- A user-friendly web data application that allows users to ask questions about the data.
- Python 3.9+
- With or without virtualenv, run
pip install -r requirements.txt. - If connection to SQL database is required, also run
pip install -r db_requirements.txt. (tested with Postgresql)
- Start application with
streamlit run main.py - (if using csv data) Upload the input csv. The encoding might need to be manually configured.
- (if using Postgresql) Set the connection details in the provided text boxes.
- Or set the default values in constants.py
- Different LLM models have different capabilities. This is set in constants.py
- gpt-4 may be the most capable overall
- gpt-3.5-turbo-16k may be needed to increase token limit
- Temperature of 0 or close to it is advised so model doesn't make things up.
- After uploading csv or establishing connection, previews will be available in preview tab.
- Enter anything in the chat input box at the bottom.
- Chat history is stored in the app, but not sent to the LLM, this is done considering token limit.
- Each new prompt is a new LLM conversation.
- Clear chat history button is at the bottom on the chat tab.
- Streamlit
- Streamlit's primary design is that it always reruns code from top to bottom with any user interaction.
- It used to be stateless too, meaning everything had to be reloaded and re-computed each time.
- Statefulness has since been added to the tool.
- But one still needs to be careful to design it to not redo expensive computations or API call.
- Streamlit's primary design is that it always reruns code from top to bottom with any user interaction.
- Overall Design
- Although probably unnecessary, the app is designed to not persist any unnecessary object in memory.
- So LLM and class objects are instantiated to perform its functions, then recycled away, functional programming style.
- Although probably unnecessary, the app is designed to not persist any unnecessary object in memory.
- NaturalQuery
- Though the Streamlit app calls it by creating new recyclable instances for each call, the Class is designed so that persisting Class Instance is a valid usage too.
- So LLM engine is a Class attribute.
- 2 methods
- df_run connects to the inputted pandas.DataFrame object
- db_run connects to the pre-connected Database connection
- Though the Streamlit app calls it by creating new recyclable instances for each call, the Class is designed so that persisting Class Instance is a valid usage too.
- If csv is poorly formatted, the app will fail at csv upload.
- LLM is not yet involved in csv load, which still uses
pd.read_csv()
- LLM is not yet involved in csv load, which still uses
- LLM may be capable of interpreting column names and data, or may not.
- Getting accurate answers or any answer at all relies on the capabilities of the LLM.
- Sometimes rewording the question can get the correct answer if the original one cannot.
- Any database system other than Postgres was not tested, and even Postgres was tested only on my configurations.