How do you handle LLM hallucinations in your pipeline?

RAG retrieves the right documents, but LLMs still make up details not present in the source data.

I kept hitting this wall, so I built Kremis. It shifts the approach from probabilistic guessing to deterministic verification.

Here is what it actually does:

1. Ingests your data to build a strict knowledge graph.

2. Evaluates every LLM query result against that graph.

3. Outputs a concrete classification for every claim: FACT, INFERENCE, or UNKNOWN.

No fuzzy confidence scores. If a claim isn't explicitly supported by the graph, the system flags it as UNKNOWN immediately.

I'm curious how other makers are handling hallucination in production. Do you:

- Post-process LLM outputs with a secondary verification step?

- Rely entirely on citations/attributions from the model?

- Accept a certain rate of hallucinations as the cost of doing business?

Would love to hear what's working (or failing) for you.

(For context, the open-source project is here: github.com/TyKolt/kremis)