An IT-service assistant answers from a single vector index of knowledge articles and resolved tickets. Conceptual questions ("why does VPN drop after sleep?") work well, but queries containing exact error codes such as 0x80070005 often return articles about unrelated errors with similar wording. What is the best change?

Answer: B — Add lexical search alongside vector search and fuse the two result sets so exact codes match reliably

Dense embeddings blur near-identical strings such as error codes; lexical matching captures them exactly. Hybrid retrieval keeps the semantic strength for conceptual questions and adds exact matching. (a) doesn't change how codes are matched. (c) targets the generator when the failure is in retrieval. (d) adds noise and cost without making exact matching reliable.