Arxiv preprint | DSPy Implementation

AvaTaR is a novel and automatic framework that optimizes an LLM agent to effectively use the provided tools and improve its performance on a given task/domain. During optimization, we design a comparator module to iteratively provide insightful and holistic prompts to the LLM agent via reasoning between positive and negative examples sampled from training data.

[July 2024] 🔥 Avatar is integrated into DSPy - Credit to Herumb Shandilya! You can try out the example on jupyter notebook.

Avatar is now integrated with DSPy as Avatar Module for agent execution and AvatarOptimizer for Actor optimization. To use Avatar you'll need: Task Signature and Tools.

• Task Signature is a dspy.Signature class defining the structure of your task. So if your task is of QA type you can create a signature with question input field and answer output field.
• Tools is a list of dspy.Tools containing all the tools of langchain tool format.

Here is an example

from dspy.predict.avatar import Tool, Avatar
from langchain_community.utilities import GoogleSerperAPIWrapper, ArxivAPIWrapper

tools = [
    Tool(
        tool=GoogleSerperAPIWrapper(),
        name="WEB_SEARCH",
        desc="If you have a question, you can use this tool to search the web for the answer."
    ),
]

agent = Avatar(
    tools=tools,
    signature="question->answer",
    verbose=True,
)

You can execute it like any other DSPy module by passing the inputs you specified in your task signature:

answer = agent(question)

You can optimize the Actor for optimal tool usage using AvatarOptimizer which optimizes it using the comparator module:

from dspy.teleprompt import AvatarOptimizer

def metric(example, prediction, trace=None):
    ...

teleprompter = AvatarOptimizer(
    metric=metric,
    max_iters=10,
    max_negative_inputs=10,
    max_positive_inputs=10,
)

optimized_arxiv_agent = teleprompter.compile(
    student=agent,
    trainset=trainset
)

For a detailed walkthrough, you can refer to the notebook in DSPy repo.

conda create -n avatar python=3.11
pip install stark-qa typeguard

• Specify API keys in command line

  export ANTHROPIC_API_KEY=YOUR_API_KEY

  export OPENAI_API_KEY=YOUR_API_KEY
  export OPENAI_ORG=YOUR_ORGANIZATION

• Embeddings: Download all embeddings by running the following script:

  sh scripts/emb_download_all.sh

• Raw data: STaRK data will be downloaded automatically when running the code. For Flickr30k Entities, submit form at Flickr 30k & Denotation Graph data to request access. Then organize the data as follows:

  data
  ├── flickr30k_entities
  │   ├── raw
  │   │   ├── Annotations
  │   │   │   ├── 36979.xml
  │   │   │   ├── ...
  │   │   ├── flickr30k-images
  │   │       ├── 36979.jpg
  │   │       ├── ...
  │   ├── split
  │   │   ├── test.index
  │   │   ├── train.index
  │   │   ├── val.index
  │   ├── qa.csv
  ├── ...
  

We already include the VSS results locally under output/eval and the grouping (for STaRK only) under output/agent. With these files, you should be able to optimize actor actions directly following the AvaTaR pipeline.

• Optimization: Following the default settings at config/default_args.json, run the following command to optimize the actor actions for a group of queries:

  sh scripts/run_avatar_stark.sh

  You can specify the dataset name and group in scripts/run_avatar_stark.sh.

  sh run_avatar_flickr30k_entities.sh

• Evaluation: Run the following command to evaluate the optimized actor actions:

  sh scripts/run_eval_avatar_stark.sh

  or

  sh scripts/run_eval_avatar_flickr30k_entities.sh

@article{wu24avatar,
    title        = {AvaTaR: Optimizing LLM Agents for Tool-Assisted Knowledge Retrieval},
    author       = {
        Shirley Wu and Shiyu Zhao and 
        Qian Huang and Kexin Huang and 
        Michihiro Yasunaga and Kaidi Cao and 
        Vassilis N. Ioannidis and Karthik Subbian and 
        Jure Leskove and James Zou
    },
    eprinttype   = {arXiv},
    eprint       = {2406.11200},
  year           = {2024}
}