offers a unified framework for explicitly constructing a complex human "thought process" from simple natural language prompts.
The user puts together chains of nodes, like stacking LEGO pieces. The chains of nodes can be designed to explicitly enforce a naturally structured "thought process".
Different arrangements of nodes could represent different functionalities, allowing the user to integrate various functionalities to build multifunctional agents.
A basic agent could be implemented as simple as a list of prompts for the subtasks and therefore could be designed and tuned by someone without any programming experience.
- Installation
- Getting Started
- Using Built-in LLM_API
- Using AgentKit without Programming Experience
- Node Components
- Commonly Asked Questions
- Citing AgnetKit
Installing the AgentKit stable version is as simple as:
pip install agentkit-llmTo install AgentKit with wandb:
pip install agentkit-llm[logging]To install AgentKit with OpenAI and Claude LLM-API support:
pip install agentkit-llm[proprietary]To install AgentKit with full built-in LLM-API support (including llama):
pip install agentkit-llm[all]Otherwise, to install the cutting edge version from the main branch of this repo, run:
git clone https://github.com/holmeswww/AgentKit && cd AgentKit
pip install -e .The basic building block in AgentKit is a node, containing a natural language prompt for a specific subtask. The nodes are linked together by the dependency specifications, which specify the order of evaluation. Different arrangements of nodes can represent different different logic and throught processes.
At inference time, AgentKit evaluates all nodes in specified order as a directed acyclic graph (DAG).
import agentkit
from agentkit import Graph, BaseNode
import agentkit.llm_api
LLM_API_FUNCTION = agentkit.llm_api.get_query("gpt-4-turbo")
LLM_API_FUNCTION.debug = True # Disable this to enable API-level error handling-retry
graph = Graph()
subtask1 = "What are the pros and cons for using LLM Agents for Game AI?"
node1 = BaseNode(subtask1, subtask1, graph, LLM_API_FUNCTION, agentkit.compose_prompt.BaseComposePrompt(), verbose=True)
graph.add_node(node1)
subtask2 = "Give me an outline for an essay titled 'LLM Agents for Games'."
node2 = BaseNode(subtask2, subtask2, graph, LLM_API_FUNCTION, agentkit.compose_prompt.BaseComposePrompt(), verbose=True)
graph.add_node(node2)
subtask3 = "Now, write a full essay on the topic 'LLM Agents for Games'."
node3 = BaseNode(subtask3, subtask3, graph, LLM_API_FUNCTION, agentkit.compose_prompt.BaseComposePrompt(), verbose=True)
graph.add_node(node3)
# add dependencies between nodes
graph.add_edge(subtask1, subtask2)
graph.add_edge(subtask1, subtask3)
graph.add_edge(subtask2, subtask3)
result = graph.evaluate() # outputs a dictionary of prompt, answer pairsLLM_API_FUNCTION can be any LLM API function that takes msg:list and shrink_idx:int, and outputs llm_result:str and usage:dict. Where msg is a prompt (OpenAI format by default), and shrink_idx:int is an index at which the LLM should reduce the length of the prompt in case of overflow.
AgentKit tracks token usage of each node through the LLM_API_FUNCTION with:
usage = {
'prompt': $prompt token counts,
'completion': $completion token counts,
}The built-in agentkit.llm_api functions require installing with [proprietary] or [all] setting. See the installation guide for details.
Currently, the built-in API supports OpenAI and Anthropic, see https://pypi.org/project/openai/ and https://pypi.org/project/anthropic/ for details.
To use the OpenAI models, set environment variables OPENAI_KEY and OPENAI_ORG. Alternatively, you can put the openai 'key' and 'organization' in the first 2 lines of ~/.openai/openai.key.
To use the Azure OpenAI models, set environment variables AZURE_OPENAI_API_KEY, AZURE_OPENAI_API_VERSION, AZURE_OPENAI_ENDPOINT, and AZURE_DEPLOYMENT_NAME. Alternatively, you can store the Azure OpenAI API key, API version, Azure endpoint, and deployment name in the first 4 lines of ~/.openai/azure_openai.key.
To use the Anthropic models, set environment variable ANTHROPIC_KEY. Alternatively, you can put the anthropic 'key' in 3rd line of ~/.openai/openai.key.
To use Ollama models, see https://github.com/ollama/ollama for installation instructions. Then set OLLAMA_URL and OLLAMA_TOKENIZER_PATH, or store OLLAMA_TOKENIZER_PATH, OLLAMA_URL in the first 2 lines of ~/.ollama/ollama_model.info.
LLM_API_FUNCTION = agentkit.llm_api.get_query("ollama-llama3")
First, follow the installation guide to install AgentKit with [all] setting.
Then, set environment variables OPENAI_KEY and OPENAI_ORG to be your OpenAI key and org_key.
Finally, run the following to evoke the command line interface (CLI):
git clone https://github.com/holmeswww/AgentKit && cd AgentKit
cd examples/prompt_without_coding
python generate_graph.py
Inside each node (as shown to the left of the figure), AgentKit runs a built-in flow that preprocesses the input (Compose), queries the LLM with a preprocessed input and prompt
To support advanced capabilities such as branching, AgentKit offers API to dynamically modify the DAG at inference time (as shown to the right of the figure). Nodes/edges could be dynamically added or removed based on the LLM response at some ancestor nodes.
Q: I'm using the default agentkit.llm_api, and graph.evaluate() seems to be stuck.
A: The LLM_API function catches and retries all API errors by default. Set verbose=True for each node to see which node you are stuck on, and LLM_API_FUNCTION.debug=True to see what error is causing the error.
@article{wu2024agentkit,
title={AgentKit: Flow Engineering with Graphs, not Coding},
author={Yue Wu and Yewen Fan and So Yeon Min and Shrimai Prabhumoye and Stephen McAleer and Yonatan Bisk and Ruslan Salakhutdinov and Yuanzhi Li and Tom Mitchell},
year={2024},
journal={arXiv preprint arXiv:2404.11483}
}