Skip to content

[8.x] Rework RRF to be evaluated during rewrite phase (#112648) #114000

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 2 commits into from
Oct 3, 2024
Merged

[8.x] Rework RRF to be evaluated during rewrite phase (#112648) #114000

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
29f0dfc
Rework RRF to be evaluated during rewrite phase (#112648)
pmpailis Oct 3, 2024
c6996bc
Merge branch '8.x' into backport/8.x/pr-112648
elasticmachine Oct 3, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
7 changes: 1 addition & 6 deletions docs/reference/search/rrf.asciidoc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -300,13 +300,12 @@ We have both the ranker's `score` and the `_rank` option to show our top-ranked
"value" : 5,
"relation" : "eq"
},
"max_score" : null,
"max_score" : ...,
"hits" : [
{
"_index" : "example-index",
"_id" : "3",
"_score" : 0.8333334,
"_rank" : 1,
"_source" : {
"integer" : 1,
"vector" : [
Expand All @@ -319,7 +318,6 @@ We have both the ranker's `score` and the `_rank` option to show our top-ranked
"_index" : "example-index",
"_id" : "2",
"_score" : 0.5833334,
"_rank" : 2,
"_source" : {
"integer" : 2,
"vector" : [
Expand All @@ -332,7 +330,6 @@ We have both the ranker's `score` and the `_rank` option to show our top-ranked
"_index" : "example-index",
"_id" : "4",
"_score" : 0.5,
"_rank" : 3,
"_source" : {
"integer" : 2,
"text" : "rrf rrf rrf rrf"
Expand Down Expand Up @@ -499,7 +496,6 @@ Working with the example above, and by adding `explain=true` to the search reque
"_index": "example-index",
"_id": "3",
"_score": 0.8333334,
"_rank": 1,
"_explanation":
{
"value": 0.8333334, <1>
Expand Down Expand Up @@ -608,7 +604,6 @@ The response would now include the named query in the explanation:
"_index": "example-index",
"_id": "3",
"_score": 0.8333334,
"_rank": 1,
"_explanation":
{
"value": 0.8333334,
Expand Down
90 changes: 45 additions & 45 deletions ...nternalClusterTest/java/org/elasticsearch/search/retriever/RankDocRetrieverBuilderIT.java
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -36,6 +36,7 @@
import org.elasticsearch.search.sort.FieldSortBuilder;
import org.elasticsearch.search.sort.NestedSortBuilder;
import org.elasticsearch.search.sort.ScoreSortBuilder;
import org.elasticsearch.search.sort.ShardDocSortField;
import org.elasticsearch.search.sort.SortBuilder;
import org.elasticsearch.search.sort.SortOrder;
import org.elasticsearch.test.ESIntegTestCase;
Expand Down Expand Up @@ -189,8 +190,10 @@ public void testRankDocsRetrieverBasicWithPagination() {
SearchSourceBuilder source = new SearchSourceBuilder();
StandardRetrieverBuilder standard0 = new StandardRetrieverBuilder();
// this one retrieves docs 1, 4, and 6
standard0.queryBuilder = QueryBuilders.constantScoreQuery(QueryBuilders.queryStringQuery("quick").defaultField(TEXT_FIELD))
.boost(10L);
standard0.queryBuilder = QueryBuilders.boolQuery()
.should(QueryBuilders.constantScoreQuery(QueryBuilders.idsQuery().addIds("doc_1")).boost(10L))
.should(QueryBuilders.constantScoreQuery(QueryBuilders.idsQuery().addIds("doc_4")).boost(9L))
.should(QueryBuilders.constantScoreQuery(QueryBuilders.idsQuery().addIds("doc_6")).boost(8L));
StandardRetrieverBuilder standard1 = new StandardRetrieverBuilder();
// this one retrieves docs 2 and 6 due to prefilter
standard1.queryBuilder = QueryBuilders.constantScoreQuery(QueryBuilders.termsQuery(ID_FIELD, "doc_2", "doc_3", "doc_6")).boost(20L);
Expand All @@ -205,8 +208,8 @@ public void testRankDocsRetrieverBasicWithPagination() {
null
);
// the compound retriever here produces a score for a doc based on the percentage of the queries that it was matched on and
// resolves ties based on actual score, rank, and then the doc (we're forcing 1 shard for consistent results)
// so ideal rank would be: 6, 2, 1, 4, 7, 3 and with pagination, we'd just omit the first result
// resolves ties based on actual score, and then the doc (we're forcing 1 shard for consistent results)
// so ideal rank would be: 6, 2, 1, 3, 4, 7 and with pagination, we'd just omit the first result
source.retriever(
new CompoundRetrieverWithRankDocs(
rankWindowSize,
Expand All @@ -227,9 +230,9 @@ public void testRankDocsRetrieverBasicWithPagination() {
assertThat(resp.getHits().getTotalHits().relation, equalTo(TotalHits.Relation.EQUAL_TO));
assertThat(resp.getHits().getAt(0).getId(), equalTo("doc_2"));
assertThat(resp.getHits().getAt(1).getId(), equalTo("doc_1"));
assertThat(resp.getHits().getAt(2).getId(), equalTo("doc_4"));
assertThat(resp.getHits().getAt(3).getId(), equalTo("doc_7"));
assertThat(resp.getHits().getAt(4).getId(), equalTo("doc_3"));
assertThat(resp.getHits().getAt(2).getId(), equalTo("doc_3"));
assertThat(resp.getHits().getAt(3).getId(), equalTo("doc_4"));
assertThat(resp.getHits().getAt(4).getId(), equalTo("doc_7"));
});
}

