I kept running into the same problem with RAG pipelines: the system gives confident answers with no way to distinguish "this is in the data" from "this was filled in from model weights." So I built Kremis.

The core idea: data goes in as EAV signals (entity, attribute, value). Kremis builds a weighted graph from co-occurrence. Every query result is then classified:

  FACT      — the path exists directly in the graph
  INFERENCE — derived via traversal, not a direct link
  UNKNOWN   — not in the data at all

Same input -> same output, always. No embeddings involved in core state.

What it is today (v0.16.1, Apache 2.0):

- kremis-core: pure Rust library, no async, no network dependencies - kremis CLI + HTTP API (16 endpoints, axum, optional Bearer auth) - kremis-mcp: MCP bridge (9 tools, stdio) -- works in Claude Desktop/Cursor - Persistent backend via redb (ACID, crash-safe) - Binary releases + Docker (~136MB)

Two recent fixes that might interest Rust devs: - Storage errors were being silently collapsed into Ok(None), making I/O failures look like "node not found" -- now properly propagated as Err(...) - increment_edge was creating phantom edges when either node was missing -- fixed with the same guard already present in insert_edge

Still experimental. The main friction is EAV ingestion: you have to structure your data as signals before it can be queried. Curious whether that's a dealbreaker in practice or a reasonable tradeoff for the grounding guarantees.