Skip to content

Latest commit

 

History

History
228 lines (136 loc) · 6.11 KB

README.md

File metadata and controls

228 lines (136 loc) · 6.11 KB
Raw

LLM Query Engine

A query engine in the form of a Flask API which implements an LLM on the queries of a Qdrant vector database sourced from the BigBasketProducts dataset.

The architecture of this query engine can be seen in the below representation: image

Table of Contents

Tech Stack

To implement the query engine, the following tech was used:

Embedding: BERT

DB: Qdrant

LLM: T5

API Server: Flask

The preprocessing was done with pandas, followed by a bert embedding to convert it to vectors. These vectors were then stored in the Qdrant database. An instance is hosted on Qdrant cloud. A Flask app API is created to handle get_answer POST requests as sent by a test python script to use the flan-t5-small LLM on the closest vector determined by the DB and the query.

Run Locally

Clone the project:

Create a local copy of this repository on your machine. You can use github cli/ssh keys as some other options!

  git clone https://github.com/rashmigr01/llm-query-engine.git

Go to the project directory:

Switch to the installed directory with a simple cd command.

  cd llm-query-engine

Create a virtual environment:

This step is recommended when you plan to run multiple projects on a system to organize package versions and requirements. However, you can choose to skip it if you are confident otherwise. Read more here.

  python -m venv venv
  source venv/bin/activate

Install dependencies:

The requirements.txt file contains the necessary dependencies. Install the packages to your system with the following command.

  pip install -r requirements.txt

Start the API:

To host the API on a local server, first navigate to the API folder with this command.

  cd API

Begin the local Flask API with this command:

  python3 api.py

You can now interact with the API by running the following program:

  python3 test_api.py "Your query goes here!"

If you choose to not enter a question, a default question is chosen. To query the API, make sure to pass a string question in the above command. The JSON answer is printed in the terminal!

Development Operations

After the above local installation, you can perform development operations to generate intermediate files of this application.

To preprocess the BigBasketProducts.csv dataset, navigate to the Preprocess directory. Run the preprocess script to generate a CSV file and a pickled version in the embeddings directory.

  cd ../Preprocess
  python3 preprocess.py

To generate vector embeddings, navigate to the Embeddings directory. Run the embeddings script, follwed by the modify_csv script to get the embeddings and their ids. Run the embeddings_json script to generate the main JSON embeddings file in the qdrant directory.

  cd ../Embeddings
  python3 generate_embeddings.py
  python3 modify_csv.py
  python3 embeddings_json.py

To generate the required metadata which will serve as the payload, navigate to the Metadata directory. Run the script to get the JSON file in the qdrant directory.

  cd ../Metadata
  python3 generate_metadata.py

To operate on the Qdrant database, navigate to the Qdrant directory. Run the qdrant_interaction script to create an index and upsert vectors in batches.

  cd ../Qdrant
  python3 qdrant_interaction.py

The changes made to the cloud hosted Qdrant DB will reflect in the Flask API calls to generate different query answers.

Examples

Here are some query examples as output by the Flask API:

Test 1

python3 test_api.py "What is the cost of orange juice?"
{'answer': '306.0'}

Test 2

python3 test_api.py "Category of coconut oil"
{'answer': '"sub_category": "Edible Oils & Ghee"'}

Test 3

python3 test_api.py "What is the rating of Rich Creme Hair Colour"
{'answer': '4.3'}

Test 4

python3 test_api.py "What is the market price of Mountain Dew?"
{'answer': '105.84'}

Query 5

python3 test_api.py "Brand of Jeera Powder?"
{'answer': '"brand": "Dabur"'}

Roadmap

The following steps were taken to complete this project:

  • Preprocess the CSV dataset to remove spaces, convert to lower case string and handle numeric data.

  • Generate vector embeddings using the BERT model to convert these strings to float vectors.

  • Add indices to get unique ids for each of the embeddings.

  • Generate a metadata JSON file and an embeddings JSON file.

  • Create Qdrant DB and upsert vectors to the cloud cluster.

  • Create a Flask API mask and a test script to interact with the API.

  • Implement a T5 LLM on the search query to the Qdrant DB with the question and the question to generate the API response.

  • Refactor codebase, add env variables, add readme.

Environment Variables

The following environment variables are used and stored in the .env file. It is currently included in the repository for submission purposes but would otherwise be added to gitignore.

QDRANT_URL - The URL to the cloud hosted vector DB.

QDRANT_API_KEY - The API access key to the Qdrant DB.

Author

This query engine was built by me, @rashmigr01, a senior undergraduate pursuing a B. Tech in CSE at IIT Kanpur. Please feel free to reach out to me with any queries at rashmigr20@iitk.ac.in.

Acknowledgements

This project wouldn't be possible without the wonderful opportunity by Chaabi. Through brainstorming and exploration, I learnt more than I ever expected all while having lots of fun!