sever3d/hugginface-embedding-api repository overview

⁠Embeddings & Vector Search

With Python, HuggingFace and MongoDB Atlas.

⁠Running app manually

flask run --host=0.0.0.0 --port=9050

⁠What is going on ?

Requirements, dreams & desires: I am at the stage of building my MSC Advanced Computing project. It is distributed app for education with some whistles, fountains and pretty flowers (but not really). I wanted to create Question/Post suggesting system. I could build that bases on Question/Post (Q/P) tags, but that isn't very accurate, modern and hot. I wanted to use my (still poor) existing NLP skills and use something much modern that SQL query to find Questions/Posts by tags.

Solution (?) : Well ... I ended up (by now) with an attempt of using word embeddings and vector search to find related Q/P.

How it works (or how it should work): On the UI side of adding a Q/P, when one is being added the user has to type Q/P title (summary) and then he goes to the body of the question. The Q/P title is being sent (in the background) to this API here to the '/embedding/search endpoint. This endpoint creates the Q/P title embedding and searches for similar ones in the MongoDB created for this purpose.

Other possibilities:

• Q/P title can be sent to the /embedding endpoint just to get the embeddings
• Q/P can be sent to /embedding/mongostore endpoint to store it's embedding in MongoDB

⁠MongoDB Atlas Index

Index Name: QuestionSematicSearchIndex

Pick the correct db and collection and set the index JSON as below:

{ "mappings": { "dynamic": true, "fields": { "question_embedding": { "dimensions": 384, "similarity": "dotProduct", "type": "knnVector" } } } }

⁠Environment Variables & Settings

Create .env file. This file will hold all environment variables required for docker-compose.yml file.

It should contain below variables:

# Env variables for docker-compose
HF_TOKEN= 
API_TOKEN=
MONDO_CONN_STRING=
SEARCH_LIMITS_COUNT=

• HF_TOKEN - HuggingFace token (I got one but free version allow to make few FREE embeddings)
• API_TOKEN - token to secure api. One is set when running the Docker container based on the image. Create your own and set and docker image env.var or env var in your system if you're running app manually. Example: export API_TOKEN=hf_myHaPpYToKeNgOesHeRe or set API_TOKEN=hf_myHaPpYToKeNgOesHeRe Same should be done for HF_TOKEN.
• SEARCH_LIMITS_COUNT - how many MongDB records should be displayed in '/embedding/search' results'

When making the API request , set header x-access-tokens to match your API_TOKEN.

⁠How to run it ?

1. Add your values in .env file . For Hugging Face token, make a free account, go to Your Profile -> Settings -> Access Tokens -> New Token
2. MongoDB Atlas connection string - Create free MongoDB Atlas account. Create Cluster. Create Database. Click Connect and copy connection string.
3. For API_TOKEN - create any password you want, but use the same in you API request headers as mentioned above.
4. Run command docker compose up -d (you must have Docker and Compose in your system)