knowhere

Why we included this project

Unstructured documents tend to defeat naive parsing: a table gets flattened, a heading loses its place in the hierarchy, and an agent ends up retrieving text out of context. Knowhere addresses this by turning messy PDFs, Office files, and scanned images into a navigable tree of chunks, keeping headings, tables, and section paths intact. The API and worker pipeline bundle ingestion, multi-modal parsing, graph construction, and retrieval, and you can point it at OpenAI, DeepSeek, or Qwen models instead of being tied to one vendor. The retrieval layer lets agents walk section trees and cross-document links to drill into the relevant region and return citations tied to the original source. That is what makes it attractive for anyone building agentic or vector-based RAG who needs traceable, context-rich answers from long or complex documents.