Overview#

A production RAG platform in three moves: ingest documents through a parsing and chunking pipeline, retrieve with hybrid dense + sparse search, and generate grounded answers that cite their sources — streamed to your users over SSE.

How the pieces fit#

| Stage | What happens | Swappable via | | --- | --- | --- | | Ingestion | Parse, OCR, chunk, embed, index | Parser & chunking strategy | | Retrieval | Dense + BM25 in parallel, RRF fusion, rerank | Retrieval mode, vector backend | | Generation | Grounded prompt, token streaming, citations | Generation provider |

Every stage sits behind a registry: implementations register themselves and are selected with an environment variable. Adding a new vector store, embedding model, or chunker never touches calling code.

Design principles#

• Grounded by default — the model is instructed to answer only from retrieved context, and every claim carries a [n] citation
• Local-first development — embeddings and generation run locally; the same pipeline scales to managed infrastructure unchanged
• Observable everywhere — per-stage latency, token usage, and eval scores ship with the platform, not as an add-on

New here? The Quickstart gets you from zero to a cited, streamed answer in about five minutes.