IRerankingService
Overview
The IRerankingService interface delivers reranking functionality to tools and adapters that need to reorder a list of documents—represented as text segments—based on their relevance to a given query. This service enables the sorting of documents according to their similarity to the input query, allowing for more accurate and contextually relevant results within applications using Wisej.AI.
There are fundamentally two approaches to reranking a list of text results in relation to a user query. The first approach is to leverage a general-purpose Large Language Model (LLM) by constructing a prompt, as demonstrated in the built-in features of Wisej.AI (see example below). This method relies on the reasoning and comprehension capabilities of LLMs to assess relevance.
The alternative approach—which is typically faster and more cost-effective—is to utilize a dedicated reranking model, such as a bi-encoder. Bi-encoders are specialized for information retrieval tasks and can efficiently compare query-text pairs to determine the most relevant results.
#
# Prompt for the DefaultRerankingService
#
[DefaultRerankingService.Prompt]
Rerank a list of documents in order of their similarity with a query string, with the most similar on top.
Query: "{{query}}"
Max Results: {{topN}}
## Rules
- Always process all the documents
- Output a JSON array of integers containing the indexes of the documentsDefault Implementation
The available implementations of the IRerankingService interface are:
DefaultRerankingService Utilizes one of the provided endpoints along with a large language model (LLM) to re-rank a list of documents based on their relevance or content.
PineconeRerankingService Integrates with Pinecone's reranking endpoint and leverages Pinecone's models to perform document reranking.
LocalAIRerankingService Uses LocalAI's reranking endpoint and associated models to reorder documents locally, without requiring external cloud services.
Each implementation provides a different method for reranking documents, allowing developers to choose the most suitable service based on their requirements and the available infrastructure.
You can use the source code provided below as a reference to help you develop a custom reranking service. By examining the implementation details, you can better understand how to create your own class that extends the IRerankingService interface and integrates with your preferred reranking model or endpoint.
Last updated