GraphSearch: Agentic Dual-Channel Retrieval

• GraphSearch is a collection of algorithmic frameworks that extract nodes, patterns, and evidence chains from graph-structured data through query decomposition and iterative reasoning.
• It integrates dual-channel retrieval by combining semantic text searches with relational graph queries, enabling dynamic, multi-hop evidence aggregation.
• Empirical results on multi-hop QA benchmarks demonstrate significant improvements in evidence quality and retrieval efficiency compared to traditional shallow methods.

GraphSearch encompasses a family of algorithmic frameworks and methodologies for finding nodes, patterns, evidence chains, or embedded items within graph-structured data. In contemporary research, the term applies to classic algorithmic node or edge search, semantic/structural database search, vector similarity search over proximity graphs, and advanced agentic, multi-modality searching for retrieval-augmented generation. This entry synthesizes core technical advances from foundational graph search theory to recent innovations in agentic GraphSearch workflows for graph retrieval-augmented generation (GraphRAG).

1. Agentic GraphSearch Workflow for Retrieval-Augmented Generation

GraphSearch, in the context of GraphRAG, denotes an agentic multi-stage search workflow over graph-based knowledge bases. Rather than relying on single-step semantic retrieval, GraphSearch decomposes a complex query $Q$ into a sequence of atomic subqueries $\{q_1,\ldots,q_m\}$, each targeting fine-grained evidence nodes or relations (Yang et al., 26 Sep 2025).

The workflow is modular and iterative, comprising the following sequence:

• Query Decomposition (QD): Parse $Q$ into subqueries, each focusing on a single relational hop or semantic aspect.
• Context Retrieval (CR): For each $q_i$, retrieve context $C_{q_i}$ using $\psi(q_i\,|\,G)$, where $G$ is the graph knowledge base.
• Context Refinement: Post-process $C_{q_i}$ to select pertinent entities and prune redundancy. Output: $C'_{q_i} = P_{CR}(q_i, C_{q_i})$.
• Query Grounding (QG): Fill placeholders in $q_i$ using previous answers and context: $$\tilde{q}_i = P_{QG}(q_i, \{q_{<i}, C_{q_{<i}}, \hat{a}_{q_{<i}}\})$$.
• Logic Drafting (LD): Chain $\{\tilde{q}_i\}$ and associated evidence into a logic plan $$\mathcal{L} = P_{LD}(\{\tilde{q}_i, C_{\tilde{q}_i}, \hat{a}_{q_i}\})$$.
• Evidence Verification (EV): Evaluate the coherence and completeness of $\mathcal{L}$, returning $\mathcal{V} \in \{\text{Accept}, \text{Reject}\}$.
• Query Expansion (QE): If $\mathcal{L}$ is rejected, generate further subqueries $\{q_1^+, \dots, q_n^+\}$ targeting missing evidence for the answer.

This recursive, agent-driven workflow enables deep, multi-hop reasoning and dynamic adaptation to incomplete or ambiguous evidence (Yang et al., 26 Sep 2025).

2. Dual-Channel Retrieval: Semantic and Relational Integration

GraphSearch innovates with a dual-channel evidence retrieval paradigm:

• Semantic Channel: Issues queries over chunked natural language text, retrieving $K_{q_i}^{(s)}$—descriptive, contextually rich evidence. This modality is vital for facts stated in unstructured text.
• Relational Channel: Executes queries as subject-predicate-object patterns $q_i^{(r)}$ over the graph, yielding subgraph contexts $C_{q_i}^{(r)}$ of entities and their relations. This modality excels at enforcing multi-hop, logically dependent reasoning chains where entity IDs and edge types guide retrieval.

The workflow synchronizes the strengths of both modalities, ensuring that evidence chains leverage both textual detail and graph structure. For example, inferring "Which artists who died of plague..." requires (a) finding an artist node (relation), (b) extracting the place and cause of death (relations), and (c) retrieving historical details about a plague event (semantic) (Yang et al., 26 Sep 2025).

