SUPERAI
Hybrid graph + vector retrieval, native desktop.
- Architecture
- Hybrid graph + vector
- Form factor
- Native desktop
- Stage
- In development
Pure vector search returns documents that look similar to the query. It misses the structure connecting those documents to the things you already know. SUPERAI is an attempt to fix that by treating retrieval as a graph + vector problem rather than a vector-only one.
The architecture is a Tauri v2 desktop app: a Rust core with a React 19 shell. SQLite holds documents and full-text indexes; Qdrant holds embeddings; a graph layer encodes relations between entities mentioned in the corpus. Queries fan out across the three, results merge through reciprocal-rank fusion, and a cross-encoder reranks the top candidates. Multi-provider LLM orchestration sits behind a thin abstraction so the same query can be routed to whatever model is cheapest or fastest for that step.
The hard part is not retrieval. It is entity canonicalization: "Smith," "R. Smith," and "Smith et al." need to resolve to one node, and getting that right at scale is its own subproject.