What is Promptriever
Promptriever is a new retrieval model jointly launched by Johns Hopkins University and Samaya AI. It can accept natural language prompts like a language model and respond to users' search needs in an intuitive way. Promptriever is trained on an instruction training set based on the MS MARCO dataset, performs well on standard retrieval tasks, can more effectively follow detailed instructions, and improves robustness to queries and retrieval performance. Promptriever demonstrates the potential of combining prompt techniques from large language models with information retrieval.
Promptriever's main features
- Accepts natural language prompts: It can understand and respond to prompts in natural language form, allowing users to express search needs in a more natural way.
- Dynamically adjusts relevance: It dynamically adjusts the relevance of search results according to the user's specific instructions, for example, filtering and ranking documents according to the user's specific requirements for search results (such as time range and specific attributes).
- Improves retrieval robustness: Based on understanding and handling subtle differences in natural language, it enhances the model's robustness to different query expressions.
- Improves retrieval performance: It performs hyperparameter search based on prompts, improving the quality of retrieval results.
Promptriever's technical principles
- Dual-encoder architecture: Based on a dual-encoder (bi-encoder) architecture, with a large language model (such as LLaMA-2 7B) as the supporting model behind it.
- Instruction training dataset: A new instruction-level training set is selected and released from the MS MARCO dataset, and the training includes natural language instructions that define query relevance.
- Instruction generation: A language model is used to generate more specific instructions, and the instructions can add additional requirements or explicitly exclude certain types of documents.
- Instruction negative example mining: Based on generated and filtered (query, passage) pairs, negative examples are created whose relevance decreases after adding specific instructions, forcing the model to learn how to adjust relevance judgments according to instructions.
- Zero-shot prompting technique: Hyperparameter search is performed based on zero-shot prompting techniques, similar to prompting of language models, to improve retrieval performance.
Promptriever's project address
- GitHub repository：https://github.com/orionw/promptriever
- arXiv technical paper：https://arxiv.org/pdf/2409.11136
Promptriever's application scenarios
- Search engine optimization: It provides more accurate search results and improves the search experience based on understanding users' natural language queries and instructions.
- Intelligent assistants and chatbots: It understands and executes users' complex instructions and provides more personalized and context-relevant answers.
- Enterprise internal search: It quickly and accurately retrieves specific information in an enterprise knowledge base, improving work efficiency.
- Academic research and literature retrieval: It retrieves specific academic papers and literature materials according to researchers' detailed query instructions.
- E-commerce: It provides customized search results and product recommendations according to users' shopping needs and preferences.