Skip to main content
Document RAG transforms unstructured text into structured, queryable knowledge. Unlike traditional vector-only RAG, Worlds maintains the relationships between document fragments.

​
The pipeline

  1. Chunking: Break documents into semantic segments.
  2. Embedding: Generate vector representations for each segment.
  3. Extraction: Use LLMs to identify items and relationships within the text.
  4. Linking: Store segments as chunks and relationships as triples within the world memory.

​
Benefits of graph-backed documents

  • Contextual continuity: Follow relationships between different parts of a document or across multiple documents.
  • Improved recall: Retrieve specific document segments based on their logical role.
  • Verification: Cross-reference extracted facts against the original document chunks to ensure accuracy.