stored in a singleton.
It is death-like, if you have a multi-tenant application, and unnoticeable in all other cases. Multi-tenant means that your application works with multiple users.
For example, in my case, in Feeds Fun, a user can enter their API key to improve the quality of the service. Imagine what a funny situation could happen: a user entered an API key to process their news but spent tokens (paid for) for all service users.
I reported this bug only in LLamaIndex as a security issue, and there has been no reaction for 3 weeks. I'm too lazy to reproduce and report for Haystack and LangChain. So this is your chance to report a bug to a top repository. All the info will be below, reproducing is not difficult.
This error is notable for many reasons:
- The assessment of the criticality of the error depends a lot on taste, experience, and context. For me, in the projects I worked on, this is a critical security issue. However, it seems that this is not critical at all for most current projects that use LLMs. Which leads to some thoughts about mainstream near-LLM development.
- This is a good indicator of a low level of code quality control: code reviews, tests, all processes. After all, this is an integration with one of the major API providers. The problem could have been found in many different ways, but none worked.
- This is a good illustration of the vicious approach to development: "copy-paste from a tutorial and push to prod". To make such a mistake, you had to ignore both the basic architecture of your project and the logic of calling the code you are copying.
Ultimately, I gave up on these frameworks and implemented my own client over HTTP API.
My conclusion from this mess is: you can't trust the code under the hood of modern LLM frameworks. You need to double-check and proofread it. Just because they state that they are "production-ready" doesn't mean they are really production-ready.
Let me tell you more about the bug.
Grainau: hiking and beer at 3000 meters
How it all looks from the ground.
For her vacation, Yuliya decided to show me the beautiful German mountains and took me for a couple of days to Grainau — it's a piece of Bavaria that's almost like Switzerland. At least, it is similar to the pictures of Switzerland that I've seen :-D
In short, it's a lovely place with a measured pace of life. If you need to catch your breath, calm your nerves, and enjoy nature, then this is the place for you. But if you can't live without parties, you'll get bored quickly.
What's there:
- The highest mountain in Germany plus a couple of glaciers.
- There's skiing in winter. If you really need it, you can find a place to ski in summer, but the descent is short, and the lifts are turned off.
- A large clean lake and a couple of smaller ones.
- A huge number of trails for hiking.
- A huge number of waterfalls, streams, and a couple of mountain rivers.
- Restaurants with beer.
- Beautiful fallen trees in the forests, private property, fences, cows with bells, and "racing tractors" (I don't know how to name this phenomenon better, but tractors are moving fast there :-D).
This is briefly, and now in detail.
Migrating from GPT-3.5-turbo to GPT-4o-mini
Guess when I switched models.
Recently OpenAI released GPT-4o-mini — a new flagship model for the cheap segment, as it were.
- They say it works "almost like" GPT-4o, sometimes even better than GPT-4.
- It is almost three times cheaper than GPT-3.5-turbo.
- Context size 128k tokens, against 16k for GPT-3.5-turbo.
Of course, I immediately started migrating my news reader to this model.
In short, it's a cool replacement for GPT-3.5-turbo. I immediately replaced two LLM agents with one without changing prompts, reducing costs by a factor of 5 without losing quality.
However, then I started tuning the prompt to make it even cooler and began to encounter nuances. Let me tell you about them.
Dungeon generation — from simple to complex
What we should get.
This is a translation of a post from 2020
This is a step-by-step guide to generating dungeons in Python. If you are not a programmer, you may be interested in reading how to design a dungeon [ru].
I spent a few evenings testing the idea of generating space bases.. The space base didn't work out, but the result looks like a good dungeon. Since I went from simple to complex and didn't use rocket science, I converted the code into a tutorial on generating dungeons in Python.
By the end of this tutorial, we will have a dungeon generator with the following features:
- The rooms will be connected by corridors.
- The dungeon will have the shape of a tree. Adding cycles will be elementary, but I'll leave it as homework.
- The number of rooms, their size, and the "branching level" will be configurable.
- The dungeon will be placed on a grid and consist of square cells.
The entire code can be found on github.
There won't be any code in the post — all the approaches used can be easily described in words. At least, I think so.
Each development stage has a corresponding tag in the repository, containing the code at the end of the stage.
The aim of this tutorial is not only to teach how to generate dungeons but to demonstrate that seemingly complex tasks can be simple when properly broken down into subtasks.
«Slay The Princess» — combinatorial narrative
My favorite version of the Princess.
It's hard to impress me as a player and even harder as a game developer. The last time it happened with Owlcat Games in Pathfinder: Kingmaker, when they added a timer to the game's plot.
But Black Tabby Games managed to do it. And they did it not with some technological complexity but with a visual novel on a standard engine (RenPy), which is cool in itself.
I'll share a couple of thoughts about the game and its narrative structure, while I'm still under the impression. I need to think about how to adapt this approach to my projects.
ATTENTION: SPOILERS!
If you haven't played Slay The Princess yet, I strongly recommend you to catch up — the game takes 3-4 hours. You'll not regret it.