elena-kalinina/legal_agent

legal_agent

General info

The legal agent is an LLM that can analyze contracts and respond if a certain requests is in compliance with the contract clauses. The LLM is Mixtral MoE 8x7B Instruct v1 available through Nvidia API (https://build.nvidia.com/mistralai/mixtral-8x7b-instruct). The system takes the contract text, splits it into paragraph, embeds each paragraph and indexes them with FAISS index. Embeddings model used is embded-qa-4, available through Nvidia API (https://build.nvidia.com/nvidia/embed-qa-4). You can provide queries to the system as rows taken from the task file, e.g.: 'Team building event in Colorado;$2,200'. The system will analyze its compliance with the retrieved relevant contract clauses and provide the response. In case of ambiguity, when it is unclear whether the task conforms or contradicts the contract stipulations, "human in the loop" is introduced: you can provide follow up clarifications and explanations to the system to help it come up with the conclusion.

Disclaimer

Because of time pressure and constraints concerning the number of free credits for Nvidia API, the system was tested with a small portion of the contract (file contract.txt). Conversation sequence it was tested for:

• Conference in New York with weekend travel; $3,000
• The flight to New York was booked in business class
• Strategy retreat in the Swiss Alps during winter;$2,400

Also, a few shortcurs were used, e.g.: the file is not uploaded in the app, but the filename is given. Finally, whe extracting key terms in a json, the LLM sometimes produced unparsable json outputs. While there are tools to make the LLMs output better jsons, for the current application they were not used because of time pressure.

Installation and requirements

• Install requirements:

    pip install -r requirements.txt      

• To run the app, you will need an API key from Nvidia. Register an account with Nvidia (https://build.nvidia.com/explore/discover) and go to models (https://build.nvidia.com/explore/reasoning). Click on any model of your choice (e.g. https://build.nvidia.com/meta/llama3-70b) and choose Get API key. Copy the key and export it as and environment variable, e.g.:

    export NVIDIA_API_KEY=YOUR_KEY      

or

    set NVIDIA_API_KEY=YOUR_KEY      

Also, add your key to the config.yml file:

api:
    - nvidia_key: YOUR_KEY 

Test the system

• Run the app:

   python app.py      

The app will run on localhost:5450. To reproduce the results, indicate contract.txt as file name and introduce queries as shown above, e.g. Task name + price in USD separated by semicolon: 'Team building event in Colorado;$2,200'. Follow-up clarifications can be added as text.

• While task analysis is done through the conversational interface, the system can also extract key terms from each paragraph. To perform key term extraction, just put the contract file in the root directory as contract.txt and run the file contract_agent.py:

   python contract_agent.py      

The extracted keywords will be saved in the file named topics.json in the root directory.