Probability measures how likely an event or outcome is. In retrieval-augmented generation, it turns raw scores into normalized weights and represents how likely a model is to produce or support an answer. section 4.2.1 section 14.3.1
Two practical uses illustrate the distinction between scores, probabilities, and probability-weighted results:
- Normalize retrieval scores so p(z|x) sums to 1; this preserves ranking while making the values usable as document weights. section 4.2.1
- Combine retrieval relevance with generation likelihood; a lower-ranked document can dominate when it supports the answer much better. section 4.2.1
- Represent rating uncertainty with token probabilities; weighting possible scores produces a continuous result instead of a potentially misleading integer. section 14.3.1
A lower-ranked document becomes decisive
Five documents receive normalized retrieval weights. Document C ranks third at 0.199, but its generation score is 0.92 because it directly explains the answer's root causes. Multiplication gives 0.183, exceeding Document A's 0.034 and Document B's 0.043; summing the products yields p(y|x) = 0.500 for the answer string. section 4.2.1
Use the probability-weighted score equation to see how likely rating tokens combine into a fine-grained evaluation. section 14.3.1





















