This is the first reranker that can follow custom instructions about how to rank retrievals (e.g. recency, source, metadata, etc.). Our reranker is the most accurate in the world, outperforming competitors by large margins on the industry-standard benchmarks.
We've built something we think is pretty cool—a reranker that can follow natural language instructions about how to rank retrieved documents. To our knowledge, it's the first of its kind. We’re offering it for free as part of our product launch, and would love for the Product Hunt community to try it and share your feedback!
🎯 The problem we were solving: RAG systems constantly run into conflicting information within the knowledge base. Traditional rerankers only consider relevance, which doesn't always capture use case-specific preferences.
🛠️ What we built: Our reranker lets you specify ranking preferences through instructions like: - "Prioritize recent documents over older ones" - "Prefer PDFs to other sources" - "Give more weight to internal-only documents"
This means your RAG system can now make prioritization decisions based on criteria that matter to you, not just semantic similarity.
📊 Performance details: We've tested it extensively against other rerankers on the BEIR benchmark and our own customer datasets. It achieves state-of-the-art performance and consistently outperforms alternatives. The performance improvement was particularly noticeable when dealing with ambiguous queries or conflicting information sources.
👉 If you want to try it: We've made the reranker available through a simple API. You can start experimenting with the first 50M tokens for free by creating an account and using the /rerank endpoint.
There's also documentation for the API, Python SDK, and Langchain integration:
We've been working on this for a while and would love to hear feedback from folks building RAG systems. What types of instruction capabilities would be most useful to you? Any other ranking problems you're trying to solve?
