Type-Specific Weighted RRF

• Type-Specific Weighted Reciprocal Rank Fusion (wRRF) is a framework that extends standard RRF by incorporating type-specific weights for adaptive and interpretable ranking.
• It utilizes a piecewise scoring formula where each candidate's rank is modulated by a type-dependent weight, improving retrieval performance in multimodal and graph-based scenarios.
• Empirical results show that wRRF outperforms unweighted methods, offering efficient control over trade-offs between precision and recall in large-scale retrieval pipelines.

Type-Specific Weighted Reciprocal Rank Fusion (wRRF) is a scoring and reranking framework for information retrieval that extends the standard Reciprocal Rank Fusion (RRF) method by incorporating type- or modality-specific weights. Unlike ordinary RRF, which fuses ranked lists from multiple sources or modalities under uniform assumptions, wRRF assigns a distinct weight—based on item type, signal informativeness, or modality quality—to each candidate or class of candidates before aggregation. This enables adaptive, interpretable, and empirically superior fusion in retrieval pipelines spanning multimodal video, large-scale knowledge graphs, and agent/tool repositories (Samuel et al., 26 Mar 2025, Nizar et al., 22 Nov 2025).

1. Mathematical Formulation

In type-specific wRRF, each item $e$ from a corpus $\mathcal{C}$—partitioned into distinct types (e.g., modalities or node classes)—receives a fusion score that modulates reciprocal rank by a type- or data-dependent weight. The formulation generalizes standard RRF, transitioning from uniform aggregation to piecewise, type-sensitive scoring.

For item $e$ of type $t$: $s_{\mathrm{wRRF}}(e) = \frac{\alpha_t}{k + r(e)}$ where:

• $r(e)$ is the global rank of $e$ within the candidate set, based on initial retrieval,
• $k>0$ is a damping constant (e.g., $k=60$ for agent/tool reranking (Nizar et al., 22 Nov 2025), $k=0$ in multimodal video (Samuel et al., 26 Mar 2025)),
• $\mathcal{C}$0 is the type-specific weight.

In multiretriever or multimodal scenarios, this model can incorporate per-item weights $\mathcal{C}$1 (for modality $\mathcal{C}$2 and item $\mathcal{C}$3), giving: $\mathcal{C}$4 with $\mathcal{C}$5 and $\mathcal{C}$6 the set of modalities (Samuel et al., 26 Mar 2025).

2. Weight Estimation Strategies

Modality-Informativeness in Multimodal Retrieval

For the MMMORRF system (Samuel et al., 26 Mar 2025), "text-informativeness" weights $\mathcal{C}$7 for each video $\mathcal{C}$8 are estimated at indexing time by issuing a fixed, semantically targeted query (e.g., "news anchor live coverage...") to the visual retrieval pipeline. The resulting SigLIP frame-based retrieval scores $\mathcal{C}$9 are normalized via min–max scaling: $e$0 This process correlates $e$1 with the likelihood that text-based retrieval (ASR+OCR) will be effective for $e$2.

Type-Specific Constants for Graph-Based Retrieval

Agent-as-a-Graph (Nizar et al., 22 Nov 2025) defines two scalars, $e$3 and $e$4, representing agent and tool node preferences, respectively. These are tuned over a discrete grid (e.g., $e$5), optimizing for recall or nDCG. The unweighted baseline is $e$6; empirical maximum achieved at $e$7.

3. Algorithmic Outline

Type-specific wRRF typically involves the following steps, as instantiated in both video and agent/tool retrieval applications:

1. Retrieve separate ranked candidate lists for each type or modality (e.g., vision, text; agent, tool).
2. Merge candidate lists and assign global ranks for all candidates.
3. Assign each candidate a type- or data-dependent weight (either $e$8 or $e$9).
4. Compute wRRF fusion scores using the above formula.
5. Sort candidates by fusion score; break ties by original rank if needed.
6. In graph contexts, traverse ownership or parent edges to select the final set of representatives (e.g., agents owning top-ranked tools).

