wehub-resource-sync
e768098d0e
tools_continuous_delivery / Private PyPI non-main branch release (push) Has been skipped
tools_continuous_delivery / Private PyPI main branch release (push) Failing after 2m42s
Publish Promptflow Doc / Build (push) Has been cancelled
Publish Promptflow Doc / Deploy (push) Has been cancelled
Flake8 Lint / flake8 (push) Has been cancelled
Spell check CI / Spell_Check (push) Has been cancelled
66 lines
3.8 KiB
Markdown
66 lines
3.8 KiB
Markdown
# Autonomous Agent
|
|
|
|
This is a flow showcasing how to construct a AutoGPT agent with promptflow to autonomously figures out how to apply the given functions
|
|
to solve the goal, which is film trivia that provides accurate and up-to-date information about movies, directors,
|
|
actors, and more in this sample.
|
|
|
|
It involves inferring next executed function and user intent with LLM, and then use the function to generate
|
|
observation. The observation above will be used as augmented prompt which is the input of next LLM inferring loop
|
|
until the inferred function is the signal that you have finished all your objectives. The functions set used in the
|
|
flow contains Wikipedia search function that can search the web to find answer about current events and PythonREPL
|
|
python function that can run python code in a REPL.
|
|
|
|
For the sample input about movie introduction, the AutoGPT usually runs 4 rounds to finish the task. The first round
|
|
is searching for the movie name, the second round is searching for the movie director, the third round is calculating
|
|
director age, and the last round is outputting finishing signal. It takes 30s~40s to finish the task, but may take
|
|
longer time if you use "gpt-3.5" or encounter Azure OpenAI rate limit. You could use "gpt-4" or go to
|
|
https://aka.ms/oai/quotaincrease if you would like to further increase the default rate limit.
|
|
|
|
Note: This is just a sample introducing how to use promptflow to build a simple AutoGPT. You can go to
|
|
https://github.com/Significant-Gravitas/Auto-GPT to get more concepts about AutoGPT.
|
|
|
|
## What you will learn
|
|
|
|
In this flow, you will learn
|
|
- how to use prompt tool.
|
|
- how to compose an AutoGPT flow using functions.
|
|
|
|
## Prerequisites
|
|
|
|
Install prompt-flow sdk and other dependencies:
|
|
```bash
|
|
pip install -r requirements.txt
|
|
```
|
|
|
|
## Getting Started
|
|
|
|
### 1 Create Azure OpenAI or OpenAI connection
|
|
```bash
|
|
# Override keys with --set to avoid yaml file changes
|
|
pf connection create --file ../../../connections/azure_openai.yml --set api_key=<your_api_key> api_base=<your_api_base>
|
|
```
|
|
Note that you need to use "2023-07-01-preview" as Azure OpenAI connection API version when using function calling.
|
|
See <a href='https://learn.microsoft.com/en-us/azure/ai-services/openai/how-to/function-calling' target='_blank'>How to use function calling with Azure OpenAI Service</a> for more details.
|
|
|
|
### 2. Configure the flow with your connection
|
|
`flow.dag.yaml` is already configured with connection named `open_ai_connection`. It is recommended to use "gpt-4" model for stable performance. Using "gpt-3.5-turbo" may lead to the model getting stuck in the agent inner loop due to its suboptimal and unstable performance.
|
|
|
|
### 3. Test flow with single line data
|
|
|
|
```bash
|
|
# test with default input value in flow.dag.yaml
|
|
pf flow test --flow .
|
|
```
|
|
|
|
### 4. Run with multi-line data
|
|
|
|
```bash
|
|
# create run using command line args
|
|
pf run create --flow . --data ./data.jsonl --column-mapping name='${data.name}' role='${data.role}' goals='${data.goals}' --stream
|
|
```
|
|
|
|
You can also skip providing `column-mapping` if provided data has same column name as the flow.
|
|
Reference [here](https://aka.ms/pf/column-mapping) for default behavior when `column-mapping` not provided in CLI.
|
|
|
|
## Disclaimer
|
|
LLM systems are susceptible to prompt injection, and you can gain a deeper understanding of this issue in the [technical blog](https://developer.nvidia.com/blog/securing-llm-systems-against-prompt-injection/). As an illustration, the PythonREPL function might execute harmful code if provided with a malicious prompt within the provided sample. Furthermore, we cannot guarantee that implementing AST validations solely within the PythonREPL function will reliably elevate the sample's security to an enterprise level. We kindly remind you to refrain from utilizing this in a production environment. |