2026
PolyRAG
Local federated RAG
Picks the right store (graph, vector or relational) with arithmetic, not an agent, and shows the route and every pipeline step. Runs locally on AMD GPUs via Vulkan.
Drop a spreadsheet, a scanned receipt and a policy document into one folder. PolyRAG reads each file, splits it into chunks and decides, chunk by chunk, whether it belongs in a SQL table, a vector index or a knowledge graph. When you ask a question, the same decision picks where to look.
The premise is the opposite of agent-based RAG: use deterministic maths wherever it can decide, and keep the model only for writing the answer. Every step reports itself live, so you can watch the route being chosen and the numbers behind it.
Routing by arithmetic
Cosine similarity with a margin against configured anchors, a tabularity heuristic before it, and each store's section headings as evidence only when two routes tie.
Three stores, one engine
SQLite with Text-to-SQL, Qdrant with HNSW and a networkx graph queried with Personalized PageRank.
Fully observable
OpenTelemetry spans stream over WebSocket to the interface: each route's score, the margin that decided it and every step's latency.
100% local
Three models served by llama.cpp on an AMD GPU through Vulkan, with no CUDA, Docker or cloud. Works across languages thanks to BGE-M3 embeddings.