31 lines
1.3 KiB
Markdown
31 lines
1.3 KiB
Markdown
<h1 align="center">miniLLMFlow</h1>
|
|
|
|
|
|
|
|
Minimalist LLM framework in [100 lines](minillmflow/__init__.py). Express popular paradigms like agents, task decomposition, RAG, and more.
|
|
|
|
- Install via ```pip install minillmflow```. Or just copy the [source codes](minillmflow/__init__.py) (it's only 100 lines)
|
|
- We **strongly recommened** using LLMs (e.g., Claude, ChatGPT, Cursor) to build LLM applications with [this prompt](minillmflow/docs/prompt)
|
|
|
|
## Why miniLLMFlow?
|
|
|
|
The future of programming will be heavily LLM-assited, and LLMs:
|
|
|
|
1. **👍 Excel at Feature Implementation**:
|
|
Give proper context/docs, they know how to use APIs, tools, text chunking, prompt wrapping, etc.
|
|
These shouldn't be part of an LLM framework, as they're hard to maintain, update, and optimize.
|
|
|
|
2. **👎 Suck at Paradigm Design**:
|
|
Paradigms like MapReduce, DAG workflows, and recent agents are powerful for reasoning about problems.
|
|
However, designing elegant paradigms is challenging, and LLMs tend to write redundant code.
|
|
|
|
Can we build a framework that:
|
|
(1) removes redunant feature implementations,
|
|
(2) but keep core paradigms for LLMs to program against?
|
|
Turns out that we just need 100 lines of codes.
|
|
|
|
<div align="center">
|
|
<img src="./docs/minillmflow.jpg" width="400"/>
|
|
</div>
|
|
|