Hacktoberfest 2026:メンテナが10月に向けて印を付けた、オープンで初心者向けの issue。 Hacktoberfest の issue を見る

[Feature] Support hybrid search with rrf, weighted_score and mrr rankers

オープン
#407 コメント 0 件 リアクション 0 件 担当者 0 名 GitHub で見る

メンテナーはふだん 1 日以内に返信

まだ誰も着手していません。

評価

難易度
5/5
見積もり時間
1週間以上
初心者へのやさしさ
35/100
issue の種類
機能追加
明瞭さ
おおむね明確
活発さ
活発
技術スタック
cpp
領域
databases, search

調査の方向性

Start with the Java HybridSearchBuilder and HybridSearchRanker design described in the issue, then inspect the existing table-level full-text search work in #404 and vector-search prerequisite. Implement the rankers, route validation, builder, and snapshot behavior described here, with tests for ranking, ties, weights, empty routes, validation, and snapshot pinning; primary-key support also depends on #410.

索引モデルが issue の本文から書いたものです。

説明

enhancement
Search before asking
  • I searched in the issues and found nothing similar.
Motivation

Sub-issue of #399 (step 4: hybrid search).

Java Table#newHybridSearchBuilder() combines vector and full-text routes into one ranked result. It was built up over several PRs: apache/paimon#8271, apache/paimon#8288, apache/paimon#8294, apache/paimon#8324, apache/paimon#8327, apache/paimon#8336, apache/paimon#8348 and apache/paimon#8351. Paimon C++ has neither a hybrid search API nor a table-level vector search builder.

Java design:

  • Builder. HybridSearchBuilder (HybridSearchBuilderImpl) provides:

    • withPartitionFilter, withFilter
    • addVectorRoute(column, vector, limit[, weight[, options]]); the weight defaults to 1.0
    • addFullTextRoute(field, query, limit, weight)
    • withLimit
    • withRanker(String), withRrfRanker(), withWeightedScoreRanker()
    • executeLocal()

    All routes are pinned to one snapshot, and the filters are forwarded to every route. Full-text routes accept row filters since apache/paimon#9855.

  • Routes. HybridSearchRoute is either VECTOR or FULL_TEXT.

    • Weights must be finite and greater than 0.
    • Full-text routes reject route options.
    • The route builder rejects mixing vector and full-text settings.
  • Rankers. HybridSearchRanker normalizes the ranker name: blank means rrf; otherwise the name is trimmed and lower-cased.

    • rrf: Σ weight / (60 + rank + 1). Ties are broken by score descending, then row id ascending.
    • weighted_score: min-max normalize each route's scores to [0, 1] (all 1.0 when the range is 0), then compute Σ weight × normalized score.
    • mrr: Σ weight / (rank + 1).
    • The final top-k uses a heap.
  • Primary-key tables. They rank physical positions with PrimaryKeySearchRanker.weightedRrf, weightedScore and weightedMrr. Mixing primary-key physical results with global row-id results is rejected.

Solution
  1. Port the rankers and route validation. They are self-contained and can land first with unit tests.
  2. Add the hybrid search builder on top of the table-level full-text search (#404) and a table-level vector search builder.
  3. Add tests:
    • each ranker
    • tie-breaking
    • weights
    • an empty route result
    • route validation
    • snapshot pinning across routes
Anything else?
  • Prerequisite: a table-level vector search builder (Java VectorSearchBuilder / scan / read). It does not exist in Paimon C++ yet, and could be tracked separately or delivered as part of this issue.
  • Depends on #404. Primary-key hybrid search depends on #410.
Are you willing to submit a PR?
  • I'm willing to submit a PR!
主要言語
C++
スター
65
フォーク
31
平均マージ
1日 23時間
マージ済み PR(30日)
64

環境構築

はじめの一歩

  1. issue を最後まで読み、次にプロジェクトのコントリビューションガイドを読みます。
  2. 着手することを issue にコメントします — 二人が同じ作業をするのを防げます。
  3. リポジトリをフォークし、ブランチを切って変更します。
  4. issue 番号を参照したプルリクエストを送ります。

apache/paimon-cpp のほかの issue

apache/paimon-cpp の issue をすべて見る

似ている issue

C++ の issue をもっと見る

新しい issue をメールで受け取る

初心者向けの GitHub issue を短くまとめたダイジェスト。