Expand All @@ -242,8 +245,10 @@ public void testRankDocsRetrieverWithAggs() {
SearchSourceBuilder source = new SearchSourceBuilder();
StandardRetrieverBuilder standard0 = new StandardRetrieverBuilder();
// this one retrieves docs 1, 4, and 6
standard0.queryBuilder = QueryBuilders.constantScoreQuery(QueryBuilders.queryStringQuery("quick").defaultField(TEXT_FIELD))
.boost(10L);
standard0.queryBuilder = QueryBuilders.boolQuery()
.should(QueryBuilders.constantScoreQuery(QueryBuilders.idsQuery().addIds("doc_1")).boost(10L))
.should(QueryBuilders.constantScoreQuery(QueryBuilders.idsQuery().addIds("doc_4")).boost(9L))
.should(QueryBuilders.constantScoreQuery(QueryBuilders.idsQuery().addIds("doc_6")).boost(8L));
StandardRetrieverBuilder standard1 = new StandardRetrieverBuilder();
// this one retrieves docs 2 and 6 due to prefilter
standard1.queryBuilder = QueryBuilders.constantScoreQuery(QueryBuilders.termsQuery(ID_FIELD, "doc_2", "doc_3", "doc_6")).boost(20L);
Expand All @@ -267,13 +272,15 @@ public void testRankDocsRetrieverWithAggs() {
)
)
);
source.size(1);
source.aggregation(new TermsAggregationBuilder("topic").field(TOPIC_FIELD));
SearchRequestBuilder req = client().prepareSearch(INDEX).setSource(source);
ElasticsearchAssertions.assertResponse(req, resp -> {
assertNull(resp.pointInTimeId());
assertNotNull(resp.getHits().getTotalHits());
assertThat(resp.getHits().getTotalHits().value, equalTo(1L));
assertThat(resp.getHits().getTotalHits().value, equalTo(5L));
assertThat(resp.getHits().getTotalHits().relation, equalTo(TotalHits.Relation.EQUAL_TO));
assertThat(resp.getHits().getHits().length, equalTo(1));
assertThat(resp.getHits().getAt(0).getId(), equalTo("doc_2"));
assertNotNull(resp.getAggregations());
assertNotNull(resp.getAggregations().get("topic"));
Expand All @@ -291,8 +298,10 @@ public void testRankDocsRetrieverWithCollapse() {
SearchSourceBuilder source = new SearchSourceBuilder();
StandardRetrieverBuilder standard0 = new StandardRetrieverBuilder();
// this one retrieves docs 1, 4, and 6
standard0.queryBuilder = QueryBuilders.constantScoreQuery(QueryBuilders.queryStringQuery("quick").defaultField(TEXT_FIELD))
.boost(10L);
standard0.queryBuilder = QueryBuilders.boolQuery()
.should(QueryBuilders.constantScoreQuery(QueryBuilders.idsQuery().addIds("doc_1")).boost(10L))
.should(QueryBuilders.constantScoreQuery(QueryBuilders.idsQuery().addIds("doc_4")).boost(9L))
.should(QueryBuilders.constantScoreQuery(QueryBuilders.idsQuery().addIds("doc_6")).boost(8L));
StandardRetrieverBuilder standard1 = new StandardRetrieverBuilder();
// this one retrieves docs 2 and 6 due to prefilter
standard1.queryBuilder = QueryBuilders.constantScoreQuery(QueryBuilders.termsQuery(ID_FIELD, "doc_2", "doc_3", "doc_6")).boost(20L);
Expand All @@ -307,8 +316,8 @@ public void testRankDocsRetrieverWithCollapse() {
null
);
// the compound retriever here produces a score for a doc based on the percentage of the queries that it was matched on and
// resolves ties based on actual score, rank, and then the doc (we're forcing 1 shard for consistent results)
// so ideal rank would be: 6, 2, 1, 4, 7, 3
// resolves ties based on actual score, and then the doc (we're forcing 1 shard for consistent results)
// so ideal rank would be: 6, 2, 1, 3, 4, 7
// with collapsing on topic field we would have 6, 2, 1, 7
source.retriever(
new CompoundRetrieverWithRankDocs(
Expand Down Expand Up @@ -338,7 +347,6 @@ public void testRankDocsRetrieverWithCollapse() {
assertThat(resp.getHits().getAt(1).field(TOPIC_FIELD).getValue().toString(), equalTo("astronomy"));
assertThat(resp.getHits().getAt(2).getId(), equalTo("doc_1"));
assertThat(resp.getHits().getAt(2).field(TOPIC_FIELD).getValue().toString(), equalTo("technology"));
assertThat(resp.getHits().getAt(2).getInnerHits().get("a").getHits().length, equalTo(3));
assertThat(resp.getHits().getAt(2).getInnerHits().get("a").getAt(0).getId(), equalTo("doc_4"));
assertThat(resp.getHits().getAt(2).getInnerHits().get("a").getAt(1).getId(), equalTo("doc_3"));
assertThat(resp.getHits().getAt(2).getInnerHits().get("a").getAt(2).getId(), equalTo("doc_1"));
Expand All @@ -347,17 +355,15 @@ public void testRankDocsRetrieverWithCollapse() {
});
}

