Semantic relevance filtering selects retrieved passages by their similarity to a query before they are passed to a language model. A minimum similarity threshold can keep passages that address the question directly and exclude those that are only topically related. section 3.6.5
Filtering keeps retrieved context focused. Passing many long or tangential documents to a model can bury key facts and make them harder to use, increasing the risk of hallucination. Metadata categories such as billing, troubleshooting, and features can further narrow retrieval. section 3.6.5
International roaming support
For a query about enabling international roaming, a knowledge base might return an FAQ, a rates page, a travel-coverage blog, and terms and conditions. Similarity filtering can retain the FAQ with setup instructions and exclude related documents that do not directly answer the question. section 3.6.5
