Context-Aware Bidirectional Retrieval Approach

• Context-Aware Bidirectional Retrieval is a paradigm that jointly models queries and target items using contextual cues such as user intent, schema metadata, and temporal structure.
• It employs iterative, dual-stage, or joint optimization techniques to enhance retrieval accuracy by integrating feedback between retrieved candidates and their contexts.
• This approach demonstrates significant empirical gains, including improved recall rates and lower latency, making it effective for multimodal and interactive retrieval systems.

A context-aware bidirectional retrieval approach is a class of retrieval framework in which contextual signals—ranging from user intent, document context, schema metadata, or temporal structure—inform a retrieval process that integrates information flow in both directions between the target (documents, moments, schema elements) and the query (or user/context). This paradigm is distinguished by the mutual conditioning of representations during retrieval or reranking, typically involving iterative, dual-stage, or joint optimization procedures. Notable instantiations traverse natural language passage retrieval with personalized context (Wan et al., 2023), text-to-SQL schema linking (Nahid et al., 16 Oct 2025), information retrieval from scientific documents (Rakotoson et al., 2024), video moment localization (Nguyen-Nhu et al., 12 Apr 2025), and context-enriched neural reranking (Chen et al., 2021). These methods yield substantial empirical improvements over unidirectional or context-agnostic baselines, particularly in zero-shot, long-context, or low-latency settings.

1. Bidirectional Retrieval Principles

The bidirectional retrieval paradigm posits that optimal item selection (documents, schema elements, temporal segments) is best achieved by jointly modeling the relevance between the query and candidate items, as well as between candidate items and associated context. In classical single-stage retrieval, candidate selection flows strictly from query to target, often collapsing critical contextual cues. In contrast, bidirectional retrieval frameworks employ:

• Joint or alternating passes: first retrieving passages (or tables, columns, keyframes) for a query, then retrieving contexts (scenarios, user info, neighboring schema elements, local feature clusters) conditioned on the selected candidates, or vice versa.
• Mutual conditioning: allowing retrieved items to inform, reshape, or calibrate the scoring of both the query and its candidate set, as in pseudo-relevance feedback or cross-document attention.
• Aggregation of complementary retrieval trajectories, typically via union or convex combination, to maximize recall and minimize false positives with context-dependent weighting.

In "How Can Context Help? Exploring Joint Retrieval of Passage and Personalized Context" (Wan et al., 2023), this is realized as the Personalized Context-Aware Search (PCAS) procedure, which for a given query $q$, context candidates $C$, and passage candidates $P$, seeks the pair $$(c^*, p^*) = \arg\max_{c\in C, p\in P} S_{\text{joint}}(q, c, p)$$, where $S_{\text{joint}}$ blends query–passage and context–passage relevance.

2. Methodological Instantiations

2.1 Text-to-SQL Schema Linking

In schema linking for Text-to-SQL, "Rethinking Schema Linking: A Context-Aware Bidirectional Retrieval Approach" (Nahid et al., 16 Oct 2025) advances a two-path retrieval strategy: table-first retrieval (filtering tables, then selecting columns) and column-first retrieval (filtering columns, then selecting tables), each employing LLM-based inference grounded in question keywords and decomposed subquestions. Contextual augmentation is realized through LLM-driven keyword and keyphrase extraction and question decomposition, with bidirectional aggregation via union of selected table and column sets:

$$S_{\text{retrieved}} = ( \{t\}\times(\operatorname{Att}(t)\cap C_T) \ | \ t\in T^* ) \cup ( \{t\}\times(\operatorname{Att}(t)\cap C^*) \ | \ t\in T_C ).$$

Empirical findings indicate this bidirectional synergy narrows the gap to oracle schema recall by ≈50% while halving false positives compared to unidirectional or full-schema baselines.

2.2 Video Moment Retrieval

The GRAB system ("A Lightweight Moment Retrieval System with Global Re-Ranking and Robust Adaptive Bidirectional Temporal Search" (Nguyen-Nhu et al., 12 Apr 2025)) exemplifies bidirectional retrieval in high-resolution video moment localization. Its pipeline consists of:

• Shot detection and keyframe deduplication via pHash clustering ($\tau=0.8$).
• BEiT-3 based global feature embedding for all deduplicated keyframes indexed in FAISS.
• SuperGlobal Reranking using GeM pooling over "top-M" global descriptors for both query and neighborhood, with a scoring function $S_{\mathrm{final}} = \frac{S_1 + S_2}{2}$, where $S_1$ and $S_2$ are cosine similarities involving pooled query and neighbor embeddings.
• Adaptive Bidirectional Temporal Search (ABTS), where, given a pivot, the start and end frame boundaries are localized by maximizing a composite confidence combining similarity and a temporal “stability” score within bidirectionally adaptive local windows.

This approach achieves sub-200ms interactive latencies and reduces storage requirements by ≥60% via deduplication, with recall@1 improved by ≈15–20% after global re-ranking.

