RAGFlow

InfiniFlow's engine puts most of its effort into the parsing step that many RAG stacks treat as an afterthought. Its DeepDoc component reads page layouts, runs OCR and reconstructs tables in PDFs, Word files, slides, spreadsheets, scans and web pages, then splits content with selectable chunking templates; you can inspect chunks visually and correct them by hand after parsing, and every answer cites the chunks it used. On top come agentic retrieval with adjustable reasoning depth, prebuilt agent templates, knowledge compilation into wikis, graphs, mind maps or timelines, and connectors reaching sources like BigQuery or sitemaps. The 1.0 release candidate moved the services to Go. Deployment is a Docker Compose stack with Elasticsearch or Infinity as the document store; plan on four cores, 16 GB of memory and 50 GB of disk to start. In the web interface you register your own chat, embedding and reranking providers, hosted or local, as no LLM is bundled. An optional code sandbox needs gVisor when self-managed. Code is Apache-2.0, and a hosted cloud edition exists. It fits document-heavy assistants that must show their sources.