AI Agent execution using Temporal

Work in progress.

This demo shows a multi-turn conversation with an AI agent running inside a Temporal workflow. The goal is to collect information towards a goal. There's a simple DSL input for collecting information (currently set up to use mock functions to search for events, book flights around those events then generate an invoice for those flights, see send_message.py). The AI will respond with clarifications and ask for any missing information to that goal. It uses a local LLM via Ollama.

Setup

• Requires an OpenAI key for the gpt-4o model. Set this in the OPENAI_API_KEY environment variable in .env
• Requires a rapidapi key for sky-scrapper (how we find flights). Set this in the RAPIDAPI_KEY environment variable in .env
  • It's free to sign up and get a key at RapidAPI
  • If you're lazy go to tools/search_flights.py and replace the get_flights function with the mock search_flights_example that exists in the same file.
• See .env_example for the required environment variables.
• Install and run Temporal. Follow the instructions in the Temporal documentation to install and run the Temporal server.
• Install the dependencies: poetry install

Deprecated:

• Install Ollama and the Qwen2.5 14B model (ollama run qwen2.5:14b). (note this model is about 9GB to download).
  • Local LLM is disabled as ChatGPT 4o was better for this use case. To use Ollama, examine ./activities/tool_activities.py and rename the functions.

Running the example

From the /scripts directory:

1. Run the worker: poetry run python run_worker.py

2. In another terminal run the client with a prompt.

   Example: poetry run python send_message.py 'Can you find events in march in oceania?'

3. View the worker's output for the response.

4. Give followup prompts by signaling the workflow.

   Example: poetry run python send_message.py 'I want to fly from San Francisco'

   NOTE: The workflow will pause on the 'confirm' step until the user sends a 'confirm' signal. Use poetry run python get_tool_data.py query to see the current state of the workflow.

   You can send a 'confirm' signal using poetry run python send_confirm.py

5. Get the conversation history summary by querying the workflow.

   Example: poetry run python get_history.py

6. To end the chat session, run poetry run python end_chat.py

The chat session will end if it has collected enough information to complete the task or if the user explicitly ends the chat session.

Run query get_tool_data to see the data the tool has collected so far.

API

• poetry run uvicorn api.main:app --reload to start the API server.
• Access the API at /docs to see the available endpoints.

UI

• cd frontend
• npm install to install the dependencies.
• npm run dev to start the dev server.

Customizing the agent

• tool_registry.py contains the mapping of tool names to tool definitions (so the AI understands how to use them)
• The tools themselves are defined in their own files in /tools
• Note the mapping in tools/__init__.py to each tool
• See main.py where some tool-specific logic is defined (todo, move this to the tool definition)

TODO

• Code GenerateInvoice against the Stripe API
• I should prove this out with other tool definitions outside of the event/flight search case (take advantage of my nice DSL).
• Currently hardcoded to the Temporal dev server at localhost:7233. Need to support options incl Temporal Cloud.
• UI: Make prettier
• UI: Tool Confirmed state could be better represented