Today, we're excited to announce that our newest reranking model—pinecone-rerank-v0—is available in early access through Pinecone Inference!

Designed to significantly enhance both enterprise search and retrieval augmented generation (RAG) systems, this new model brings a powerful boost in relevance and accuracy—ensuring that your search results and AI-generated content are grounded in the most relevant and contextually precise information.

Whether you're looking to improve internal search capabilities or strengthen your RAG pipelines, our new reranker is built to meet the demands of modern, large-scale applications while delivering top-tier performance.

Below, we’ll quickly cover:

• Why retrieval quality matters in the context of RAG systems
• How you can get started, quickly and easily, with our new reranker
• The results of putting our new reranker to the test against alternative models

Why RAG matters

Large language models (LLMs) are powerful tools, but they have limitations that can affect response accuracy and relevance in real-world applications. Retrieval-augmented generation (RAG) addresses this by providing the LLM with only the most relevant information, resulting in responses that are grounded in contextually precise, up-to-date data.

LLMs trained on broad datasets can’t directly access proprietary or domain-specific information, which often leads them to generate answers that may sound plausible but lack accuracy. RAG fills this gap by retrieving the right data when it’s needed, so the model’s responses are informed by specific, relevant information. At the same time, more context isn’t always better. Large input windows can lead to information overload, where key details get diluted. By retrieving only what’s essential, RAG keeps responses focused and reduces the “lost in the middle” effect that can impact output quality.

This targeted retrieval approach also brings down token costs—an important factor in production environments where tokens are a primary cost driver. By cutting token usage by 5-10x, RAG makes high-quality responses more scalable and cost-effective. Finally, for RAG to work as intended, retrieval accuracy is key. Purpose-built neural retrieval models ensure that only the most relevant information reaches the model, enabling RAG to deliver responses that meet real-world demands for precision.

Note: For a closer look at how RAG overcomes LLM limitations and optimizes for cost, see our in-depth article on Retrieval Augmented Generation (RAG)

What is a reranker?

In retrieval systems, rerankers add an extra layer of precision to ensure that only the most relevant information reaches the model. After the initial retrieval—where an embedding model and vector database pull a broad set of potentially useful documents—rerankers refine this set by re-evaluating the results with a more sophisticated model. This step sharpens the relevance of the selected documents, so the generative model receives only high-quality, context-rich input.

Within Retrieval-Augmented Generation (RAG) frameworks, this precision is essential. Since the generative model’s output relies on the quality of its input, rerankers help ensure responses that are both accurate and grounded. By combining the broad reach of neural retrieval with the targeted precision of rerankers, RAG systems can deliver more reliable and relevant answers—especially important in applications where precision matters most.

Introducing Pinecone's new reranker

Our latest model, pinecone-rerank-v0, is optimized for precision in reranking tasks using a cross-encoder architecture. Unlike embedding models, cross-encoder processes the query and document together, allowing it to capture fine-grained relevance more effectively.

The model assigns a relevance score from 0 to 1 for each query-document pair, with higher scores indicating a stronger match. To maintain accuracy, we’ve set the model’s maximum context length to 512 tokens—an optimal limit for preserving ranking quality in reranking tasks.

Getting started

The model is accessible through the API and SDKs. Note that this endpoint is intended for development—improved performance can be achieved by contacting us directly for an optimized production deployment.

from pinecone import Pinecone

pc = Pinecone("PINECONE-API-KEY")
query = "Tell me about Apple's products"
documents = [
    "Apple is a popular fruit known for its sweetness and crisp texture.",
    "Apple is known for its innovative products like the iPhone.",
    "Many people enjoy eating apples as a healthy snack.",
    "Apple Inc. has revolutionized the tech industry with its sleek designs and user-friendly interfaces.",
    "An apple a day keeps the doctor away, as the saying goes."
]

results = pc.inference.rerank(
    model="pinecone-rerank-v0",
    query=query,
    documents=documents,
    top_n=3,
    return_documents=True
)

for r in results.data:
    print(r.score, r.document.text)

Putting the reranker to the test

To quantify—as objectively as possible—the improved effectiveness of our new reranker model, we performed an evaluation on several datasets:

• The BEIR benchmark is a widely-used evaluation suite designed to assess retrieval models across various domains and tasks. It includes 18 different datasets covering a range of real-world information retrieval scenarios, allowing models to be tested on diverse challenges like biomedical search, fact-checking, and question answering.
• TREC Deep Learning is an annual competition hosted by the National Institute of Standards and Technology (NIST), focused on web search queries across a diverse web corpus.
• Novel datasets Financebench-RAG and Pinecone-RAG (proprietary) that reflect real-world RAG interactions.

