SemanTweet Search

SemanTweet Search allows you to search over all your tweets from the Twitter archive using semantic similarity. A demo is available here.

It preprocesses your tweets, generates embeddings using OpenAI's small/large embedding model, stores the data and embeddings in LanceDB vector db, and provides a web interface to search and view the results.

You can do semantic search post pre-filtering by time, likes, retweets, media only or link only tweets too.

Pre-filtering by sql operations helps not only filter but also reduce the vector search space thus speeding up the search.

You can additionally use/edit projector.py and tensorflow projector to get a visualization of your tweets using t-sne algorithm as shown here

UPDATE

• (25/3/2024) bge-small-en-v1.5 embedding support added. No API key required. It's 29th on MTEB leaderboard and 130 MB size, 384 dimension, sequence length 512. Note it's not multi-lingual.

  Technically, a lot of embeddings from sentence-transformers are possible. You can refer LanceDB docs here

• (24/3/2024) Added CLIP based image searching on tweets_media folder. Check code app_image_search.py after installing requirements i.e till step 5. Does not require to run the setup bash scripts for this, it's a standalone program. First run will take roughly 5-10 minutes as it creates the embeddings. Subsequent runs will be instant.

This is also available as a standalone light weight repository as Embeddit where you can use any images folder.

Technologies Used:

• Twitter archive for data
• OpenAI text-embedding-3-large embeddings by default for semantic search
• Lancedb for vector search and sql operations
• Hybrid search provided by Lancedb combining BM25 + embedding search
• Flask for server

Prerequisites

• Python 3.x
• OpenAI API key
• Twitter archive data

Installation

1. Clone the repository:

   git clone https://github.com/sankalp1999/semantweet-search.git
   

   cd semantweet-search/
   

2. Download your Twitter archive (takes 2 days to be available)

   Go to: More (3 dot button) > Settings and Privacy > Your Account > Download an archive of your data.

   Extract it. Put the extracted folder at the root of this project and rename it to twitter-archive.

3. Create a virtual environment:

   python3 -m venv venv
   

   Make sure you do this at the root of project.

4. Activate the virtual environment:

   • For Unix/Linux:

     source venv/bin/activate
     

   • For Windows:

     venv\Scripts\activate
     

5. Install the required dependencies:

   pip install -r requirements.txt
   

   If you want to try out image vector search, please run below command. I have not included this in requirements.txt as it downloads a 620 MB model and not everyone would like to do that by default.

   pip install open_clip_torch
   

6. (Not required if using sentence transformers) Set up your OpenAI API key as an environment variable:

   export OPENAI_API_KEY=your_api_key
   

7. By default, this repo uses openAI text-embedding-3-large

   You can change it to text-embedding-3-small. Change required at two places.

• openai/async_openai_embedding_two.py file around line 11.
• change the MODEL_NAME at openai/create_lance_db_table_openai_three.py around line 8.

8. Run the setup script:

   OpenAI

   chmod +x run_scripts.sh
   ./run_scripts.sh

   sentence-transformers

   chmod +x run_sentence_tf_scripts.sh
   ./run_sentence_tf_scripts.sh

   Uncomment line 15, 16 in app.py

   # db = lancedb.connect("data/bge_embeddings")
   # table = db.open_table("bge_table")

9. Start the application:

   python app.py
   

   or

   flask run
   

Enjoy!

Flow of the program

graph TD
    A[Twitter Archive Data] --> B[preprocess_tweets_one.py]
    B --> C[Preprocessed Tweets CSV]
    C --> D[async_openai_embedding_two.py]
    D --> E[Embeddings CSV]
    E --> F[create_lance_db_table_openai_three.py]
    F --> G[LanceDB Database]
    G --> H[Web Interface]
    H --> I[Search Tweets]
    I --> J[View Results]

Loading

The OpenAI embedding flow consists of the following steps:

1. preprocess_tweets_one.py: This script preprocesses the tweets from the Twitter archive, extracting relevant information and saving it to a CSV file.

2. async_openai_embedding_two.py: This script reads the preprocessed tweets from the CSV file, generates embeddings using OpenAI's embedding model asynchronously, and saves the embeddings to a new CSV file.

3. create_lance_db_table_openai_three.py: This script reads the generated embeddings from the CSV file, creates a LanceDB table using the specified schema, and stores the data in the database.

The run_scripts.sh script automates the execution of these steps in the correct order.

Additional Notes

• The project uses the text-embedding-3-large model by default. You can change the model by modifying the MODEL_NAME variable

• The batch size for generating embeddings is set to 32 to stay within the token limit. Adjust the batch size if needed.

• The LanceDB database is stored in the data/openai_db directory.

• The project also includes a synchronous version of the OpenAI embedding generation script (create_openai_embedding_sync_two.py), which can be used as an alternative to the asynchronous version.

License

This project is licensed under the MIT License. See the LICENSE file for details.