public void testRankDocsRetrieverWithCollapseAndAggs() {
// same as above, but we only want to bring back the top result from each subsearch
// so that would be 1, 2, and 7
// and final rank would be (based on score): 2, 1, 7
// aggs should still account for the same docs as the testRankDocsRetriever test, i.e. all but doc_5
public void testRankDocsRetrieverWithNestedCollapseAndAggs() {
final int rankWindowSize = 10;
SearchSourceBuilder source = new SearchSourceBuilder();
StandardRetrieverBuilder standard0 = new StandardRetrieverBuilder();
// this one retrieves docs 1 and 6 as doc_4 is collapsed to doc_1
standard0.queryBuilder = QueryBuilders.constantScoreQuery(QueryBuilders.queryStringQuery("quick").defaultField(TEXT_FIELD))
.boost(10L);
standard0.queryBuilder = QueryBuilders.boolQuery()
.should(QueryBuilders.constantScoreQuery(QueryBuilders.idsQuery().addIds("doc_1")).boost(10L))
.should(QueryBuilders.constantScoreQuery(QueryBuilders.idsQuery().addIds("doc_4")).boost(9L))
.should(QueryBuilders.constantScoreQuery(QueryBuilders.idsQuery().addIds("doc_6")).boost(8L));
standard0.collapseBuilder = new CollapseBuilder(TOPIC_FIELD).setInnerHits(
new InnerHitBuilder("a").addSort(new FieldSortBuilder(DOC_FIELD).order(SortOrder.DESC)).setSize(10)
);
Expand All @@ -375,8 +381,8 @@ public void testRankDocsRetrieverWithCollapseAndAggs() {
null
);
// the compound retriever here produces a score for a doc based on the percentage of the queries that it was matched on and
// resolves ties based on actual score, rank, and then the doc (we're forcing 1 shard for consistent results)
// so ideal rank would be: 6, 2, 1, 4, 7, 3
// resolves ties based on actual score, and then the doc (we're forcing 1 shard for consistent results)
// so ideal rank would be: 6, 2, 1, 3, 4, 7
source.retriever(
new CompoundRetrieverWithRankDocs(
rankWindowSize,
Expand All @@ -392,7 +398,7 @@ public void testRankDocsRetrieverWithCollapseAndAggs() {
ElasticsearchAssertions.assertResponse(req, resp -> {
assertNull(resp.pointInTimeId());
assertNotNull(resp.getHits().getTotalHits());
assertThat(resp.getHits().getTotalHits().value, equalTo(5L));
assertThat(resp.getHits().getTotalHits().value, equalTo(6L));
assertThat(resp.getHits().getTotalHits().relation, equalTo(TotalHits.Relation.EQUAL_TO));
assertThat(resp.getHits().getAt(0).getId(), equalTo("doc_6"));
assertNotNull(resp.getAggregations());
Expand Down Expand Up @@ -427,8 +433,8 @@ public void testRankDocsRetrieverWithNestedQuery() {
