RAG Platform
An open-source retrieval-augmented generation platform that separates document ingestion, multimodal indexing, hybrid retrieval and agent integration into independently deployable services, with tenant isolation, role-based access control and citation-ready results.
RAG Platform ingests knowledge from multiple sources, indexes multimodal content, and serves secure, citation-ready retrieval to applications and AI agents. Ingestion, indexing, retrieval, lifecycle management and agent integration are split into independently deployable services, and the project is still under active pre-1.0 development.
Pluggable source connectors, including Yuque and local Markdown, feed a shared download-clean-chunk-vectorize pipeline. Hybrid retrieval combines vector similarity, BM25 and title relevance with reciprocal rank fusion; results return document and chunk IDs, source URLs, scores and character locations. Ingestion jobs persist request snapshots before processing, with idempotency checks and stale-job recovery. Access control supports knowledge-base and document-level role-based access, and public collections; it also exposes remote retrieval and health checks as MCP tools, plus offline evaluation with hit rate, recall, precision, MRR and NDCG.
Built with Python, organized as a uv workspace of five packages sharing embedding and Qdrant payload contracts, each runtime component with its own container image. MIT licensed.