Promptriever - an information retrieval model that supports natural language prompts in response to user search requests.
Promptriever is a novel retrieval model jointly developed by Johns Hopkins University and Samaya AI. It can accept natural language prompts like a language model, responding to users' search needs in an intuitive way. Promptriever...
What is Promptriever?
Promptriever, a novel retrieval model jointly developed by Johns Hopkins University and Samaya AI, can accept natural language prompts like a language model, responding to users' search needs in an intuitive way. Trained on the instruction training set of the MS MARCO dataset, Promptriever performs excellently on standard retrieval tasks, more effectively following detailed instructions and improving query robustness and retrieval performance. Promptriever demonstrates the potential of combining the prompting techniques of large language models with information retrieval.
Promptriever's main functions
- Accept natural language promptsIt can understand and respond to prompts in natural language, allowing users to express their search needs in a more natural way.
- Dynamically adjust correlationDynamically adjust the relevance of search results based on specific user instructions. For example, filter and sort documents based on specific user requirements for search results (such as time range or specific attributes).
- Improve retrieval robustnessBy understanding and processing the subtle differences in natural language, the robustness of the model to different query expressions is enhanced.
- Improve search performanceBased on prompts, hyperparameter searches are performed to improve the quality of search results.
The technical principle of Promptriever
- Dual encoder architectureIt is based on a dual-encoder architecture and uses a large language model (such as LLaMA-2 7B) as its underlying support model.
- Instruction training dataset: Select and publish new instruction-level training sets from the MS MARCO dataset, which include natural language instructions that define query relevance.
- Instruction generationUse language models to generate more specific instructions, which can add additional requirements or explicitly exclude certain types of documents.
- Instruction Negative Example MiningBased on the generation and filtering (query, passage) pairs, negative examples are created that show a decrease in relevance after a specific instruction is added, forcing the model to learn how to adjust its relevance judgment according to the instruction.
- Zero-shot hinting technologyThis approach utilizes zero-shot hints to perform hyperparameter search, similar to language model hints, thereby improving retrieval performance.
Promptriever's project address
- GitHub repository:https://github.com/orionw/promptriever
- arXiv technical paper:https://arxiv.org/pdf/2409.11136
Application scenarios of Promptriever
- Search Engine OptimizationProvides more accurate search results by understanding users' natural language queries and instructions, thus improving the search experience.
- Smart assistants and chatbotsIt can understand and execute complex user commands, providing more personalized and context-sensitive responses.
- Internal Enterprise SearchQuickly and accurately retrieve specific information from the enterprise knowledge base to improve work efficiency.
- Academic research and literature retrieval: Retrieve specific academic papers and literature based on the researcher's detailed query instructions.
- e-commerceBased on users' shopping needs and preferences, we provide customized search results and product recommendations.