null
);
// the compound retriever here produces a score for a doc based on the percentage of the queries that it was matched on and
// resolves ties based on actual score, rank, and then the doc (we're forcing 1 shard for consistent results)
// so ideal rank would be: 6, 2, 1, 4, 3, 7
// resolves ties based on actual score, and then the doc (we're forcing 1 shard for consistent results)
// so ideal rank would be: 6, 2, 1, 3, 4, 7
source.retriever(
new CompoundRetrieverWithRankDocs(
rankWindowSize,
Expand Down Expand Up @@ -460,8 +466,10 @@ public void testRankDocsRetrieverMultipleCompoundRetrievers() {
SearchSourceBuilder source = new SearchSourceBuilder();
StandardRetrieverBuilder standard0 = new StandardRetrieverBuilder();
// this one retrieves docs 1, 4, and 6
standard0.queryBuilder = QueryBuilders.constantScoreQuery(QueryBuilders.queryStringQuery("quick").defaultField(TEXT_FIELD))
.boost(10L);
standard0.queryBuilder = QueryBuilders.boolQuery()
.should(QueryBuilders.constantScoreQuery(QueryBuilders.idsQuery().addIds("doc_1")).boost(10L))
.should(QueryBuilders.constantScoreQuery(QueryBuilders.idsQuery().addIds("doc_4")).boost(9L))
.should(QueryBuilders.constantScoreQuery(QueryBuilders.idsQuery().addIds("doc_6")).boost(8L));
StandardRetrieverBuilder standard1 = new StandardRetrieverBuilder();
// this one retrieves docs 2 and 6 due to prefilter
standard1.queryBuilder = QueryBuilders.constantScoreQuery(QueryBuilders.termsQuery(ID_FIELD, "doc_2", "doc_3", "doc_6")).boost(20L);
Expand Down Expand Up @@ -506,11 +514,11 @@ public void testRankDocsRetrieverMultipleCompoundRetrievers() {
assertThat(resp.getHits().getTotalHits().value, equalTo(6L));
assertThat(resp.getHits().getTotalHits().relation, equalTo(TotalHits.Relation.EQUAL_TO));
assertThat(resp.getHits().getAt(0).getId(), equalTo("doc_4"));
assertThat(resp.getHits().getAt(1).getId(), equalTo("doc_6"));
assertThat(resp.getHits().getAt(1).getId(), equalTo("doc_1"));
assertThat(resp.getHits().getAt(2).getId(), equalTo("doc_2"));
assertThat(resp.getHits().getAt(3).getId(), equalTo("doc_1"));
assertThat(resp.getHits().getAt(4).getId(), equalTo("doc_7"));
assertThat(resp.getHits().getAt(5).getId(), equalTo("doc_3"));
assertThat(resp.getHits().getAt(3).getId(), equalTo("doc_3"));
assertThat(resp.getHits().getAt(4).getId(), equalTo("doc_6"));
assertThat(resp.getHits().getAt(5).getId(), equalTo("doc_7"));
});
}

