fix: phrase query may return incorrect results if there's unknown word (

#2850)

rust/lance-index/src/scalar/inverted/builder.rs

@@ -586,6 +586,14 @@ mod tests {
             .await
             .unwrap();
         assert_eq!(row_ids.len(), Some(0));
+
+        let row_ids = invert_index
+            .search(&SargableQuery::FullTextSearch(
+                FullTextSearchQuery::new("\"lance unknown\"".to_owned()).limit(Some(10)),
+            ))
+            .await
+            .unwrap();
+        assert_eq!(row_ids.len(), Some(0));
     }
 
     #[tokio::test]

rust/lance-index/src/scalar/inverted/index.rs

@@ -108,7 +108,12 @@ impl InvertedIndex {
         let token_ids = if !is_phrase_query(&query.query) {
             token_ids.sorted_unstable().dedup().collect()
         } else {
-            token_ids.collect()
+            let token_ids = token_ids.collect::<Vec<_>>();
+            // for phrase query, all tokens must be present
+            if token_ids.len() != tokens.len() {
+                return Ok(Vec::new());
+            }
+            token_ids
         };
         self.bm25_search(token_ids, query, prefilter).await
     }