{"id":36283,"date":"2024-01-07T19:44:45","date_gmt":"2024-01-07T19:44:45","guid":{"rendered":"https:\/\/writer.com\/?page_id=36283"},"modified":"2024-07-09T20:06:30","modified_gmt":"2024-07-09T20:06:30","slug":"product-knowledge-graph","status":"publish","type":"page","link":"https:\/\/writer.com\/product\/graph-based-rag\/","title":{"rendered":"Knowledge Graph"},"content":{"rendered":"
Knowledge Graph, our graph-based retrieval-augmented generation (RAG), achieves higher accuracy than traditional RAG approaches that use vector retrieval.<\/p>\n\n<\/div>\n\n
Knowledge Graph draws on a specialized LLM that\u2019s trained to process data at scale and build valuable semantic relationships between data points. It stores data in a cost-effective, easy-to-update graph structure.<\/p>\n\n<\/div>\n\n
The alternative approach<\/strong><\/p>\n\n\n\n By converting data into vector embeddings, traditional RAG can only define similarity by distance between data points but has no context on their semantic relationships. Vector databases are also difficult and costly to maintain and update.<\/p>\n\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n\n Because graph structures retain semantic relationships, Knowledge Graph accurately retrieves relevant data for each query. Our retrieval-aware compression technique condenses data and indexes it with metadata, which gives it rich context.<\/p>\n\n<\/div>\n\n The alternative approach<\/strong><\/p>\n\n\n\n Traditional RAG converts the query into a vector embedding and uses a rough algorithm to find the closest data points to the query, without any understanding of the relationship between the data points. When data is dense, this method fails to return the most relevant data consistently.<\/p>\n\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n\n To generate a response, Knowledge Graph sends relevant data to our Palmyra LLMs, which are top-ranked<\/a> and trained with 1 trillion tokens of quality data. We apply advanced techniques to enhance performance and minimize hallucinations.<\/p>\n\n<\/div>\n\n<\/figure>\n\n<\/div>\n\n
Accurate retrieval methodology<\/h3>\n\n\n\n
<\/figure>\n\n<\/div>\n\n
State-of-the-art LLMs<\/h3>\n\n\n\n