fix: expand entity-grounding haystack to transcripts + top comments

PR #285's entity grounding checked only title + snippet. That missed:

- YouTube videos where the entity is mentioned in transcript but not
  in title (false demotion of on-topic content)
- Reddit posts where the entity is in top comments but not in title
  (false demotion of on-topic discussion)

And it also wasn't strong enough to reliably demote items like the
2026-04-19 Nate Herk "Managed Agents" video - which had no Hermes
anywhere - because the -25 penalty on rerank_score composed to only
-15 on final_score via the 0.60 weight, and engagement bonus partially
offset that.

Two fixes:

1. _candidate_haystack() now joins title + snippet +
   metadata[transcript_snippet] + metadata[transcript_highlights] +
   metadata[top_comments][*].excerpt/text + metadata[comment_insights].
   Catches entity mentions wherever they actually live. Guarded with
   isinstance checks so malformed metadata doesn't raise.
2. ENTITY_MISS_FINAL_PENALTY (20.0) applied directly in _final_score
   when candidate.explanation contains "entity-miss". This lands the
   full penalty weight on the composite signal that cluster-scoring
   consumes, instead of being diluted by the rerank_score weight.
   Combined effect: entity-miss gap grows from ~15 to ~35 points.

Tests: 8 new scenarios covering transcript match, transcript highlight
match, top-comment match, comment-insight match, empty-text skip,
no-primary-entity no-op, and the dual-penalty composition check.
This commit is contained in:
Matt Van Horn
2026-04-19 10:28:36 -07:00
parent b7df5ecd2d
commit a7d6ef051a
2 changed files with 187 additions and 6 deletions
+59 -6
View File
@@ -214,6 +214,39 @@ def _apply_fallback_scores(candidates: list[schema.Candidate], *, primary_entity
candidate.final_score = _final_score(candidate)
def _candidate_haystack(candidate: schema.Candidate) -> str:
"""Build the lowercase text blob against which entity-grounding is checked.
Expanded 2026-04-19 to include transcript snippets, transcript highlights,
and top-comment text. The prior `title + snippet` check missed YouTube
videos whose entity mentions live in transcript content and Reddit posts
whose mentions are in top comments. Now checks all text surfaces a human
would see.
"""
parts: list[str] = [candidate.title or "", candidate.snippet or ""]
metadata = candidate.metadata or {}
transcript_snippet = metadata.get("transcript_snippet") or ""
if isinstance(transcript_snippet, str):
parts.append(transcript_snippet)
for hl in metadata.get("transcript_highlights") or []:
if isinstance(hl, str):
parts.append(hl)
for tc in metadata.get("top_comments") or []:
if isinstance(tc, dict):
parts.append(str(tc.get("excerpt", "") or tc.get("text", "") or ""))
elif isinstance(tc, str):
parts.append(tc)
for insight in metadata.get("comment_insights") or []:
if isinstance(insight, str):
parts.append(insight)
return " ".join(parts).lower()
def _fallback_tuple(candidate: schema.Candidate, *, primary_entity: str = "") -> tuple[float, str]:
score = (
(candidate.local_relevance * 100.0 * 0.7)
@@ -222,12 +255,16 @@ def _fallback_tuple(candidate: schema.Candidate, *, primary_entity: str = "") ->
)
reason = "fallback-local-score"
# Entity-grounding demotion: if the primary entity (topic minus intent
# modifier) is not present in the candidate's title or snippet, subtract
# ENTITY_MISS_PENALTY. Skip for candidates with no text at all (e.g.,
# image-only TikToks) to avoid penalizing thin-text sources unfairly.
if primary_entity and (candidate.title or candidate.snippet):
haystack = f"{candidate.title} {candidate.snippet}".lower()
if primary_entity.lower() not in haystack:
# modifier) is not present anywhere in the candidate's text surfaces
# (title, snippet, transcript, transcript highlights, top comments,
# insights), subtract ENTITY_MISS_PENALTY. Skip for candidates with
# NO text anywhere (e.g., image-only TikToks) to avoid penalizing
# thin-text sources unfairly. 2026-04-19 Nate Herk "Managed Agents"
# video ranked #2 on a Hermes query despite zero Hermes mentions
# because the old haystack only checked title + snippet.
if primary_entity:
haystack = _candidate_haystack(candidate)
if haystack.strip() and primary_entity.lower() not in haystack:
score -= ENTITY_MISS_PENALTY
reason = "fallback-local-score (entity-miss demotion)"
return max(0.0, min(100.0, score)), reason
@@ -247,6 +284,17 @@ def _primary_entity(topic: str) -> str:
return stripped
#: Secondary entity-miss penalty applied directly to final_score (not just
#: rerank_score). The -25 on rerank_score composes to only -15 on final_score
#: via the 0.60 weight, which engagement bonus partially offsets on
#: high-view YouTube items. This secondary penalty lands the full weight on
#: the composite signal the cluster-scoring layer consumes. 2026-04-19
#: Nate Herk "Managed Agents" video ranked at cluster #2 with score 51
#: despite the rerank_score demotion because engagement + freshness drowned
#: the dilute penalty. This backstop makes the demotion actually decisive.
ENTITY_MISS_FINAL_PENALTY = 20.0
def _final_score(candidate: schema.Candidate) -> float:
normalized_rrf = _normalized_rrf(candidate.rrf_score)
rerank_score = candidate.rerank_score or 0.0
@@ -265,6 +313,11 @@ def _final_score(candidate: schema.Candidate) -> float:
)
if candidate.rerank_score is not None and candidate.rerank_score < 20.0:
base *= 0.3
# Secondary entity-grounding penalty: when the fallback path flagged
# entity-miss via candidate.explanation, apply an additional penalty
# at final_score level so engagement signal can't mask the demotion.
if candidate.explanation and "entity-miss" in candidate.explanation:
base = max(0.0, base - ENTITY_MISS_FINAL_PENALTY)
return base