Slingshot
AI model · Reranker / retrieval · Text → Tables / structured data by Vectara
High-performance multilingual reranking for enterprise AI agents.
Slingshot is a multilingual reranking model that reorganizes search results to prioritize the most relevant information for AI agents. It provides fine-grained sorting to improve RAG accuracy and reduce redundancy across diverse domains.
Model specs
The full spec card (context, pricing per token, modalities and more) on AntSpace. Sign up free to see the full spec card.
Category: Natural Language Processing
- Modalities
- Text → Text
- Model Type
- Multilingual Reranker
- Primary Function
- Fine-grained result sorting
- Wiki Benchmark Score
- 0.85 (nDCG)
- News Benchmark Score
- 0.57 (nDCG)
- Shopping Benchmark Score
- 0.45 (nDCG)
- 0.85 score on Wiki and Dev domain benchmarks
- Multilingual reranking for global enterprise data
- Fine-grained sorting for higher RAG accuracy
Overview
Slingshot is Vectara's high-performance multilingual reranking model designed to optimize the relevance of search results within a Retrieval-Augmented Generation (RAG) pipeline. It serves as a critical secondary sorting layer that reorganizes the initial results from a retrieval model to ensure the most pertinent information is prioritized for the generative LLM. By performing fine-grained sorting, Slingshot improves the overall accuracy and comprehensiveness of AI agent responses while reducing redundant information. Slingshot is integrated directly into the Vectara platform and can be accessed via REST APIs to build production-grade AI assistants.
Ranking and Relevance
- Result Prioritization: The model reorganizes search results to prioritize highly relevant and useful information for the generative step.
- Fine-grained Sorting: Slingshot performs fine-grained sorting of results after the initial coarse-grained retrieval by the Boomerang model.
- Diversity Optimization: Reranking can increase the diversity of search results used in the generated response to ensure it does not have redundant information.
Multilingual and Domain Performance
- Wiki Domain Performance: The model achieved a score of 0.85 on Wiki domain benchmarks as of May 2024.
- News Domain Performance: Slingshot scored 0.57 on News domain benchmarks, exceeding Cohere Rerank V3's 0.55.
- Shopping Domain Performance: The model reached a score of 0.45 on Shopping domain benchmarks.
Benchmarks
- Wiki: 0.85; Social/Forum: 0.79; News: 0.57; Shopping: 0.45; Dev: 0.85
Use cases
- Engineering teams use Slingshot to prioritize technical documentation in troubleshooting agents for higher accuracy.
- E-commerce platforms apply Slingshot to reorganize product search results based on fine-grained relevance to user queries.
- News organizations use the model to rank multilingual articles by relevance for automated summarization tools.
- Legal teams utilize Slingshot to ensure the most pertinent clauses are prioritized during document analysis workflows.
Specifications
- Modalities
- Text → Text
- Access
- Available via Vectara REST API and Console as part of the RAG pipeline
- Languages
- Multilingual support including English and major global languages
- Model Type
- Multilingual Reranker
- Primary Function
- Fine-grained result sorting
- Wiki Benchmark Score
- 0.85 (nDCG)
- News Benchmark Score
- 0.57 (nDCG)
- Shopping Benchmark Score
- 0.45 (nDCG)
- Social/Forum Benchmark Score
- 0.79 (nDCG)
- Dev Benchmark Score
- 0.85 (nDCG)
How it works
- Query Input: The system receives a question from a person or an agent.
- Initial Retrieval: The Boomerang model performs stateless hybrid search and coarse-grained sorting of results.
- Slingshot Reranking: Slingshot reorganizes the search results to prioritize the most relevant and useful information through fine-grained sorting.
- Grounded Generation: The reranked results are passed to a generative LLM like Mockingbird to form an accurate answer.
FAQ
What is Slingshot?
Slingshot is Vectara’s reranking model that reorganizes search results to prioritize highly relevant and useful information.
How does Slingshot work with Boomerang?
Slingshot performs fine-grained sorting of results after the initial coarse-grained sorting performed by the Boomerang retrieval model.
Added by AntSpace from public sources (vectara.com).