The Current State of RRF Querying in Cypher

Oct 13, 2025 by pinkandtrash.com 21 views

Introduction

In the realm of graph databases, querying nodes based on their properties is a fundamental operation. However, when dealing with complex queries that involve multiple indexes, the process can become cumbersome and performance-intensive. This is where RRF (Ranking and Relevance Feedback) semantic querying comes into play. In this article, we will delve into the world of RRF querying, exploring its benefits, and how it can be utilized to unlock deeper access to Lucene instances.

What is RRF Semantic Querying?

RRF semantic querying is a technique that enables you to query nodes based on their properties, taking into account the relevance and ranking of the results. This is achieved by combining multiple indexes, such as full-text and vector indexes, to produce a more accurate and relevant set of results.

The Current State of RRF Querying in Cypher

While it is possible to implement RRF querying in Cypher, the process is often cumbersome and performance-intensive. The current implementation involves creating separate queries for each index, joining the results, and then computing the RRF score. This can lead to complex and inefficient queries that can impact the overall performance of your application.

The Ideal RRF Querying API

To overcome the limitations of the current implementation, we propose an ideal RRF querying API that provides a more streamlined and efficient way to query nodes based on their properties. The proposed API, db.index.rrf.queryNodes, would take the following parameters:

indexNameList: a list of index names to query
queryString: the query string to search for
queryVector: the query vector to search for
options: an optional parameter to configure the RRF calculation, such as weights for a weighted RRF calculation

The API would return a list of nodes, along with their corresponding RRF scores.

Benefits of the Proposed API

The proposed API offers several benefits over the current implementation:

Simplified querying: the API provides a more streamlined way to query nodes based on their properties, eliminating the need for complex and inefficient queries.
Improved performance: the API is designed to take advantage of the underlying Lucene instances, providing faster and more accurate results.
Configurability: the API allows for configuration of the RRF calculation, enabling you to fine-tune the results to suit your specific use case.

Example Use Case

To illustrate the benefits of the proposed API, let's consider an example use case. Suppose we have a graph database containing information about users, with each user having a profile that includes a full-text description and a vector representation of their interests. We want to query the database to find users who match a specific query string and vector.

Using the proposed API, we can query the database as follows:

CALL db.index.rrf.queryNodes(["user_description", "user_interests"], "query_string", [query_vector], {weights: [0.5, 0.5]}) YIELD node, score
RETURN node, score

This query would return a list of users who match the query string and vector, along with their corresponding RRF scores.

Conclusion

RRF semantic querying is a powerful technique that enables you to query nodes based on their properties, taking into account the relevance and ranking of the results. While the current implementation in Cypher is cumbersome and performance-intensive, the proposed API provides a more streamlined and efficient way to query nodes based on their properties. By unlocking deeper access to Lucene instances, the proposed API offers several benefits, including simplified querying, improved performance, and configurability. We hope that this article has provided a comprehensive overview of RRF semantic querying and its benefits, and we look forward to seeing the impact of this technology in the world of graph databases.

RRF Querying in Cypher: A Step-by-Step Guide

While the proposed API provides a more streamlined way to query nodes based on their properties, it is still possible to implement RRF querying in Cypher. Here is a step-by-step guide to implementing RRF querying in Cypher:

Step 1: Define the Query String and Vector

First, we need to define the query string and vector that we want to search for. We can do this using the following Cypher query:

WITH "sample user query" as query,
     "some_text_index" as text_index,
     "some_vector_index" as vector_index,
     40 as k

Step 2: Perform a Full-Text Search

Next, we need to perform a full-text search on the text index. We can do this using the following Cypher query:

CALL (query, text_index, k) {
    CALL db.index.fulltext.queryNodes(text_index, query, {limit: k}) YIELD node, score
    ORDER BY score DESC
    WITH collect(node) as nodes
    UNWIND range(0, size(nodes) - 1) AS idx
    RETURN [idx, "text_search"] as rank, nodes[idx] as node
}

Step 3: Perform a Vector Search

Next, we need to perform a vector search on the vector index. We can do this using the following Cypher query:

