Hi all — I'm a database reliability engineer (Postgres/MySQL/TimescaleDB background) who got frustrated that every retriever I've used treats retrieved chunks as equally trustworthy, even when two of them flatly contradict each other.

I built TekmerDB — open source, Apache 2.0, single Rust binary — as a storage layer that runs an NLI contradiction check on every insert. If a new fact conflicts with something already stored, both get flagged, confidence drops on the outlier, and the conflict is tracked with full source provenance. It ships with an MCP server too, so an agent can query it directly.

Given LlamaIndex is the data/retrieval layer for a lot of production RAG stacks, I'm curious whether this is a problem people building indices and custom retrievers here have actually hit — conflicting source documents, stale vs. current data, that kind of thing. Not pitching a replacement for anything in the stack, more wondering if it's worth exploring as something that plugs in alongside an existing index/retriever.

Happy to share a quick demo (a real vs. fabricated energy-stats example) if it's useful context. Genuinely just looking for whether this resonates with people using LlamaIndex day to day.