gpt meeting service

gpt-meeting-service

Zero, what is it for?

1. This is a small project based on GPT API development. As we all know, the current context support of GPT is to re-send historical messages every time. The method is very "simple and crude", and this project simply does one thing: to make the context organization more flexible to support the completion of more complex tasks. some tasks.
2. The original source of inspiration for the project was "brainstorming" and "6 Thinking Hats". During the brainstorming process, each link has a different focus and requires different roles to play. This process is consistent with the use of the six thinking hats. The flexible use of the six thinking hats can achieve different purposes. After further extension, there is this project "meeting" or discussion.
3. Based on the above, the project mainly implements the following three functions:
   1. The arrangement of the task flow chart, each link completes one thing, and is strung together to complete a relatively complex task (also understood as meeting agenda design).
   2. Let GPT ask and answer its own questions and complete the association.
   3. Let GPT play different roles and discuss the same topic.
4. Possible application scenarios
   1. brainstorming
   2. Quick association
   3. Quick assessment
   4. Multi-character dialogue

Associated front-end project: gpt-meeting-web

1. Node type

Node refers to the agenda (meeting link) in the "meeting" process. It is currently divided into 5 types, and each type of node handles the context differently.

General situation: background information + AI character design + Prompt + input = output

Input

Generally, you enter the topic and goal of this time, and you can submit it directly or let GPT optimize it before submitting it.

Thinking

Lenovo includes two AI human characters, among which the Lenovo character is used to generate multiple ideas; the Questioning character is used to ask questions related to the generated ideas, and can ask multiple questions and follow up (corresponding to the "Question Round" and "number of questions"), so the overall structure appears as a tree structure, generating ideas corresponds to the root node of the tree, and asking questions about ideas corresponds to the growth of branches. In addition to automatic question and answer, manual intervention can also be performed to ask questions.

DiscussionDiscussion

The discussion node can select multiple members with different roles, and each member can participate in the discussion according to his own characteristics.

General processingProcessing

Customize prompts to process input.

Output

The output is generally used as the last node to summarize the entire discussion process, and then output in a specified format as the final conclusion.

2. Working principle

Contextual organization of link (agenda) nodes

First, the flow chart must be a directed acyclic graph, and then after the "meeting" starts, it will be executed in topological order. On each node, the "conclusions" of all upstream nodes of the current node will be regarded as its own "previous text". This can ensure that GPT obtains sufficient background information and then answers the task of the current node. Theoretically, the effect will be More ideal.

Contextual organization in association

The association of self-question and self-answer is carried out in a tree structure, so the organization of the context is naturally the dialogue records of all ancestor nodes of the current association node. This is easy to understand, such as the mark of the red box in the figure below.

Contextual organization in multi-role discussions

The beginning of the multi-role discussion will automatically bring in background information and the requirements of the current session. After that, each member's speech will bring the previous record as background information.

development related

Technology selection

• Language: golang
• Framework: kratos
• Storage: MongoDB

Recommended environment

vscode

Run locally

 # 配置configs/config.yaml

# 依赖安装（在Makefile文件中定义了相关命令）
make init

# 运行(调试建议使用vscode)
kratos run

# 导入初始模版
cd cmd/script && go run dataOp.go importData

docker image build

 docker build -t gpt-meeting-service:v1 .

docker-compose deployment

 cd docker-compose
# 配置文件
mkdir conf && cp ../configs/ ./conf

# 启动
docker-compose up -d

# 停止
docker-compose down

illustrate

• The project uses the kratos framework, you need to understand this framework first
• Since protoc does not support sse, the meeting module does not use protoc to define the interface.
• Reference article for getting started with kratos
• For kratos commands, please refer to kratos.md