Query Data from Fit or CSV Using Natural Language. Enhance Analysis with an Ability to Query Relevant Research Papers and the Google Search API for Contextual Formulas and Reasoning.

This repository contains a code for FastAPI app that utilize OpenAI's GPT-powered API to query and generate answers for a given prompt in natural language. The script exposes Natural Language Query Endpoint: /nat_lang_query that allows users to generate answers using OpenAI's API.

Project Workflow: This application is designed to process Garmin FIT files and analyze them alongside relevant research papers. The workflow consists of the following steps:

• Ingest the supplied Garmin FIT file.
• Parse the FIT file and store the data in a PostgreSQL database tables.
• Split the supplied PDF research paper into chunks and store them in an in-memory Chroma vector database.
• Accept user prompts in English to specify the desired analysis.
• Query the vector database to find matches related to the user prompt.
• Query Google Search using Serper API to find matches related to the user prompt.
• Query the Pandas DataFrame, which is generated from the PostreSQL database query, to retrieve the data specified in the user prompt.
• Send the data from both sources to the LLM (Large Language Model) for further processing and analysis.

This approach enables seamless integration of structured data (Garmin FIT files) and unstructured data (research papers) to provide comprehensive insights and context for coaches and athletes. Furthermore, it enhances analysis by incorporating text from research papers that the LLM might not have encountered during training, resulting in a more robust and informed understanding of the data.

It takes prompt, doc_path, table, activity_id, and model(optional) as inputs, and retrieves the relevant data from the specified table (csv_data or fit_data). The endpoint then creates a pandas DataFrame containing the data, cleans it, and loads the research document specified in the doc_path parameter using the UnstructuredPDFLoader. The document is split into chunks and passed to the Chroma vector database, which is then used to create a RetrievalQA instance.

The below diagrams ilustrates the process of ingestion of the relevant research documents into vector database (Chroma) and the query flow where the data from vector database is combined with the data from pandas dataframe and fed to LLM. Source: https://blog.langchain.dev/tutorial-chatgpt-over-your-data/

• Ingestion:

• Query Flow:

The endpoint combines the DataFrame the Google Search and the PDF document search tools to create a ZeroShotAgent, which is then used to generate the response based on the input prompt. This endpoint provides an efficient way to leverage natural language processing to analyze and extract information from both structured (pandas DataFrames) and unstructured data (pdf documents). A sample research paper that can be used for experimentation is included in the repo (Heart_Rate_Running_Speed_Index_May_Be_an_Efficient.4.pdf).

The package is delivered as a FastAPI web app and requires an OpenAI API key, and a PostgreSQL database to function. I am using a free tier Supabase but a local instance can also be used. The OpenAI API key needs to be stored in a 'OPENAI_API_KEY' environment variable or in a "openai_api_key" variable (not secure). The Serper API key needs to be stored in a 'SERPER_API_KEY' environment variable or in a "serper_api_key" variable (not secure).The required python libraries are listed in the requirements.txt file.

The script will install the DB schema into an existing blank database upon the first execution. Subsequently the user will need to be creaded using FastAPI /signup endpoint and loged in using the /login endpoint. The fit file is supplied for parsing and data storage using /fit_activities endpoint. A bulk upload is also available via /bulk_upload_fit endpoint suplying a zip file containing multiple fit files. A sample fit files to play around with are included in the repo.

Lot of work to be done yet, but the initial results are encouraging :-)

Requirements:

OPENAI_API_KEY as an environmental variable (Can be obtained from here: https://platform.openai.com/account/api-keys)

SERPER_API_KEY as an Environmental Variable (Can be obtained from here: https://serper.dev/api-key)

Usage:

Start the app: uvicorn main:app --host 0.0.0.0 --port 8000 --reload

FastAPI URL: *http://127.0.0.1/docs

Usage with Docker:

To build the docker image:

make build

To run the application:

make serve

This will run the FastAPI application and a PostgreSQL database using a docker-compose configuration.

Sample outputs:

Prompt (Withouth Google Search): Get the formula for the HR-Running-Speed-Index, then plug in the mean hr and mean speed data from the df, and the standing hr of 56, maximal hr of 182, and the vo2max running speed of 3.5 m/s. Calculate the results and return the HR-Running-Speed-Index value in numeric format together with the formula used.

Prompt (With Google Search): Get the formula for the HR-Running-Speed-Index, then plug in the mean hr and mean speed data from the df, and the standing hr of 56, maximal hr of 182, and the vo2max running speed of 4.5 m/s. Calculate the results and return the HR-Running-Speed-Index value in numeric format together with the formula used.

Prompt: What 3 columns correlate most with the skin_temperature column ? Provide Pearson coefficient in the response.

Prompt: At what time did the highest core temperature occured and for how long ?

Prompt: Print out standard deviation for all columns, and order from highest to lowest

Prompt: What is the average running speed in km/h and total duration of activity in minutes ?

FastAPI endpoint parameters and the corresponding response example