Abstract

GPT+artifact, simple and practical one-stop AGI architecture, built-in localization, LLM model, agent， Vector database, intelligent chain

Introduction

GPT+artifact, a simple and practical one-stop AGI architecture with built-in native vector database, supporting localized LLM models, agent， Intelligent chain, prompt prompt prompt

10 lines of code can replicate the classic GPT open source project: localGPT，privateGPT。

Single py source code module file, implementing 80% functionality of langchain+prompt prompt+llm+vector database+AGI (autoGPT, babyAGI).

Methodology

Experiments

Vector data conversion, 300 data per second, 1.5w per minute, nearly one million per hour, sufficient for small and medium-sized enterprises

Mini AGI is a free and open-source version of ZW AGI, with some business modules removed.

Results

[Completed Function]
10 lines of code to replicate the classic GPT open source project: localGPT，privateGPT，
Single py source code module file, implementing 80% functionality of langchain+prompt prompt+llm+vector database+AGI (autoGPT, babyAGI).
Single file module source code, copy working directory, import directly, no installation required.
Built in native vector database, based on pandas, with a complete ecosystem.
Support: Import data from various formats such as txt, csv, docx, ppt, pdf, markdown, etc.
Open prompts prompt word document, with multiple sets of commonly used prompts built-in.

Built in LLM call:
Support the Tsinghua CGATGLM-6B second-generation model, as it is the strongest open-source LLM model in China.
The first batch supports GGML format, Llamcpp is a pure CPU call, and ZW-VICUNA-7B, 13B-3B, and 33B have been tested and passed.
Supports most mainstream open source models, including: LLaMA 🦙、 Alpaca、 GPT4All、 Chinese LLaMA / Alpaca、 Vigogne (French)、 Vicuna、 GGML、Koala、 OpenBuddy 🐶 (Multilingual)、 Pygmalion / MetharmeB、 WizardLM、 Baichuan and its derivations (such as baichuan-sft)
Todo list, plan completion function
Agent Agent
Tool application
Intelligent Chain
Open and free web search, key free
......

Conclusion

Mini agi module library file description
The mini AGI structure is very simple, with a single Python file structure. Save in the mini xlib directory.
Most of the structure and code are original, emphasizing engineering practicality and prioritizing conventions.
Part of the code has been referenced from relevant open source projects, and we would like to express our gratitude.
[Core file description]
mini_agi.py， Core files of module library
prompt.yaml， Prompt word template, in dictionary format, saved in YAML file, users can add it themselves.
【 Plungin Plugin Description 】
The mini AGI plugin also uses a single py file mode, which is simple and clear. When released, three plugin case source codes were built-in for easy extension.

mini_glm.py， Tsinghua chatGLM2-6B plugin, supporting domestic LLM projects.
Mini_gpt. py, openAI API plugin, set the URL and key in the source code file before use
Mini_1lm.by, ggml model plugin, supports ZW-Vicoan-7B/13B/33B Chinese version, tested and approved.
Installing the llama cpp python module library requires a C language compiler.
For specific details, please refer to: https://github.com/ggerganov/llama.cpp https://github.com/abetlen/llama-cpp-python
QA Query Demo Sample Program
tur001-qa.py， Less than 10 lines of code, perfectly replicating the classic GPT open source project: localGPT，privateGPT。

Tur002-gpt-py, GPT3.5-api interface mode

Tur03-llm.py, ggml model file call, supports ZW-Vicoan-7B/13B/33B Chinese version, tested and approved.

VDB database generation demo example program
VDB vector database, using the standard pandas DataFrame format. The database file extension is pkl, and the storage is in feather format.

tur011_db8fn.py， Generate a vec vector database from a CSV file
tur012_db8dir.py， Batch generate vec vector database from directory, support subdirectories structure, support multiple file formats, sample program demo includes: doc, txt, pdf and other formats
[Model file description]
The default model is saved in the libs directory,

libs\all-MiniLM-L6-v2， Text Vector Conversion Model
libs\zw-v7cn.bin， Zhiwang ZW-VICUNA-13B-CN Chinese version
libs\zw-v33cn.bin， Zhiwang ZW-VICUNA-33B-CN Chinese version
libs\chatglm2-6b， Tsinghua ChatGLM2-6B Model
The above models can be downloaded from Zhiwang Baidu Cloud or downloaded from websites such as Huggingface.