CombSUM

CombSUM operates on the principle of linear additivity, summing normalized scores from component rankers. Each document's final score is the arithmetic sum of its scores across all input systems. This assumes score scales are comparable and that each system contributes independent evidence of relevance. The formula is straightforward: CombSUM(d) = Σ s_i(d) where s_i(d) is the normalized score from system i.

Effective CombSUM fusion critically depends on score normalization. Raw scores from different systems—such as cosine similarity from a dense retriever and BM25 from a sparse retriever—exist on incomparable scales. Common normalization techniques include:

• Min-Max Scaling: Rescales scores to the [0,1] range
• Z-Score Normalization: Centers scores around zero with unit variance
• Sum-to-Unity: Divides each score by the sum of all scores in the ranked list Without proper normalization, a system with larger raw score magnitudes will dominate the fusion.

Unlike Reciprocal Rank Fusion (RRF), which operates purely on rank positions, CombSUM leverages the actual relevance scores. This makes CombSUM more sensitive to the confidence differences between documents. RRF treats the gap between rank 1 and rank 2 identically to the gap between rank 100 and 101, while CombSUM preserves the score deltas. However, RRF requires no normalization, making it more robust when score distributions are unknown or miscalibrated.

CombSUM belongs to a family of linear combination operators. Two common variants modify the aggregation function:

• CombMAX: Takes the maximum score across systems for each document, favoring documents strongly supported by at least one system
• CombMIN: Takes the minimum score, requiring consensus across all systems CombSUM represents a middle ground, rewarding documents that perform well across multiple systems while not penalizing strong single-system performance as severely as CombMIN.

A natural extension is Weighted CombSUM, where each system's contribution is scaled by a learned or heuristic weight: CombSUM(d) = Σ w_i * s_i(d). Weights can be determined by:

• Grid search over validation queries to optimize NDCG
• Linear regression using relevance judgments as targets
• Domain expertise assigning higher weight to more reliable systems This transforms CombSUM from an unsupervised fusion method into a simple learned combination.

CombSUM is frequently employed in two-stage retrieval architectures to fuse results from multiple first-stage retrievers before re-ranking. For example, a pipeline might:

1. Retrieve candidates using both dense vector search and sparse BM25
2. Apply CombSUM to merge the two candidate sets into a single pool
3. Pass the fused candidates to a computationally expensive cross-encoder for final re-ranking This approach maximizes recall by combining complementary retrieval signals before the precision-focused re-ranking stage.

CombSUM is widely used to fuse results from diverse retrieval models to improve robustness. By summing the normalized scores from a dense vector retriever, a sparse lexical retriever (BM25), and a cross-encoder re-ranker, the final ranking leverages the strengths of each.

• Semantic Coverage: Dense models capture conceptual meaning.
• Precision: Sparse models ensure exact keyword matching.
• Depth: Cross-encoders provide fine-grained relevance assessment.

This aggregation often outperforms any single model, reducing the risk of catastrophic failure from one retrieval strategy.

In a two-stage retrieval architecture, CombSUM acts as the final aggregation layer. A fast first-stage retriever (e.g., Bi-Encoder) returns hundreds of candidates. Multiple specialist re-rankers then score this candidate pool independently.

• Score Aggregation: CombSUM sums the normalized scores from each re-ranker.
• Efficiency: Avoids running a single, massive ensemble model over the entire corpus.
• Flexibility: New re-rankers can be added to the pool without retraining the entire system.

The final merged list provides a consensus view that balances diverse relevance signals.

E-commerce platforms use CombSUM to merge rankings optimized for different business objectives into a single, balanced results page.

• Semantic Relevance Score: How well the product matches the user's query.
• Business Metric Score: Normalized values for Click-Through Rate (CTR), conversion rate, or profit margin.
• Personalization Score: Affinity based on user purchase history.

By summing these normalized scores, the final ranking balances pure relevance with business goals, avoiding a results page that is either irrelevant or unprofitable.

For complex research queries, a single string may not capture the user's full intent. CombSUM fuses results from multiple automatic query rewrites.

• Process: The original query is expanded into paraphrases and keyword variants.
• Parallel Retrieval: Each variant retrieves a ranked list independently.
• Fusion: CombSUM aggregates the scores, boosting documents that are consistently relevant across different query formulations.

This technique increases Mean Reciprocal Rank (MRR) by ensuring the top result is robust to variations in phrasing.

When no training data exists for a new document corpus, CombSUM provides a powerful, parameter-free baseline for hybrid search.

• Setup: Combine scores from a dense model (e.g., all-MiniLM-L6-v2) and BM25.
• Normalization: Apply min-max normalization to bring both score distributions to a [0, 1] range.
• Aggregation: A simple sum produces a consensus ranking without any labeled data.

This zero-shot application often rivals early-stage supervised models, making it a critical tool for rapid prototyping and cold-start scenarios.

While both are unsupervised fusion techniques, they have distinct operational profiles.

• CombSUM: Requires score normalization to be effective. It is sensitive to outlier scores and the scale of input distributions. Best when raw relevance scores are meaningful and calibrated.
• Reciprocal Rank Fusion (RRF): Operates purely on rank positions, ignoring the magnitude of relevance scores. It is inherently robust to score scale differences and requires no normalization.

Choose CombSUM when you have well-calibrated, normalized scores. Choose RRF when merging lists from black-box systems with unknown or noisy score distributions.

An unsupervised fusion alternative to CombSUM that scores documents based on the reciprocal of their rank position across multiple input lists. Unlike CombSUM, RRF does not require score normalization, making it robust to outliers and heterogeneous score distributions. It is particularly effective when combining results from dense and sparse retrievers where raw scores are not directly comparable.

A critical pre-processing step for CombSUM that transforms raw relevance scores onto a common scale. Common techniques include:

• Min-Max Scaling: Rescales scores to a [0,1] range.
• Z-Score Normalization: Centers scores around zero with unit variance.
• Sum Normalization: Divides each score by the sum of all scores. Without proper normalization, retrieval systems with larger raw score magnitudes can dominate the CombSUM aggregation.

A supervised alternative to unsupervised fusion methods like CombSUM. LTR models are trained to optimize the final ordering using pointwise, pairwise, or listwise loss functions. While CombSUM applies a fixed linear combination, LTR learns the optimal weighting of features, including the individual ranker scores themselves, often yielding superior precision at the cost of requiring labeled training data.

The architectural pattern where CombSUM typically operates. A fast first-stage retriever (e.g., vector search) fetches hundreds of candidates, and a slower second-stage re-ranker (e.g., a cross-encoder) scores them. CombSUM fuses the relevance signals from both stages to produce the final merged ranking, balancing the semantic breadth of the retriever with the precision of the re-ranker.

The standard listwise metric for evaluating the output of a CombSUM fusion. NDCG measures ranking quality by discounting relevance gains logarithmically by position, emphasizing that highly relevant documents must appear at the top. It normalizes against an ideal ranking, providing a score between 0 and 1 that directly reflects the quality of the final merged list.

The broader retrieval strategy that CombSUM enables. It fuses relevance signals from dense vector search (semantic similarity) and sparse lexical retrieval (keyword matching via BM25). By summing normalized scores from both systems, CombSUM creates a consensus ranking that captures conceptual meaning while respecting exact terminology, mitigating the weaknesses of either approach in isolation.