Overview

Reranking is a powerful technique that enhances search systems and RAG (Retrieval-Augmented Generation) setups by intelligently reordering initial search results based on semantic relevance. By integrating reranking into your workflow, you can significantly improve result quality.

What's Reranking, and Why Should You Care?

Imagine you've got a killer search system or a fancy RAG (Retrieval-Augmented Generation) setup. It's doing okay, but sometimes the results are... meh. That's where reranking comes in to save the day!

Reranking is like having a super-smart assistant that takes your initial search results and says, "Hold up, let me sort these in a way that actually makes sense." It considers the semantic meaning of your query and the content of each result, then reshuffles everything to put the most relevant stuff at the top.

How Reranking Fits into Semantic Search and RAG

Here's a quick breakdown of how reranking fits into the bigger picture:

Initial Retrieval: Your system fetches a bunch of potentially relevant documents or chunks.
Reranking Magic: Our reranking models analyze the query and each result, considering their semantic relationship.
Optimized Results: The most relevant content bubbles up to the top, improving the quality of your search or RAG output.

Two-stage search flow including rerank

Two-stage search flow including reranking

For RAG systems, this is huge. By feeding your language model more relevant information, you're essentially giving it a cheat sheet of accurate, contextual data. The result? More accurate responses and fewer hallucinations. It's like putting guardrails on your AI's imagination!

Why Mixedbread Reranking Models?

Open-source: Want to host the models yourself? Go for it!
State-of-the-art: They outperform many closed-source competitors.
Easy to integrate: Use them with existing SDKs, HTTP Clients, or our SDK.
Continuously updated: Our API models are trained on new data every month, ensuring they understand ongoing developments without a knowledge cutoff.

Available Models

We offer several reranking models to fit your specific needs:

Model	Parameter Count	Context Length	API Availability	Key Features
mxbai-rerank-large-v2	1.5B	up to 32K	Available	Highest accuracy, multilingual (100+ languages)
mxbai-rerank-base-v2	0.5B	up to 32K	Unavailable	Balance of size and performance, multilingual
mxbai-rerank-large-v1	435M	512	Available	High accuracy for English text
mxbai-rerank-base-v1	184M	512	Unavailable	Balanced performance for English text
mxbai-rerank-xsmall-v1	70.8M	512	Unavailable	Capacity-efficient while maintaining performance

Our v2 models use reinforcement learning techniques to deliver exceptional performance across benchmarks, with BEIR Accuracy scores of 74.9 for both base and large models.

Quick Start

Here's a quick example using our Reranking API to get you started:

Best Practices for Reranking Success

Diverse Initial Results: Make sure your first-stage retrieval casts a wide net. Reranking can't magically create relevant results if they're not in the initial set.
Experiment with Reranking Depth: Try reranking different numbers of initial results (e.g., top 10, top 50, top 100) to find the sweet spot for your use case.
Combine with Other Techniques: Use reranking alongside other relevance boosting methods like query expansion or faceted search for even better results.
Monitor and Iterate: Keep an eye on your search metrics and user feedback. Reranking might reveal insights about your content or user needs that you can use to further improve your system.
Consider Specialized Use Cases: Our reranking models excel beyond document search, including code snippets, SQL ranking, LLM tool selection, and e-commerce product ranking.

We are currently investigating finetuning and domain adaptation with a limited number of beta-testers. Please contact us if you are interested in using a reranking model tailored to your data.

Ready to take your search and RAG systems to the next level? Dive in, experiment, and let us know how it goes. Happy baking! 🍞🔍