MothRag
A deterministic multi-hop retrieval-augmented generation system that reasons over a plain, append-only dense index instead of a knowledge graph, so it stays current on data that changes daily without any re-indexing pass.
MothRag answers questions that need multi-hop reasoning across a corpus that keeps changing, such as live standings, prices, support tickets or internal documents, by treating the reasoning as query-time orchestration over a plain dense index rather than building and maintaining a knowledge graph the way GraphRAG or HippoRAG do. Updating the index is embed-and-append: a changed chunk is embedded and appended with one embedding call, with no LLM extraction pass and no graph rebuild or retraining required, so the system keeps answering correctly as new data arrives.
It runs entirely on commodity, pay-per-call APIs, using a hosted Llama-3.3-70B model as the reader and a hosted embedder, with a reported cost per query in the low cents and no local GPU or fine-tuning required. The same input always produces the same output, with zero run-to-run variance, and every answer ships an inspectable proof tree; a paper reports that it outscores graph-based multi-hop RAG systems on the HotpotQA, 2WikiMultiHop and MuSiQue benchmarks.