advent-ai v1.1.4

Advent

Advent is a ⚡️quick start kit ⚡️for Microsoft Semantic Kernel.

Advent automatically discovery both semantic and native skills from a folder, and exposes all the discovery skills via a REST API. And Advent uses Qdrant as the persistent semantic memory.

Get Started

Install

Using your favorite Node.js package management tool(mine is pnpm), run:

pnpm i advent-ai

API server

To use Advent API server for you awesome AI app, you can put your semantic and native skills under the skills folder. Provide a simple config file call advent.json:

{
  "Logging": {
    "LogLevel": {
      "Default": "Information"
    }
  },
  "Port": 6666,
  "Skills": [
    "./skills"
  ],
  "Models": {
    "Text": "text-davinci-003",
    "Chat": "gpt-3.5-turbo",
    "Embedding": "text-embedding-ada-002"
  },
  "Memory": {
    "Type": "Qdrant",
    "Host": "http://localhost",
    "Port": "6333"
  }
}

Make sure you have .NET Core and Qdrant installed, then run the following command to start the server:

npx advent api

The API of the server is really simple:

• List all available skills as HAL JSON
  • GET https://localhost:6666/api/skills
• Get detailed skill description
  • GET https://localhost:6666/api/skills/{skill}/{function}
• Execute functions
  • POST https://localhost:6666/api/asks?iterations={iterations}

In order to execute functions, the following JSON must be provided.

{
  "variables": [
    {
      "key": "INPUT",
      "value": "...."
    }
  ],
  "pipeline": [],
  "skills": []
}

variables is an array of kay value pair, for the input to the kernel.

pipeline is the chained or piped functions would like to run. For example, the following json will run TextSkill.Uppercase and TextSkill.TrimEnd as piped functions:

{
  "variables": [
    {
      "key": "INPUT",
      "value": " lowercase"
    }
  ],
  "pipeline": [
    {
      "skill": "TextSkill",
      "name": "Uppercase"
    },
    {
      "skill": "TextSkill",
      "name": "TrimEnd"
    }
  ],
  "skills": []
}

If no functions specified, it will run PlannerSkill.CreatePlan and PlannerSkill.ExecutePlan by default (a.k.a, archive goal). And the iterations query parameter will be used to determine how many times should the kernel try before the plan execute successfully.

skills indicates which skills will be used during the execution. Since the planner tend to use most of the available functions, the result plan might be too long. And can't fit within the token limits. Then skills could be used to tell the kernel exactly which skills should the plan be based on.

Every API call should provide OpenAI API key via HTTP headers:

And if you use the same key for different purposes, you only need to provide x-advent-text-completion-key.

Embeddings indexing

Semantic memory with "embeddings" is growing in popularity when a set of documents needed to be provide for LLM. To index documents, run the following command:

npx advent index <path to folder> -c <collection name> -i .md .txt 

That's it. Have fun with AI 🧗‍!

License

Advent is licensed under the MIT License.