Text Embeddings in Elasticsearch

This repository provides a simple example of how Elasticsearch can be used for similarity search by combining a sentence embedding model with the dense_vector field type.

Important note: Using text embeddings in search is a complex and evolving area. We hope this example gives a jumping off point for exploration, but it is not a recommendation for a particular search architecture or implementation.

Description

The main script src/main.py indexes the first ~20,000 questions from the StackOverflow dataset. Before indexing, each post's title is run through a pre-trained sentence embedding to produce a dense_vector.

After indexing, the script accepts free-text queries in a loop ("Enter query: ..."). The text is run through the same sentence embedding to produce a vector, then used to search for similar questions through cosine similarity.

Currently Google's Universal Sentence Encoder is used to perform the vector embedding. This is a fully pre-trained model, and no 'fine tuning' is performed on the StackOverflow dataset.

Usage

Make sure that pip and python installed (Python version 3), then install the script's dependencies:

pip3 install -r requirements.txt

The script assumes that a local Elasticsearch node is running and able to connect. Instructions on how to download and run Elasticsearch can be found here. Note that Elasticsearch 7.3 or higher is required in order to use the vector functions, and that the cluster must be running the default (not oss) distribution.

Finally, the script can be run through

python3 src/main.py

Example Queries

The following queries return good posts near the top position, despite there not being strong term overlap between the query and document title:

• "zipping up files" returns "Compressing / Decompressing Folders & Files"
• "determine if something is an IP" returns "How do you tell whether a string is an IP or a hostname"
• "translate bytes to doubles" returns "Convert Bytes to Floating Point Numbers in Python"

Note that in other cases, the results can be noisy and unintuitive. For example, "zipping up files" also assigns high scores to "Partial .csproj Files" and "How to avoid .pyc files?".

Running from Docker

Instead of downloading Elasticsearch and cloning this repository, you can run the following commands to download and run from a Docker container:

docker run --name text_embeddings  -p 9200:9200 -p 9300:9300 -e "discovery.type=single-node"  -d shantanuo/textembeddings
docker exec -it text_embeddings bash
cd text-embeddings/
python3.6 src/main.py