3. Overcoming Shallow Retrieval and Graph Data Utilization Limits

Classic GraphRAG systems suffer from two core issues:

1. Shallow Retrieval: Standard single-pass retrieval often fails to assemble all evidence hops; critical entities or facts may be absent from the answer context. This leads to broken logical chains and incorrect generations.
2. Inefficient Graph Utilization: Structural graph data may be present but is used only with heuristic or co-occurrence-based retrieval, neglecting deeper multi-relational chains or precise navigation of knowledge graphs.

GraphSearch's agentic, modular workflow enables iterative evidence aggregation, query expansion, and evidence verification modules to mitigate these deficits. Each module is designed to explicitly ground subqueries, verify logical chains, and ensure that both simple and multi-hop dependencies are resolved, often requiring multiple retrieval rounds to surface missing nodes or paths (Yang et al., 26 Sep 2025).

4. Experimental Results and Empirical Performance

Across six multi-hop QA benchmarks—including HotpotQA, MuSiQue, and 2WikiMultiHopQA—GraphSearch demonstrated substantial performance gains:

• On MuSiQue, integrating GraphSearch with LightRAG increased SubEM (exact match on sub-answers) from 35.00 to 51.00, with corresponding improvements in A-Score and E-Score (metrics for answer and evidence quality).
• GraphSearch maintained its advantage even as the retrieval budget (Top-$K$) decreased, indicating efficient evidence prioritization per budget (Yang et al., 26 Sep 2025).
• When combined with different graph KB retrievers, it consistently outperformed native retrieval methods on both open-domain and domain-specific (Medical, Agriculture, Legal) datasets.

These results confirm the generality and practical utility of GraphSearch for advanced retrieval-augmented generation pipelines.

5. Technical Formulations

Key algorithmic and formal components underpin the workflow:

• Subquery Evidence Selection: $C_{q_i} = \psi(q_i | G)$,
• Query Grounding Step: $$\tilde{q}_i = P_{QG}(q_i, \{q_{<i}, C_{q_{<i}}, \hat{a}_{q_{<i}}\})$$,
• Logic Draft Plan: $$\mathcal{L} = P_{LD}(\{\tilde{q}_i, C_{\tilde{q}_i}, \hat{a}_{q_i}\})$$.

All modules act as composable prompt-based or control-flow procedures, with evidence selection, grounding, and chaining formalized to enable plug-and-play integration in retrieval-augmented LLM contexts.

6. Implications, Applications, and Future Directions

GraphSearch advances GraphRAG by enabling:

• Deep, iterative, and multi-hop retrieval, directly addressing the limitations of shallow, non-reflective pipelines.
• Hybrid reasoning over both structured and unstructured data, improving factual QA grounded in both graph and text modalities.
• Domain and retriever generality, with applications spanning open-domain QA, scientific and medical knowledge integration, and legal reasoning.

Future work may further explore:

• Integration with reinforcement learning and fine-tuning for agentic component selection.
• Extension to multimodal corpora (e.g., image-graph-text fusion).
• Increasing efficiency and scalability in large graph KBs while maintaining high recall.
• Advanced prompt engineering and module optimization for more complex logical dependencies.

7. Comparative Summary Table

Feature	Traditional GraphRAG	GraphSearch Framework
Retrieval Depth	Single-step, shallow	Multi-turn, iterative
Evidence Modalities	Predominantly semantic (text)	Semantic + relational (graph KB)
Query Decomposition	Atomic or none	Explicit, agentic
Reasoning Chain Drafting	Minimal or implicit	Modular logic drafting and verification
Retrieval Adaptivity	Static (1 pass)	Iterative, dynamic expansion
Empirical Performance	Baseline	Consistent, significant improvements

This progression from traditional approaches to agentic, dual-channel deep search marks a significant step forward in efficient, accurate, and contextually grounded reasoning in graph-augmented generative systems.