Metarank

Open-source, real-time secondary re-ranker and personalization service written in Scala. It sits between a search/recommendation engine and the user: ingest behavioral events (impressions, clicks, purchases), train a LTR model, and rerank candidate sets at serving time. Created/maintained by Roman Grebennikov and Vsevolod Goloviznin.

• Website: https://metarank.ai/
• GitHub: https://github.com/metarank/metarank
• Docs: https://docs.metarank.ai/

What it does

• Implements LambdaMART over an embedded feature-engineering pipeline (other models: Shuffle, Trending, MF/ALS collaborative filtering, BERT semantic; BPR & BERT4Rec on the roadmap).
• Retrieval-agnostic: integrates with the application, not the search engine — works downstream of Elasticsearch, OpenSearch, Solr, or Pinecone.
• YAML DSL for 20+ feature extractors: CTR/conversion rates, sliding-window counts, UA/Referer/GeoIP parsing, one-hot/label encoding, scalar item fields.
• Full feedback loop: event import → training → serving, runnable as one stateless Docker service (standalone mode) or split sub-commands (import, train, serve).
• Trains on historical click-through events aggregated into Implicit Judgments.

Architecture notes

• fs2 + cats-effect streaming pipeline; ImpressionInject synthesizes negative signals (shown-but-not-clicked).
• http4s POST /rank/{model}; ?explain=true returns per-item feature values.
• Redis state store in production (in-memory for demos); ~20–30 ms re-ranking latency budget.
• Datasets: RankLens, MSRD.

Use cases in this vault

• Hybrid Search and Learning-to-Rank with Metarank — fusing BM25 + vector retrievers via LTR
• Learn-to-Rank with OpenSearch and Metarank — external re-ranker on OpenSearch
• Metarank - Personalized Ranking That Actually Reads Your Clicks — code-level repo review

Related

• Learning to Rank · LambdaMART · Reranking · Retrieval Pipeline · Personalization · Feature Store · Implicit Judgments · Interleaving
• Implementations it builds on: XGBoost · LightGBM
• OpenSearch · Pinecone Vector DB
• Elasticsearch Learning to Rank — the in-engine alternative; Metarank is the external, retrieval-agnostic re-ranker by contrast

People

• Roman Grebennikov · Vsevolod Goloviznin — creators