2.3 Neural Contextual Rerankers

Co-BERT ("A Context-Aware BERT Retrieval Model Incorporating Local and Query-specific Context" (Chen et al., 2021)) incorporates bidirectional context at two levels: (1) pseudo-relevance feedback (PRF) refines the query–document interaction vector with context from top-m PRF candidates, via a small transformer; (2) groupwise scoring incorporates local context from neighboring candidates using groupwise transformer scoring with overlapping batches and softmax normalization. Empirical improvements are observed: +2.0% nDCG@20 on Robust04 and +14.7% on GOV2 over BERT-Base (MS MARCO fine-tuned).

2.4 Scientific Document Retrieval

In "Science Checker Reloaded: A Bidirectional Paradigm for Transparency and Logical Reasoning" (Rakotoson et al., 2024), document retrieval leverages a combination of ontology-driven query expansion (with entity synonym/hypernym augmentation) for sparse retrieval and multi-hop (bidirectional) chunk routing inside target documents, fusing the results with reciprocal rank fusion and re-ranking. User, document, and field context are injected at each stage, and bidirectionality is realized both in dual-stage pipeline structure and in feedback from document-level signals back to upstream retrieval.

3. Context Sensitivity and Augmentation

Context-aware mechanisms underpin all bidirectional frameworks by maximizing the retrieval model's ability to leverage:

• Query-side context: derived from subquestion decomposition, keyword/keyphrase extraction, or user intent modeling.
• Target-side context: document structure (sections, fields), schema metadata (sample values), local candidate neighborhoods, or historical interaction signals.
• User/contextual signals: provided via explicit user context sets (Wan et al., 2023), domain ontologies (Rakotoson et al., 2024), or scenario descriptions.

These signals are fed into ranking, aggregation, or calibration modules to refine selection at each stage. Notably, (Nahid et al., 16 Oct 2025) shows LLM-based decomposition improves schema recall and FPR by ~3–5pp and 2–4pp respectively.

4. Empirical Results and Evaluation

Context-aware bidirectional retrieval approaches consistently outperform unidirectional and context-agnostic baselines across modalities, as shown by:

Method	Task/Domain	Key Gains	Ref
PCAS	Passage+context	+8.1% R@1, +17.3% MAP@5 over baseline	(Wan et al., 2023)
GRAB (video ABTS+SuperGlobal)	Video retrieval	−60% storage, +15–20% recall@1, <200 ms latency	(Nguyen-Nhu et al., 12 Apr 2025)
Co-BERT	Ad-hoc text ranking	+6.6% MAP@1K, +14.7% nDCG@20 on GOV2	(Chen et al., 2021)
Schema Bidirectional	Text-to-SQL schema	Halves FPR (to ≤25%) with recall ≥90%	(Nahid et al., 16 Oct 2025)
Science Checker Reloaded	Sci. doc. retr.	nDCG@10=64.8 (sparse, no embeddings)	(Rakotoson et al., 2024)

These gains reflect improved ranking accuracy, storage/memory efficiency, and interactive latency, attributable to the mutual flow of contextual signals.

5. Complexity, Scalability, and Resource Considerations

The shift to bidirectional and context-aware pipelines introduces tradeoffs:

• Memory/computation: SuperGlobal reranking (O(M·D)), PRF/groupwise transformers (~2–10% extra FLOPs), and LLM calls for context augmentation yield modest overhead relative to gain in retrieval quality.
• Inference efficiency: e.g., schema retrieval in (Nahid et al., 16 Oct 2025) achieves ~2.6 s average latency with only 6 LLM calls, compared to >47 s for more monolithic approaches.
• Storage: Keyframe deduplication in video moment retrieval reduces storage by ≥60% (Nguyen-Nhu et al., 12 Apr 2025).
• Index footprint: Sparse, expansion-based approaches (e.g., Science Checker Reloaded) require ~32 GB for 10M documents, versus 0.1–0.6 TB for dense retrieval.

6. Limitations and Future Directions

Despite empirical gains, context-aware bidirectional retrieval faces several limitations:

• LLM-based inference stages introduce non-determinism and hallucination risk (Nahid et al., 16 Oct 2025).
• Multiple retrieval passes or joint scoring may raise inference latency relative to minimal single-shot retrieval.
• Highly similar or abstract context elements can evade accurate linkage or calibration, as seen in schema-linking error analysis (Nahid et al., 16 Oct 2025).
• Most examined approaches employ fixed (zero-shot or template) augmentation; integration of learned, end-to-end ranking is an open direction (Wan et al., 2023, Nahid et al., 16 Oct 2025).

A plausible implication is that advances in compressed or hierarchical representation learning, denser retrieval pipelines, and adaptive context modeling will further drive bidirectional retrieval efficacy and efficiency.

7. Summary and Significance

Context-aware bidirectional retrieval approaches unify the strengths of context modeling, dual-path selection, and joint aggregation to advance state-of-the-art in retrieval across natural language, multimedia, and structured domains. This retrieval paradigm, observed in PCAS (Wan et al., 2023), GRAB (Nguyen-Nhu et al., 12 Apr 2025), bidirectional schema linkers (Nahid et al., 16 Oct 2025), and hybrid dense/sparse rerankers (Rakotoson et al., 2024), offers scalable, explainable, and empirically superior alternatives to unidirectional or context-agnostic pipelines, and is poised to become foundational in next-generation multimodal and interactive retrieval systems.