CALM Reimplementation of LangGraph's customer service bot

This is a reimplementation of langgraph's customer support example in Rasa's CALM paradigm There's a YouTube video that provides a walkthrough of langgraph's implementation.

Skills

The bot has the following skills:

• Showing the user's booked flights
• Changing flight bookings
• Booking a rental car
• Booking a hotel
• Booking an excursion (e.g. museum visit) on the trip

Running in a Github Codespace

Follow these steps to set up and run the Rasa assistant in a GitHub Codespace.

Prerequisites

• You'll need a Rasa Pro license and an OpenAI API key.
• You should be familiar with Python and Bash.

Steps to run the CALM assistant

To run the CALM assistant, you can watch the video below and/or follow the instructions:

CALM-assistant-setup-codespace.mp4

1. Create a Codespace:

   • Navigate to the repository on GitHub.
   • Click on the green "Code" button, then scroll down to "Codespaces".
   • Click on "Create codespace on main branch".
   • This should take under two minutes to load.

2. Set Up Environment:

   • Once the Codespace loads, it will look like VSCode but in your browser!
   • First, open two terminal windows.
   • In both terminals, run source .venv/bin/activate
   • Open /calm_llm/.env file and add the required keys to that file.

     export RASA_PRO_LICENSE='your_rasa_pro_license_key_here'
     export OPENAI_API_KEY='your_openai_api_key_here'
     

   • In both terminals, set your environment variables by running:

     source calm_llm/.env
     

3. Create the Database:

   • In one of the terminals, run the command to create the database: python scripts/create_db.py.

4. Train the Model:

   • In the first terminal, cd calm_llm and then run:

     rasa train
     

5. Start the Action Server:

   • In the second terminal, cd calm_llm and then run:

     rasa run actions
     

6. Launch the Rasa Inspector:

   • In the first terminal, run:

     rasa inspect --debug
     

7. Access the Inspector:

   • When prompted to open in browser, click the link.

8. Chat with your customer support assistant about flights, hotels, cars, and/or excursions!

Notes

• Keyboard bindings may not map correctly in the Codespace, so you may not be able to copy and paste as you normally would!
• The database creation is done separately to manage memory usage.
• The repository is compatible with Rasa Pro versions >=3.10.0.
• You'll also notice that there are several subdirectories: calm_llm is the CALM implementation, calm_nlu combines CALM with intent based NLU, langgraph_implementation is the implementation inspired from langgraph's tutorial, calm_self_hosted is the CALM implementation but a fine-tuned model such as Llama 3.1 8B working as the command generator, and calm_nlu_self_hosted is CALM working with intent based NLU and a fine-tuned model as the command generator.

Quantitative Evaluation

We provide scripts to evaluate the assistant on 3 measures:

• number of tokens used per user turn (proxy for measuring LLM cost per user turn)
• latency (time to get a response)
• accuracy

To do so, we construct a test set to evaluate the following capabilities:

• Happy paths - Conversations with minimal complexity sticking to one skill.
• Slot corrections - Conversations where a user changes their mind in between and corrects themselves.
• Context switches - Conversations with a switch from one skill to another and coming back to the former skill.
• Cancellations - Conversations where the user decides to not proceed with the skill and stops midway.
• Multi Skill - Conversations where the user tries to accomplish multiple skills one after the other.

Run end-to-end tests

Ensure you have set up the environment in two active terminals by following the instructions in this section

Next, execute the following in the calm_llm directory:

1. Run action server - In one of the terminals, execute the command -

rasa run actions

2. While the action server is running, in the second terminal execute:

MAX_NUMBER_OF_PREDICTIONS=50 python run_eval.py

This will print the results to your terminal. You can also pipe the results to a text file MAX_NUMBER_OF_PREDICTIONS=50 python run_eval.py > results.txt.

Once the script finishes you will see runtime stats on input and output tokens consumed and latency incurred. These stats are grouped by the folder which contained the tests -

Running tests from ./e2e_tests/happy_paths
=============================
COST PER USER MESSAGE (USD)
---------------------------------
Mean: 0.031122631578947374
Min: 0.026789999999999998
Max: 0.038040000000000004
Median: 0.03162
---------------------------------

COMPLETION TOKENS PER USER MESSAGE
---------------------------------
Mean: 10.368421052631579
Min: 6
Max: 26
Median: 9.0
---------------------------------

PROMPT TOKENS PER USER MESSAGE
---------------------------------
Mean: 1016.6842105263158
Min: 881
Max: 1248
Median: 1021.0
---------------------------------

LATENCY PER USER MESSAGE (sec)
---------------------------------
Mean: 2.567301022379022
Min: 1.5348889827728271
Max: 4.782747983932495
Median: 2.067293882369995
---------------------------------

============================================================== short test summary info ===============================================================
================================================================= 0 failed, 5 passed
==================================================================

LangGraph assistant

Navigate to langgraph_implementation folder and then set up the environment with -

# Step 1: Create a new virtual environment
python -m venv new_env

# Step 2: Activate the virtual environment
source new_env/bin/activate

# Step 3: Install the packages from requirements.txt
pip install -r requirements.txt

Next, set up the necessary keys by opening the .env file in that folder and filling the values for requested variables

TAVILY_API_KEY - Access key for Tavily, used for making search queries
LANGCHAIN_API_KEY - Langsmith access key, for monitoring and tracing LLM calls.
OPENAI_API_KEY - API key for OpenAI platform, for invoking the LLM.

Load the keys by running source .env in the terminal window.

Then execute -

python run_eval.py

This will print the results to your terminal. You can also pipe the results to a text file python run_eval.py > results.txt.

Run tests to create figures

To create the figures in our blog post

1. Generate data for CALM assistant

• Follow steps 2-5 from Steps to run CALM assistant section.
• On a separate terminal, navigate to calm_llm directory, run python run_tests_for_plots.py to generate data for figures.
• Restructure the data for plotting with cd results and then python combine_data.py

2. Generate data for CALM + NLU assistant

• Follow steps 2-5 from Steps to run CALM assistant section but in Steps 4 and 5, cd calm_nlu instead of cd calm_llm.
• On a separate terminal, navigate to calm_nlu directory, run python run_tests_for_plots.py to generate data for figures
• Restructure the data for plotting with cd results and then python combine_data.py

3. Generate data for LangGraph assistant

• Run steps 1-5 from LangGraph assistant above
• In langgraph_implementation folder, run python run_tests_for_plots.py to generate data for figures
• Restructure the data for plotting with cd results and then python combine_data.py

Load and visualize results in a Jupyter Notebook

• Open metrics.ipynb (in root directory)
• In the top-right of your screen, you should see 'Select Kernel', click on it
• Once prompted, install necessary extensions
• Once the extensions are installed, click "select Kernel' again and select 'Python Environments...'
• select the .venv environment for running the kernel:
• execute all cells!