Release Summary 25.08 | Apr 17, 2025

ENG-30035

Science:

Query Enrichment Mapping Flow for Manual Entries

A new flow has been created to populate client-provided manual query enrichment mappings into the find.query_enrichment_mappings PSQL table. The flow captures fields such as query, enrichment terms, language, type, source, and timestamps, enabling structured storage and future retrieval.

ENG-30016

Science:

PSQL Table for Query Enrichment Configuration

A new PSQL table find.query_enrichment_config has been created to store site-level query enrichment configurations. This includes fields such as language, suffixes, and synonym mapping preferences in JSON format. The table supports scalable and customizable enrichment behavior for each site.

ENG-30034

Science:

Find Query Enrichment Pipeline Enhancement

The enrich_query() function has been integrated just before the Find query embedding stage. This enhancement ensures that enrichment is applied using site-specific configurations from the PSQL tables. Additionally, the output file format has been updated to include a third column—[enriched_document]—capturing the enriched query alongside the original document and its embedding.

ENG-29897

Science:

Ensemble AI Model Build: Performance Optimizations

Several improvements have been made to optimize the Ensemble AI model build process. These include better Spark task distribution, reduced network IO, and simplified logic to minimize processor time. Key enhancements include repartitioning data for even distribution, using data broadcasting, caching, and removing redundant computations. These updates ensure faster, more efficient model execution and retain older models when a new build fails.

ENG-29394

Science:

Ensemble AI: Shopper Feedback Score Calculation

A feedback score is now computed for each ensemble to enhance personalization based on shopper interactions such as "more like this" or "less like this." The score is calculated using default attributes including brand and total price bucket, which is based on the combined prices of products within the ensemble. A decay-based formula is applied using multipliers stored in the user profile system (UPS) to reflect recent shopper preferences.

The feedback score helps identify ensembles that align more closely with a shopper’s interests and will be used to influence future ensemble rankings during recommendation delivery.

ENG-30051

Social Proof:

Social Proof Messaging API: Force Variation ID

The Social Proof Messaging API now supports a force_varId parameter, allowing you to always play a specific variation’s messaging—bypassing traffic allocation rules. This is especially useful for running external A/B tests using tools like Adobe. If optimization is enabled, the forced variation will return optimized messages as well.

ENG-30001

Find:

Monitoring: Catalog Vector Fields in Solr

A new hourly monitor script now checks for catalog vector fields (vector_1024_*) in Solr for configured sites. If any fields are missing, it automatically triggers a Slack alert. This ensures timely detection of vector indexing issues.

ENG-30122

Add Enriched Query Field to Query Vector PDC

To improve debugging and transparency in embedding generation, a new enriched_query field has been added to the default property definition collection (PDC) for queryvector. This ensures that the exact enriched input text used for vector generation is stored and visible in Solr's MetaStore. New PDCs will now include this field by default.

ENG-30123

Query Vector Job Updated to Push Enriched Query

The Query Vector job has been enhanced to include the enriched query text alongside the query vector during streaming ingestion. This allows visibility into the exact enriched input used for embedding, aiding in debugging and ensuring enriched queries are properly captured in Solr's MetaStore Collection.

ENG-30124

Patch Existing Query Vector PDCs to Support Enriched Query

The existing Query Vector Property Definition Collections (PDCs) have been patched to include the enriched_query field. This ensures enriched queries are ingested without requiring new MetaStore snapshots, aligning with recent updates to query vector ingestion.

ENG-30128

Standardized Index and Primary Key Setup for Catalog Embedding Tables

A changescript has been added to update the primary key and indexes for find_catalog_embedding tables across all sites, including those where hybrid search was not previously enabled.

ENG-30005

SFI Compatibility Check for Streaming Catalog Update

Sanity and regression tests were run on streaming catalog v1.26.0 to verify the impact of recent site partition logic changes. No issues were found with SFI consuming from the streaming.engine.out topic.

ENG-30042

Recommend:

Consistent Rule Application on Backfill Strategies

Recommendation restriction rules, such as attribute diversity, are now applied uniformly across both base and backfill strategies. This ensures consistent filtering of products, regardless of the strategy source.

ENG-30149

Accurate Timing for Ebclient Search Service

The Ebclient timer has been updated to measure only the actual search service time, excluding any preprocessing overhead. This ensures more precise monitoring of search performance.

ENG-29967

Enterprise Dashboard:

Update Support Portal Link in Dashboard-Dynamic

The Support customer portal link in the RR Portal has been updated in the header.jsp of the dashboard-dynamic.

PLAT-3988

Externalized Kafka Partition Allocation to Postgres

The site partition allocation logic for streaming has been moved from code to a Postgres configuration table. This allows easier assignment of dedicated Kafka partitions to sites without requiring code changes. Existing partition logic and hash-based allocation for shared partitions remain unchanged.

PLAT-3975

Streaming-recommend:

Improved Identity Lookup in Streaming Itemstore Consumer

Fixed an issue where association identities weren’t retrieved locally due to separate MapDB instances, causing unnecessary identity service calls and slowing ingestion. Local lookups now work as expected, improving performance.

ENG-30012

Enterprise Dashboard:

Seed Validation Error in Configurable Strategies for Top Sellers

Resolved an issue where users encountered a "no seed specified" error in Configurable Strategies, even after switching to a model like Top Sellers that doesn't require a seed. Users can now switch models and view results without being blocked by incorrect validation.

ENG-29979

Enterprise Dashboard:

Variables Section Display Issue in Comparison Placement Template

Fixed an issue where the Variables sections in the Comparison placement template were not displaying correctly, preventing users from expanding or collapsing them.

ENG-29985

Find:

Package Upload Failure on Fresh Solr Setup

Resolved a NullPointerException that occurred during fresh Solr installations where SFI and Find Deployer failed to upload packages to the Solr package store. The issue was caused by malformed JSON handling in the package helper logic.

ENG-29854

Recommend:

Localization Fallback to Default Region

The fallback mechanism to return default region (English) is now correctly applied across all relevant areas.

ENG-30145

Build Failures Due to Missing language_tag Column in find_catalog_embeddings

Fixed an issue causing multiple build and PR failures related to missing language_tag column in the find_catalog_embeddings table. The failure occurred during test setups where the altered primary key and index creation expected this column to exist. The fix ensures that the column is added if missing, allowing builds to proceed without errors.

ENG-30133

Find:

Disable Reranking When Sort Order Is Not by Score

Resolved an issue where reranking was incorrectly applied even when users selected a sort order other than relevance (score descending), such as price low to high. This caused the results to appear out of order. Reranking is now skipped unless the sort order is explicitly set to score descending.

ENG-30130

Boosting Rules Not Fully Applied

Fixed an issue where not all boosting rules were being applied during search execution. Although rules were saved successfully, only a subset were reflected in the Solr query. Now, all configured boost rules are correctly incorporated into the search query as expected.