This procedure combines the interpretability of mixture-of-experts approaches with the robustness of reciprocal rank damping (Samuel et al., 26 Mar 2025, Nizar et al., 22 Nov 2025).

4. Theoretical Rationale

Type-specific wRRF extends the robustness properties of RRF—where late ranks are heavily damped and fusion is insensitive to individual noise—by allowing the system to adapt fusion emphasis according to item type or per-item informativeness. This "lightweight mixture-of-experts" paradigm handles strong or weak signals heterogeneously, promoting recall and specificity in retrieval tasks with heterogeneous evidence sources.

In multimodal settings, this avoids over-reliance on a dominant modality and mitigates benchmark-induced modality bias (e.g., overprioritization of vision-language signals in video retrieval) (Samuel et al., 26 Mar 2025). In agent/tool selection, it balances coarse-grained (agent-description) versus fine-grained (tool-functionality) evidence, providing independent control over both (Nizar et al., 22 Nov 2025).

5. Empirical Results and Benchmarks

Type-specific wRRF demonstrates consistent gains over both standard and unweighted RRF baselines.

Multimodal Video Retrieval (Samuel et al., 26 Mar 2025)

Fusion Scheme	nDCG@10	Recall@10
RRF (text+vision)	0.562	0.600
Modality-aware wRRF	0.586	0.611

• Absolute improvements: nDCG@10 +4.2%, Recall@10 +1.1%.
• All improvements statistically significant (paired $t$0-test, Bonferroni-corrected).
• On TVR, Recall@10 improved from 0.537 to 0.540 (smaller but consistent).

Agent/Tool Retrieval (Nizar et al., 22 Nov 2025)

$t$1	Recall@5	nDCG@5
RRF (unweighted)	0.79–0.80	0.44–0.46
wRRF (optimal $t$2)	0.85 (+6 pp)	0.47 (+1 pp)
Unweighted graph retriever	0.83	0.46
MCPZero (prior best)	0.70	0.39 (inferred)*

*A plausible implication is that standard RRF can be considerably outperformed by type-specific weighting, especially as system scale and diversity increase.

Additional Observations

• Cross-model robustness: Recall@5 improvements consistent ($t$3 over MCPZero) with low variance across 8 embedding families (Nizar et al., 22 Nov 2025).
• wRRF reranking is computationally lightweight—no iterative or LLM-based reranking required.

6. Integration in Retrieval Pipelines

wRRF is integrated into several retrieval architectures:

• Multimodal video: MMMORRF queries vision and OCR+ASR independently, uses indexed per-video modality weights, and applies wRRF for search result fusion (Samuel et al., 26 Mar 2025).
• Knowledge-graph retrieval: Agent-as-a-Graph builds a bipartite knowledge graph, retrieves candidates from a shared embedding index, applies one-pass type-specific wRRF reranking, and performs agent selection via graph traversal (Nizar et al., 22 Nov 2025).

These integrations maintain $t$4 per-query cost in candidate retrieval and negligible cost in fusion and traversal, preserving scalability to large corpora.

7. Interpretability, Tuning, and Limitations

Type-specific wRRF provides transparent interpretability. The exclusive use of two (or, for per-item regimes, dynamically assigned) weights gives direct and independent control over how different types contribute to final ranking. This is beneficial for domain experts seeking to diagnose retrieval failures or optimize trade-offs between coverage and specificity.

Weight tuning is grid-based and computationally straightforward. No hyperparameter tuning was required for modality weight estimation in (Samuel et al., 26 Mar 2025); scalar weights in (Nizar et al., 22 Nov 2025) are determined by sweep.

A plausible implication is that wRRF's effectiveness relies on meaningful heterogeneity in type signal. In degenerate cases (where all types are equally informative, or type labels are unreliable), standard RRF may be sufficient.

• "MMMORRF: Multimodal Multilingual Modularized Reciprocal Rank Fusion" (Samuel et al., 26 Mar 2025)
• "Agent-as-a-Graph: Knowledge Graph-Based Tool and Agent Retrieval for LLM Multi-Agent Systems" (Nizar et al., 22 Nov 2025)