3 ms·
Good to see others are thinking similar - I've implemented a lot of what this is discussing into DoctorGPT: https://github.com/FeatureBaseDB/DoctorGPT https://g
by kordlessagain 3y ago
Good to see others are thinking similar - I've implemented a lot of what this is discussing into DoctorGPT: https://github.com/FeatureBaseDB/DoctorGPT https://github.com/FeatureBaseDB/DoctorGPT. FeatureBase is great for storing sets of things and vectors, and combining graphs/keyterms to filter the vector space.
We're close to releasing a hosted tool that will take plain text and convert it to semantic graphs and embeddings, which will allow users to use this basic approach for searching their document fragments, across many documents. Should be out in a few weeks.
- kordlessagain 3y agoIndexing the paper with some of the strategy of the paper implemented in the indexing. system> User entered, `what are knowledge graphs and how does the paper describe using them?`. bot> Knowledge graphs are data structures that represent entities and their relationships in a structured manner. In the paper, the authors propose a method called Knowledge Graph Prompting (KGP) for Multi-Document Question Answering (MD-QA). They create a knowledge graph over multiple documents, where nodes represent passages or document structures (such as pages or tables), and edges represent the semantic or lexical similarity between passages or the intra-document structural relations. The constructed graph acts as a global ruler to regulate the transitional space among passages and reduce retrieval latency. It also serves as a prompt for Language Models (LLMs), assisting them in answering questions by providing contextual information. The authors demonstrate through experiments that leveraging knowledge graphs enhances the prompt design for LLMs in MD-QA.