• Written by Azhar Aliyar, Director

A purchasing manager checks the inventory levels for a specific product, and the query results from AsscherAi include a list of relevant passages from the purchase order table, the inventory management system, and the supplier contracts database. To ensure the most relevant information is presented first, the system uses a process called reranking, which is a second pass that reorders the retrieved passages to prioritize the most relevant ones.

Definition of Reranking

Reranking is a process that takes the initial set of retrieved passages and reorders them based on their relevance to the query, using a different set of ranking criteria or weights. This process is necessary because the initial retrieval is often based on a simple keyword search, which may not capture the nuances of the query. Reranking allows the system to consider additional factors, such as the context of the query, the relationships between the different data sources, and the specific requirements of the user.

The reranking process involves re-scoring the retrieved passages using a more sophisticated algorithm, which takes into account the specific characteristics of the query and the data sources. This algorithm may use techniques such as natural language processing, entity recognition, and semantic analysis to better understand the context and intent of the query.

Difference from Retrieval

Reranking is often confused with the initial retrieval process, but it is a distinct and separate step. Retrieval is the process of finding the relevant passages from the various data sources, whereas reranking is the process of reordering those passages to prioritize the most relevant ones. While retrieval is focused on finding the relevant information, reranking is focused on presenting that information in the most useful and relevant way.

The key difference between retrieval and reranking is the level of analysis and processing involved. Retrieval is typically a simpler process that involves keyword searching and basic filtering, whereas reranking involves more advanced analysis and processing to understand the nuances of the query and the relationships between the data sources.

Applications in a Large Organisation

Reranking has a number of applications in a large organisation, particularly in situations where the query results need to be presented in a specific order or where the user needs to quickly identify the most relevant information. For example, a shift supervisor may use AsscherAi to query the maintenance log and retrieve a list of relevant passages, which can then be reranked to prioritize the most critical maintenance issues.

In another example, the category manager may use AsscherAi to compare two suppliers and retrieve a list of relevant passages from the supplier contracts database, which can then be reranked to prioritize the most relevant information about each supplier. In both cases, the reranking process helps to present the information in a more useful and relevant way, allowing the user to quickly identify the most important information and make informed decisions.

For more information on how AsscherAi can be used in a large organisation, please visit our website.

Limitations of Reranking

While reranking is a powerful tool for presenting query results in a more useful and relevant way, it is not a magic solution that can solve all the problems of information retrieval. Reranking is only as good as the initial retrieval process, and if the retrieval process is poor, the reranking process will not be able to compensate. Additionally, reranking requires a good understanding of the query and the data sources, as well as a sophisticated algorithm for re-scoring the passages.

Most people get wrong the idea that reranking is a simple process that can be done automatically without any human intervention. However, this is not the case, and reranking often requires careful tuning and adjustment to ensure that the results are accurate and relevant. For more information on how to implement reranking in your organisation, please contact us.

When Reranking is Not the Right Choice

Reranking is not always the right choice, particularly in situations where the query results are simple and straightforward, or where the user is looking for a specific piece of information. In these cases, a simple retrieval process may be sufficient, and reranking may add unnecessary complexity and overhead. Additionally, reranking may not be suitable for very large datasets, where the computational resources required for reranking may be prohibitive.

In general, reranking is most useful in situations where the query results need to be presented in a specific order, or where the user needs to quickly identify the most relevant information. However, the decision to use reranking should be based on a careful analysis of the specific requirements of the user and the characteristics of the data sources.