LLMs 的记忆和信息检索服务器 Motorhead

1. 为什么使用 Motorhead？

使用 LLMs构建聊天应用程序时，每次都必须构建记忆处理。Motorhead是协助该过程的服务器。

它提供了 4 个简单的 API：

• GET /sessions/:id/memory 返回最多 MAX_WINDOW_SIZE 的 messages

{
    "messages": [
        {
            "role": "AI",
            "content": "Electronic music and salsa are two very different genres of music, and the way people dance to them is also quite different."
        },
        {
            "role": "Human",
            "content": "how does it compare to salsa?"
        },
        {
            "role": "AI",
            "content": "Electronic music is a broad genre that encompasses many different styles, so there is no one \"right\" way to dance to it."
        },
        {
            "role": "Human",
            "content": "how do you dance electronic music?"
        },
        {
            "role": "AI",
            "content": "Colombia has a vibrant electronic music scene, and there are many talented DJs and producers who have gained international recognition."
        },
        {
            "role": "Human",
            "content": "What are some famous djs from Colombia?"
        },
        {
            "role": "AI",
            "content": "Baum opened its doors in 2014 and has quickly become one of the most popular clubs for electronic music in Bogotá."
        }
    ],
    "context": "The conversation covers topics such as clubs for electronic music in Bogotá, popular tourist attractions in the city, and general information about Colombia. The AI provides information about popular electronic music clubs such as Baum and Video Club, as well as electronic music festivals that take place in Bogotá. The AI also recommends tourist attractions such as La Candelaria, Monserrate and the Salt Cathedral of Zipaquirá, and provides general information about Colombia's diverse culture, landscape and wildlife.",
    "tokens": 744 // tokens used for incremental summarization
}

• POST /sessions/:id/memory - 向 Motorhead 发送数组 messages 进行存储

curl --location 'localhost:8080/sessions/${SESSION_ID}/memory' \
--header 'Content-Type: application/json' \
--data '{
    "messages": [{ "role": "Human", "content": "ping" }, { "role": "AI", "content": "pong" }]
}'

存储消息时，可以使用现有会话或新 SESSION_ID 会话，如果会话以前不存在，则会自动创建会话。

（可选） context 如果需要从其他数据存储加载，则可以将其送入。

• DELETE /sessions/:id/memory - 删除会话的消息列表。

A max window_size is set for the LLM to keep track of the conversation. Once that max is hit, Motorhead will process (window_size / 2 messages) and summarize them. Subsequent summaries, as the messages grow, are incremental.

为跟踪对话设置了 LLM 最大值 window_size 。一旦达到最大值，Motorhead 将处理（ window_size / 2 messages）并汇总它们。随着消息的增长，后续摘要是增量的。

• POST /sessions/:id/retrieval - 使用 VSS 按文本查询进行搜索

curl --location 'localhost:8080/sessions/${SESSION_ID}/retrieval' \
--header 'Content-Type: application/json' \
--data '{
    "text": "Generals gathered in their masses, just like witches in black masses"
}'

2. 通过 Docker 启动 Motorhead

docker run --rm --name some-redis -p 6379:6379 -d redis
docker run --rm --name motorhead -p 8080:8080 -e PORT=8080 -e REDIS_URL='redis://some-redis:6379' -d ghcr.io/getmetal/motorhead:latest

3. Github 地址

https://github.com/getmetal/motorhead

4. python 使用示例地址

https://github.com/getmetal/motorhead/tree/main/examples/chat-py

p.s. 暂时使用的可能性不大，所以先不做深入研究。

完结！