Expand Down Expand Up @@ -545,9 +553,9 @@ public void testRankDocsRetrieverDifferentNestedSorting() {
assertThat(resp.getHits().getTotalHits().relation, equalTo(TotalHits.Relation.EQUAL_TO));
assertThat(resp.getHits().getAt(0).getId(), equalTo("doc_4"));
assertThat(resp.getHits().getAt(1).getId(), equalTo("doc_1"));
assertThat(resp.getHits().getAt(2).getId(), equalTo("doc_7"));
assertThat(resp.getHits().getAt(2).getId(), equalTo("doc_2"));
assertThat(resp.getHits().getAt(3).getId(), equalTo("doc_6"));
assertThat(resp.getHits().getAt(4).getId(), equalTo("doc_2"));
assertThat(resp.getHits().getAt(4).getId(), equalTo("doc_7"));
});
}

Expand Down Expand Up @@ -673,22 +681,14 @@ private RankDoc[] getRankDocs(SearchResponse searchResponse) {
for (int i = 0; i < size; i++) {
var hit = searchResponse.getHits().getAt(i);
long sortValue = (long) hit.getRawSortValues()[hit.getRawSortValues().length - 1];
int doc = decodeDoc(sortValue);
int shardRequestIndex = decodeShardRequestIndex(sortValue);
int doc = ShardDocSortField.decodeDoc(sortValue);
int shardRequestIndex = ShardDocSortField.decodeShardRequestIndex(sortValue);
docs[i] = new RankDoc(doc, hit.getScore(), shardRequestIndex);
docs[i].rank = i + 1;
}
return docs;
}

public static int decodeDoc(long value) {
return (int) value;
}

public static int decodeShardRequestIndex(long value) {
return (int) (value >> 32);
}

record RankDocAndHitRatio(RankDoc rankDoc, float hitRatio) {}

/**
Expand Down
1 change: 1 addition & 0 deletions server/src/main/java/org/elasticsearch/TransportVersions.java
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -231,6 +231,7 @@ static TransportVersion def(int id) {
public static final TransportVersion CCS_REMOTE_TELEMETRY_STATS = def(8_755_00_0);
public static final TransportVersion ESQL_CCS_EXECUTION_INFO = def(8_756_00_0);
public static final TransportVersion REGEX_AND_RANGE_INTERVAL_QUERIES = def(8_757_00_0);
public static final TransportVersion RRF_QUERY_REWRITE = def(8_758_00_0);

/*
* STOP! READ THIS FIRST! No, really,
Expand Down
2 changes: 2 additions & 0 deletions server/src/main/java/org/elasticsearch/action/search/TransportSearchAction.java
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -502,6 +502,8 @@ void executeRequest(
});
final SearchSourceBuilder source = original.source();
if (shouldOpenPIT(source)) {
// disabling shard reordering for request
original.setPreFilterShardSize(Integer.MAX_VALUE);
openPIT(client, original, searchService.getDefaultKeepAliveInMillis(), listener.delegateFailureAndWrap((delegate, resp) -> {
// We set the keep alive to -1 to indicate that we don't need the pit id in the response.
// This is needed since we delete the pit prior to sending the response so the id doesn't exist anymore.
Expand Down
3 changes: 0 additions & 3 deletions server/src/main/java/org/elasticsearch/common/lucene/Lucene.java
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -74,7 +74,6 @@
import org.elasticsearch.index.analysis.NamedAnalyzer;
import org.elasticsearch.index.fielddata.IndexFieldData;
import org.elasticsearch.lucene.grouping.TopFieldGroups;
import org.elasticsearch.search.retriever.rankdoc.RankDocsSortField;
import org.elasticsearch.search.sort.ShardDocSortField;

import java.io.IOException;
Expand Down Expand Up @@ -553,8 +552,6 @@ private static SortField rewriteMergeSortField(SortField sortField) {
return newSortField;
} else if (sortField.getClass() == ShardDocSortField.class) {
return new SortField(sortField.getField(), SortField.Type.LONG, sortField.getReverse());
} else if (sortField.getClass() == RankDocsSortField.class) {
return new SortField(sortField.getField(), SortField.Type.INT, sortField.getReverse());
} else {
return sortField;
}
Expand Down
2 changes: 1 addition & 1 deletion ...er/src/main/java/org/elasticsearch/lucene/search/uhighlight/CustomUnifiedHighlighter.java
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -260,7 +260,7 @@ public void visitLeaf(Query leafQuery) {
* KnnScoreDocQuery and RankDocsQuery requires the same reader that built the docs
* When using {@link HighlightFlag#WEIGHT_MATCHES} different readers are used and isn't supported by this query
*/
if (leafQuery instanceof KnnScoreDocQuery || leafQuery instanceof RankDocsQuery) {
if (leafQuery instanceof KnnScoreDocQuery || leafQuery instanceof RankDocsQuery.TopQuery) {
hasUnknownLeaf[0] = true;
}
super.visitLeaf(query);
Expand Down
Loading