For both TREC Deep Learning and BEIR, we retrieved and reranked 200 documents and finally measured Normalized Discounted Cumulative Gain (specifically, NDCG@10). For Pinecone-RAG and Financebench-RAG, we retrieved and reranked up to 100 documents (some collections were smaller than this) and finally measured Mean Reciprocal Rank (specifically, MRR@10).

MRR@10 and NDCG@10 were calculated with the respective metrics (MRR and NDCG) considering only the top 10 ranked candidates.

We compared with an extremely large set of reranking models, but here we present a selection of those that we consider more competitive:

• Cohere’s latest models are cohere-v3-english and cohere-v3-multilingual, for English and Multilingual reranking spectfully.
• Voyage AI’s generalist reranker optimized for quality with multilingual support named voyageai-rerank-2.
• The model offered by Jina AI built for Agentic RAG named jina-v2-base-multilingual.
• The high performance, multilingual model by Beijing Academy of Artificial Intelligence (BAAI) named .
• google-semantic-ranker-512-003 is the latest reranking model part of Google’s Retrieval Augmented Generation (RAG) experience in Vertex AI Agent Builder.

The models for Cohere, Voyage AI and Google were tested using the official APIs, while the both models from Jina AI and BAAI were hosted locally since they are publicly available.

BEIR

For the BEIR benchmark, we focused on 12 datasets, excluding the four that are not publicly available to ensure reproducibility. We also excluded MS MARCO, as it is evaluated separately, and ArguAna, since its task of finding counter-arguments to the query contrasts with the purpose of a reranker.

• Pinecone-rerank-v0 had the highest average NDCG@10, notably higher than the alternatives
• Pinecone-rerank-v0 performed the best on 6 out of the 12 datasets; the second-best reranking model performed the best on only 3

TREC

For our TREC evaluation, we merged the datasets from the 2019 and 2020 editions, resulting in a total of 97 queries tested against a collection of 8.8 million documents.

Once again, pinecone-rerank-v0 outperformed the other reranking models.

Financebench-RAG

The Financebench-RAG dataset is generated from the open source financebench question-answering dataset. Text is extracted from the raw PDFs with pypdf, and then split into chunks of up to 400 tokens prior to embedding.

The initial retrieval stage is done using openai-large-3 embeddings scored with dot product. The top 100 candidates from the first retrieval stage are then reranked, and chunks are expanded out to the entire page. If multiple chunks from the same page are retrieved, the page’s score is taken to be the average of the chunks on it. These pages are then ranked based on score.

Pinecone-RAG

The Pinecone-RAG dataset reflects real-world RAG interactions and is annotated by Pinecone. This data is not publicly available and was not used in training, thus showcasing 0-shot performance of the model.

After reranking the final candidates with the corresponding reranker, we calculate the standard MRR@10 metric (described above) based on the final rank of the top 10 final candidate documents after reranking.

MRR@10 on Pinecone-RAG

Wrapping up

Unlike building your own foundation model, fine-tuning an existing model, or solely performing prompt engineering, RAG simultaneously addresses recency and context-specific issues cost-effectively and with lower risk than alternative approaches—with the primary purpose of providing context-sensitive, detailed answers to questions that require access to private data to answer correctly.

However, the overall effectiveness of a RAG system is highly influenced by its retrieval model. That’s why we’re so excited to debut pinecone-rerank-v0!

As outlined above, this new reranker demonstrated superior performance when evaluated via both standard datasets and novel datasets—strong indicators that it can help take your LLM-powered applications to the next level.

Pinecone enables you to integrate RAG within minutes. Check out our examples repository on GitHub for runnable examples, and please don’t hesitate to get in touch with us if you’d like to learn more.