AssistantPrefulfillmentRankerPrefulfillmentSignals

Whether this is a fulfillable, dominant Media intent.

Whether or not the intent was generated by a legacy Aqua domain that PFR should care about.

Whether the interpretation has a Search answer group object, signifying it came from Search resolution.

Whether the intent is from ORBIT.

Whether the interpretation should run through grounding box or not.

The total effective length of the spans for the arguments used to construct the parse. May include vertical specific adjustments. Eg: For the query [delete my 7 p.m. alarm called chicken] and intent Delete_alarm(alarm_object=RD(category=AlarmObject( label="chicken", trigger_time_datetime=<< 7 PM >>))), the effective argument span is "7 p.m." + "chicken" (total length of 13).

Query understanding phase the intent was produced in.

Whether the intent is fully grounded.

Whether the intent is produced by the Sage IntentGenerator invoked by the NSP intent-generator (thus, Sage-in-NSP-invoking_Sage, or Sage-in-Nage).

Whether the intent is an intent marked disabled by the Sage IG.

Whether the intent is dominant according to NSP deep-media.

Whether the intent is a media object video intent.

A ID corresponding to which bucket a given parsing score belongs in.

Intent confidence predicted by the AssistantVerticalClassifier QRewrite servlet.

Parse score generated by NSP. If NSP intent has been pruned due to dededuplication, It will have highest parse score for deduplicated intent.

Signals as proposed in go/improved-grounding-signals. Sum of the number of constraints used by the Grounding Box to ground each variable.

Whether the intent is a podcast intent.

Whether this intent is a score-based intent, relying on PFR for scoring and pruning to the top intent.

Whether the intent comes from the Sage IntentGenerator's "platinum" source, signifying high-confidence in quality.

Rank of the intent as reported by NSP.

Indicates interpretation dominance predicted by KScorer

Whether the intent is a PlayMedia radio intent.

The determination made by the SearchDispatchingConfig as to whether and how this interpretation should be dispatched to Search.

Used for PFR manual rule to prefer high confidence podcast intent (e.g. topical, genre) over generic podcast intents.

A parsing score that is independently calibrated by each parser/IG.

Assistant User Interaction Score which is aggregated using intent name from exp laelaps.

This is a cross-intent feature which is calculated by iterating all intent candidates. This feature should be populated in post-IG stage (before GB).

Cosine similarity between predicted query-to-term model and assistant intent-type-based salient terms. This is intended to be only used for ACE ranking and only populated for assistant traffic.

Whether the intent is a media object tvm intent.

Number of grounded arguments the intent has, populated by the Grounding Box.

Grounding Signals. Number of groundable arguments the intent has, populated by the Grounding Box.

intent_type differentiates between intents that share the top level intent name. For eg: for TV_FALLBACK_SEARCH_INTENT, the top level intent name must be "Find_media" and the media_object argument within it must be of type "Media_unspecified".

Cosine similarity between predicted query-to-term model and intent-type-based salient terms. This is intended to be used as a backoff to pq2t_vs_qibst_cosine if it is missing.

Whether this interpretation was genearted by NSP.

Whether this intent was generated by Sage.

QUS intent-based ranking signals. Assistant User Interaction Score which is aggregated using intent name.

Used for PFR manaul rule to prefer high kscore radio intent. The k-score comes from twiddled_ranking_scores in query interpretations.

Whether the intent aqua intent from media domain enabled in QBT

Whether the intent is a Smarthome-domain intent with valid device arguments.

SmartHome intent metadata used for the SmartHome business-rules twiddler.

Sum of the number of constraints satisfied for each variable. Depending on the match score for a constraint, this number can be fractional and is in the range [0, num_constraints]. Populated by the Grounding Box.

Whether the intent is a media control intent.

Used for PFR manual rule to prefer high confidence podcast intent (e.g. topical, genre) over generic podcast intents.

The maximum score assigned by the Horizontal Grounding Ranker (HGR) across all of the intent's binding sets.

intent_name is used by PFR ensemble model. See go/pfr_ha_launch_doc

The rank order of the interpretation as determined by kscorer. The kscorer-determined dominant interpretation, if any, gets a rank of 0. The remaining N interpretations get a rank of 1 through N.

Whether the intent is a PlayGenericMusic-type intent.