← Back to Blogs
Skip to main content

Improving Text Embeddings with Large Language Models

· 2 min read
Zain Hasan

A preview of the paper

❓Your RAG workflow is only as good as the retrieved context. Can you use LLMs to improve recall and search relevance for dense retrievers?🤔

📜Work from Microsoft uses synthetic data + LLMs as embedding models to achieve SOTA on the MTEB benchmark.

⏩In Short:

  1. They generate a multilingual synthetic retrieval dataset using GPT-4 which includes {queries, positive matches, hard negatives}.

  2. They use this synthetic dataset along with 13 other public datasets and embed the queries & positives/negatives using the last layer vectors of Mistral-7B.

  3. They tune the Mistral-7B embedding model using a contrastive loss along with embeddings from step 2.

  4. Using this fine-tuned Mistral-7B as an embedding model then achieves SOTA(+2.4%) on MTEB.

img1

❌ Limitations/Short-coming: Potential data contamination? Didn't Mistral-7B have access to all MTEB benchmark datasets? - The MTEB was released(Oct 2022) before the training cutoff(~2023) of the model. So there might be some contamination since we are using the LLM to embed this same data.

  • maybe I just don't understand!🤷

📑The details:

  1. The synthetic data generation consists of prompting GPT-4 to brainstorm a list of potential retrieval tasks followed by getting GPT-4 to generate (query, positive, hard negative) triplets for each task.

img2

  1. This synthetic data captures text embedding tasks in 93 languages, covering 1000's of embedding tasks. - see prompt templates for how this diversity is obtained and the different tasks that are generated

img3

  1. Fine-tune Mistral-7B using standard contrastive loss with a temperature-scaled cosine similarity with LoRA with rank 16.

  2. For Mistral-7B based models, contrastive pre-training has negligible impact on the model quality. This is surprising since it's one of the key factors behind the success of existing text embedding models.

img4

🚀Pretty cool work overall, anytime an LLM augments a well studied field and beats SOTA it's exciting times! They show that language modeling and text embeddings are two sides of the same coin.🪙

Given an embedding task prompt template, a robust LLM should be able to generate training data and then be transformed into an embedding model through light-weight fine-tuning.

🔗 arXiv Link

📜 Download paper

Ready to start building?​

Check out the Quickstart tutorial, or build amazing apps with a free trial of Weaviate Cloud (WCD).

GitHub

Forum

Slack

X (Twitter)

Don't want to miss another blog post?

Sign up for our bi-weekly newsletter to stay updated!


By submitting, I agree to the Terms of Service and Privacy Policy.