Authors: Philippe Laban, Jesse Vig, Marti A. Hearst, Caiming Xiong, Chien-Sheng Wu
Paper Link: Beyond the Chat: Executable and Verifiable Text-Editing with LLMs
Contact: [email protected]
- Install the requirements.txt
pip install -r requirements.txt
-
(Optional) Modify the app's parameters, defined in the
app_params.py. This can be used to enable/disable certain components of the app, and set which LLM to use to serve edit recommendations from. -
Set an environment variable for
OPENAI_API_KEYwhich will be used to bill for the LLM API usage. You can get an API key from OpenAI. It should be set in the environment before launching the app. -
Launch the app! For development purposes, you can simply launch the app this way:
python wsgiis.py
This is not stable for a "production environment" (or a long-term usability study), but can help during development.
For a more stable deployment, you can use NGINX & Gunicorn, and start the app with the following command:
gunicorn --timeout 60 --bind unix:/path/to/inksync/app.sock -m 777 -w 5 wsgiis:app --reload
This requires more advanced setup and configuration on a server, but is more stable and secure.
- Head to the app's URL in the browser:
By default, typically: http://127.0.0.1:5000/
app_*files contain the files to the backend (Flask) of the app.app_inksync.pyis the main file which defines all the backend routes.app_params.pyhas all basic parameters to enable/disable components without having to get deep in the code.
prompts/folder contains all the prompts used by LLM-based components of the app.model_recommender.pyhas all the LLM-related code, dealing with receiving inputs, and generating/validating edit suggestions by the LLM.- The front end is distributed in:
static/folder contains all the static files (CSS, JS, images, etc.)templates/folder contains all the HTML files.
documents/folder contains the full trace of documents edited by users. Each document corresponds to a JSON file in there in the format:user_<USERID>_<UNIQUE_DOC_ID>.jsonlogs/folder contains several important log files useful for analysis:api_log.jsonlcontains all the API calls made to the app.llm_stats_log.jsonlcontains high-level information about LLM-queries made: which component, the number of tokens, the cost (USD), and the latency.query_log.jsonlcontains the exact queries that were sent to the LLM, for debugging purposes.
Read the paper, listed above for details on the system and implementation.