Abstract
Large-scale dense retrieval systems often suffer from performance degradation due to representational space misalignment between query and item towers and inconsistencies in the retrieval index. We present the Symmetric Consistent Indexing (SCI) framework, which addresses these issues through a novel input-swapping training mechanism (SymmAligner) and a dual-tower indexing strategy (Consistent Indexing). SCI achieves a 9.9% relative improvement in MRR@10 on the MS MARCO benchmark without increasing inference latency or parameter count.
Introduction
Modern search and recommendation engines rely on dense retrieval, typically implemented via dual-tower neural networks. These models encode queries and items into a shared vector space, enabling millisecond-level searches across billions of items through clustering-based indexing (e.g., IVF-PQ). However, two fundamental flaws often limit their effectiveness:
- Representational Space Misalignment: Since query and item towers are often trained on asymmetric data, they may learn disparate "languages," placing matching pairs in different neighborhoods of the embedding space.
- Retrieval Index Inconsistency: Indexing typically uses the item tower for cluster construction, while retrieval uses the query tower. This discrepancy leads to "drift," where queries are routed to incorrect clusters, particularly for long-tail queries.
The SCI Framework
The SCI framework introduces two synergistic modules designed to unify the representation space and ensure indexing consistency.
SymmAligner: Symmetric Representation Alignment
To bridge the gap between towers, SymmAligner employs an input-swapping mechanism during training. In addition to standard training, queries are passed through the item encoder and items through the query encoder. A Swap Loss ($L_{swap}$) is introduced to enforce consistent similarity relationships regardless of which encoder is used. This process encourages the model to converge into a unified, isotropic semantic space without the need for shared parameters, which often degrade performance due to the inherent differences between query and item text.
Consistent Indexing (CI)
Standard indexing strategies rely solely on the item tower. SCI proposes a "Dual-Vector Strategy" to resolve retrieval-time inconsistency:
- Structural Vectors: The query tower is used to encode items for coarse clustering. This ensures the "map" used during retrieval is drawn in the same "language" as the incoming query.
- Representation Vectors: The item tower encodes fine-grained residuals within clusters, preserving rich, item-specific metadata.
Experimental Evaluation
The framework was evaluated on two scales: the industry-standard MS MARCO (MS300k) benchmark and a proprietary Industrial E-commerce Dataset from JD.com featuring 10 million user-item interactions.
Key Results
On the MS MARCO benchmark, SCI (with an alignment weight $\lambda=0.3$) achieved an MRR@10 of 0.448, a significant improvement over the 0.408 baseline. In the industrial setting, the framework delivered a 4.0% relative gain in Recall@100 and a 9.3% gain in NDCG@100. Notably, the "worst-case" alignment (minimum similarity for ground-truth pairs) improved by over 50%, indicating significantly higher reliability for difficult matches.
Discussion and Analysis
A key advantage of SCI is its avoidance of the "Seesaw Effect"—the performance drop often seen when forcing query and item towers to share parameters. By using loss-based alignment rather than architectural constraints, SCI preserves the specialized encoding capabilities of each tower. Furthermore, the Swap Loss acts as a regularizer that reduces anisotropy, ensuring embeddings are more uniformly distributed across the vector space.
Crucially, SCI introduces zero latency overhead during online inference. The improvements are baked into the model weights and the offline index structure, making it a "free" upgrade for production environments.
Conclusion
The SCI framework provides a robust solution to the alignment and consistency problems in dense retrieval. While the input-swapping mechanism increases training complexity, the resulting gains in accuracy and reliability are substantial. Future work may explore the application of SCI to Generative Recommendation, where high-quality semantic IDs are critical for performance.
Summary updated