client wants rag with "confidence scores" on every citation - how do you even calculate this

shipping a rag system for legal document search. client wants a confidence score (0-100) on every citation the model returns. they want to show it in the UI next to each source. the obvious approach is cosine similarity between query embedding and retrieved doc embedding, but that doesn't actually measure whether the citation supports the claim being made. you can have high similarity but the model still hallucinates the specific detail. tested using the model's logprobs on the citation token sequence but that just measures "how confident is the model in this text" not "how well does this source support this claim". anyone actually solved this in production or is everyone just faking it with retrieval similarity scores