Linh Phuong Dao

Grounding LLMs with search evidence reduces hallucination through system design, not model size.

The infrastructure choice that determines what content your LLM pipeline can actually reason over.

Preserve agent context through stateful orchestration, not stateless replay.

Architectural patterns determine whether search-grounded agents retrieve and reason reliably.

Split retrieval and generation scoring to catch hallucinations end-to-end evaluation misses.

Query-time fetching forces a rethink of how web pages get split.

Semantic search can't tell old from new, and that gets costly.