🌱 seedling

A simple and supposedly scalable python library for intent recognition using Large Language Models (LLMs).

The Idea

Seedling 🌱 gives you a very simple and scalable way to detect intents from user messages using LLMs. In order to make things scalable, seedling forces you to group your intents to topics. This also keeps your prompt length short, reduces LLM hallucination, and makes LLM output more accurate.

Example image showing intents grouped under six different topics

Seedling supports OpenAI/OpenAI-Complatible APIs out of the box. The OpenAI-Compatible API server must support tools use. It is also very easy to write your own LLM interface for prediction purposes.

Defining Topic YAML

Following is a simple topic YAML file for the map topic with the related intents grouped under it. Please note that both the topic and the intent description must be as detailed as possible, as these get fed into the LLM.

name: "map"
description: "Get directions, find places such as restaurants, shops, places of interest, etc."

intents:
  - name: "get_directions"
    description: "Gets directions between two locations."
  - name: "find_place"
    description: "Finds a specific place like a restaurant, gas station, or store."
  - name: "get_traffic_information"
    description: "Gets real-time traffic information for a route."

Each YAML files must NOT contain more than 1 topic.

One-shot entity extraction

In addition to detecting intents, seedling also allows you to attempt to extract entities in one-shot. For enitity extraction, the smarter your LLM is, the better.

name: "travel"
description: "Plan and manage travel arrangements. Search for hotels, flights, and rental cars."

intents:
  - name: "search_flights"
    description: "Searches for flights based on origin, destination, and dates."
    entities:
      - name: "origin"
        type: "string"
        description: "The origin of the flight. Write 'none' if not provided."
      - name: "destination"
        type: "string"
        description: "The destination of the flight.  Write 'none' if not provided."
      - name: "date"
        type: "string"
        description: "The date of the flight. Write 'none' if not provided."

Basic Usage

from seedling import YAMLreader, LanguageModel, predict

# Path to the directory containing all your topic/intent YAML files
all_topics = YAMLreader("directory_path")

# OpenAI/OpenAI-compatible LLM server (must support tools use)
# check example/main.py for Ollama config
llm = LanguageModel(
    base_url='http://baseurl',
    api_key='apikey',
    model='model_name',
)

# Predict the intent of the user message
user_message = "do some magic on spotify!"
predicted = predict(llm, all_topics, user_message)

print(predicted)

Check out the example

The example is great for wrapping your head around everything. Follow these steps to run the example:

1. Start a virtual environment

   python -m venv venv
   source venv/bin/activate
   

2. Build the seedling 🌱 library

   pip install -e .
   

3. Install the latest version of Ollama and pull down the llama3.1 model. Alternatively, you can also update the example file under example/main.py with your OpenAI/OpenAI-compatible API server information.

   ollama pull llama3.1
   

4. Run the example

   cd example
   python main.py
   

Local Development

• Only tested on python 3.11 so far
• Currently, the actual library source code can be found under src/seedling
• To install all dependencies do pip install ."[dev]" in your virtual environment
• Use make lint to run the linter