The 2026-04-19 Hermes Agent Use Cases run had a Nate Herk YouTube video
titled "I Tested Claude's New Managed Agents" score 51 and rank #2
with zero Hermes content. The reranker had intent-specific scoring hints
but no entity-grounding check, so topic-vicinity matches (one offhand
OpenClaw mention) drifted to the top.
Add _primary_entity(topic) that strips intent-modifier suffixes ("use
cases", "workflows", etc.) so "Hermes Agent use cases" yields
primary_entity="Hermes Agent". Pass the entity through to both the LLM
and fallback scoring paths.
Fallback path: if primary_entity is not found (case-insensitive) in
title + snippet, subtract ENTITY_MISS_PENALTY (25 pts). Skip the
demotion for candidates with no text at all (image-only TikToks etc.)
to avoid false negatives on thin-text sources.
LLM path: add a "Primary entity grounding" hint to _build_prompt when
primary_entity is non-empty. Instructs the LLM to score candidates
without the entity at <=30.
Tests: 24 rerank tests pass, including 8 new entity-grounding tests.
Matt Van Horn
a709d66e2a