CALL (query, vector_index, k) {
    WITH [query] AS batch
    CALL genai.vector.encodeBatch(batch, 'OpenAI', { token: $openAiKey }) YIELD index, vector
    CALL db.index.vector.queryNodes(vector_index, k, vector) YIELD node, score
    WITH collect(node) as nodes
    UNWIND range(0, size(nodes) - 1) AS idx
    RETURN [idx, "vec_search"] as rank, nodes[idx] as node
}

Step 4: Join and Process Results

Finally, we need to join and process the results of the full-text and vector searches. We can do this using the following Cypher query:

UNWIND (text_res + vec_res) AS pair
WITH pair[0][0] AS rawRank, pair[0][1] AS source, pair[1] AS node1
WITH  node1, collect(rawRank) AS ranks

// 4. Compute RRF score
RETURN 
  reduce(
    sumScore = 0.0, 
    r IN ranks | sumScore + 1.0 / (60 + r) //hardcoded rrf const to 60
  ) AS rrf,
  node1
ORDER BY rrf DESC

Introduction

In our previous article, we explored the concept of RRF (Ranking and Relevance Feedback) semantic querying and its benefits in unlocking deeper access to Lucene instances. We also provided a step-by-step guide to implementing RRF querying in Cypher. In this article, we will answer some of the most frequently asked questions about RRF semantic querying.

Q: What is RRF semantic querying?

A: RRF semantic querying is a technique that enables you to query nodes based on their properties, taking into account the relevance and ranking of the results. This is achieved by combining multiple indexes, such as full-text and vector indexes, to produce a more accurate and relevant set of results.

Q: What are the benefits of RRF semantic querying?

A: The benefits of RRF semantic querying include:

Simplified querying: RRF semantic querying provides a more streamlined way to query nodes based on their properties, eliminating the need for complex and inefficient queries.
Improved performance: RRF semantic querying is designed to take advantage of the underlying Lucene instances, providing faster and more accurate results.
Configurability: RRF semantic querying allows for configuration of the RRF calculation, enabling you to fine-tune the results to suit your specific use case.

Q: How does RRF semantic querying work?

A: RRF semantic querying works by combining multiple indexes, such as full-text and vector indexes, to produce a more accurate and relevant set of results. The process involves:

Defining the query string and vector: The query string and vector are defined based on the user's search query.
Performing a full-text search: A full-text search is performed on the text index to retrieve a list of nodes that match the query string.
Performing a vector search: A vector search is performed on the vector index to retrieve a list of nodes that match the query vector.
Joining and processing results: The results of the full-text and vector searches are joined and processed to produce a final list of nodes.
Computing the RRF score: The RRF score is computed based on the relevance and ranking of the results.

Q: What are the challenges of implementing RRF semantic querying?

A: The challenges of implementing RRF semantic querying include:

Complexity: RRF semantic querying involves complex queries and calculations, which can be challenging to implement and optimize.
Performance: RRF semantic querying requires efficient querying and processing of large datasets, which can impact performance.
Configurability: RRF semantic querying requires configuration of the RRF calculation, which can be challenging to fine-tune.

Q: How can I implement RRF semantic querying in my application?

A: To implement RRF semantic querying in your application, you can follow these steps:

Define the query string and vector: Define the query string and vector based on the user's search query.
Perform a full-text search: Perform a full-text search on the text index to retrieve a list of nodes that match the query string.
Perform a vector search: Perform a vector search on the vector index to retrieve a list of nodes that match the query vector.
Join and process results: Join and process the results of the full-text and vector searches to produce a final list of nodes.
Compute the RRF score: Compute the RRF score based on the relevance and ranking of the results.

Q: What are the best practices for implementing RRF semantic querying?

A: The best practices for implementing RRF semantic querying include:

Use efficient querying and processing techniques: Use efficient querying and processing techniques to minimize the impact on performance.
Configure the RRF calculation: Configure the RRF calculation to fine-tune the results to suit your specific use case.
Monitor and optimize performance: Monitor and optimize performance to ensure that the RRF semantic querying implementation is efficient and effective.

Conclusion

RRF semantic querying is a powerful technique that enables you to query nodes based on their properties, taking into account the relevance and ranking of the results. By understanding the benefits, challenges, and best practices of RRF semantic querying, you can implement this technique in your application to improve performance, configurability, and accuracy.