Prompt-Driven GraphRAG Framework

• Prompt-Driven GraphRAG Framework is a retrieval-augmented generation architecture that leverages explicit prompt design for graph construction and LLM retrieval to achieve state-of-the-art multi-hop reasoning.
• It couples symbolic graph traversal with prompt-guided entity extraction and fact filtering to reduce noise and enhance accuracy in answer synthesis.
• Empirical evaluations on HotpotQA and 2WikiMultiHopQA demonstrate superior performance with marked improvements in precision and reduced duplication in knowledge graph construction.

Prompt-Driven GraphRAG Framework is a class of retrieval-augmented generation (RAG) architectures in which prompt design is a first-class component governing both graph construction and runtime retrieval, enabling precise combinatorial reasoning for tasks such as multi-hop question answering, causal inference, and knowledge graph construction. Distinguishing itself from vanilla RAG and non-prompted GraphRAG, this approach tightly couples LLMs with symbolic/semantic graph traversal, leveraging prompt templating and selective prompt execution to drive entity extraction, fact selection, passage reranking, and answer synthesis. Empirical results underline its superiority on complex tasks, notably achieving state-of-the-art metrics on HotpotQA and 2WikiMultiHopQA and demonstrating marked reductions in noise and duplication in real-world knowledge graph construction (Sarnaik et al., 3 Nov 2025, Meher et al., 20 Jun 2025).

1. Framework Overview and Structural Principles

Prompt-Driven GraphRAG fundamentally operates in two phases: (a) offline graph construction from text corpora, and (b) online, prompt-mediated graph retrieval and answer generation. The workflow involves:

• Text Encoding: All passages and candidate fact triples are embedded into a shared vector space (using encoders such as text-embedding-3-large). Queries are likewise embedded at runtime.
• Symbolic Graph Construction: Entities (persons, organizations, locations) and passages become nodes. Factual triples, extracted via LLM prompts, form typed edges (subject–predicate–object). Additional edges (synonymy, co-occurrence) are introduced for dense semantic connectivity. Nodes are scored by inverse passage frequency.
• Online Prompt-Oriented Retrieval: Queries first retrieve top-k candidate triples by embedding similarity. LLM fact-filtering (keep/drop via prompt) refines these to yield high-precision seed entities. These seeds instantiate the restart vector for entity-guided Personalized PageRank (PPR) traversal on the heterogeneous graph, culminating in a reranked passage set for answer synthesis.

This framework sets prompt formulation and execution as integral operators rather than auxiliary aids, directly impacting every major system stage.

2. Prompt Design: Templates, Execution, and Optimization

Prompt-driven GraphRAG approaches employ explicit, context-aware prompt templates, systematically tuned for high extraction and retrieval precision (Sarnaik et al., 3 Nov 2025, Meher et al., 20 Jun 2025):

• Entity Extraction Prompt: Enforces high-precision recognition and format consistency.

  You are an entity recognizer. Given the passage: {passage_text} List all PERSON, ORGANIZATION, and LOCATION entities as a JSON array.

• Fact Filtering Prompt: Drives semantic selection of relevant graph edges.

  Question: {q} Candidate Facts: 1. (s1, p1, o1) ... For each fact, respond ‘KEEP’ if relevant or ‘DROP’ otherwise.

• Passage Reranking Prompt: Improves online passage selection using LLM-based scoring.

  You are a QA assistant. Question: {q} Passages: ... Rank passages in order of relevance for supporting an answer, and assign a confidence score (0-1).

Prompts are carefully structured and include explicit output formats (JSON, KEEP/DROP, ranked list). Softmax interpretation of LLM binary logits (keep/drop) allows for probabilistic fact selection:

$$p_{\mathrm{keep}}(f) = \frac{\exp(s_{\mathrm{keep}})}{\exp(s_{\mathrm{keep}}) + \exp(s_{\mathrm{drop}})}$$

with $\tau \approx 0.5$ as the selection threshold.

In pipelines targeting type-aware extraction (e.g., CORE-KG), sequential type-specific prompts are employed for coreference resolution, followed by domain-guided extraction with explicit type definitions and strict extraction order, eliminating legal boilerplate during KG construction (Meher et al., 20 Jun 2025).

3. Mathematical Formulation: Graph Construction and Traversal

Factual relationships are encoded as triples $(h, r, t)$, where $h$ and $t$ are entities and $r$ is the relation. The graph $G = (V, E)$ comprises:

• $V =$ entities and passage nodes,
• $E =$ mention-in edges, triple edges, synonymy edges, contextual co-occurrence edges.

Node and edge types are embedded by learned linear maps $W_v$ and $W_r$, with a token-sequence encoder $\varphi(\cdot)$:

$$\forall v \in V,\, \vec{v} = W_v \varphi(v) \in \mathbb{R}^d,\quad \forall r,\, \vec{r} = W_r \varphi(r) \in \mathbb{R}^d$$

For traversal, PPR is defined as:

$$v^{(t+1)} = (1-\alpha)A^\top v^{(t)} + \alpha v^{(0)}$$

where $A$ is the normalized adjacency matrix, $\alpha$ is the restart probability, and $v^{(0)}$ is the seed entity distribution.

Passage reranking incorporates both graph-derived and prompt-derived features:

$$\mathrm{score}(p|q) = v_p + \lambda_r \cdot \mathrm{overlap}(p, q)$$

where $\mathrm{overlap}(p, q)$ quantifies shared entities or facts.

4. Integration and Execution of Prompts in Online Retrieval

Prompt execution is interleaved with retrieval operations, specifically at:

• Fact Filtering: Post top-k triple retrieval, to weed out semantically irrelevant facts.
• Passage Reranking (optional): Using LLM for further semantic refinement of candidate context.
• Answer Generation: Employing evidence-first prompts composed of filtered triples, seed entities, and most relevant passages.

Combined scoring for passage relevance may integrate dense similarity, PPR score, and LLM-based confidence:

$$\mathrm{relevance}(p, q) = \beta \cos(\mathrm{emb}(p), \mathrm{emb}(q)) + (1-\beta) v_p + \gamma p_{\mathrm{prompt}}(p, q)$$

where $p_{\mathrm{prompt}}(p, q)$ is the normalized LLM-assigned confidence.

In KG construction pipelines like CORE-KG, coreference resolution is performed in sequence for each entity type using strict prompts, followed by domain-filtered extraction, post-chunk aggregation, and exact-string/type node merging (Meher et al., 20 Jun 2025).

5. Empirical Results and Ablation Analyses

State-of-the-art performance is achieved on multi-hop QA datasets:

Dataset	F1	EM	Recall@5
HotpotQA	80.7	79.9	97.1
2WikiMultiHopQA	78.9	76.4	98.1

Ablation studies (HotpotQA):

System	F1
Full PROPEX-RAG	80.7
–EntityPrompt	79.1
–FactPrompt	77.2
–RerankPrompt	78.5

Fact filtering yields the greatest marginal benefit (≈1–3 F1 points per module), confirming the importance of prompt-driven semantic selection.

For KG construction (CORE-KG):

Metric	GraphRAG Baseline	CORE-KG Adaptation	Relative Improvement
Node Duplication (%)	30.38	20.27	33.28%
Legal Noise (%)	27.41	16.89	38.37%

This suggests prompt-aware type-specific extraction and filtering significantly improve graph coherency and relevance.

6. Effects, Limitations, and Research Directions

Precision in seed entity selection sharply reduces traversal initialization noise, while fact-level filtering and LLM reranking amplify recall@5 and answer consistency. Prompt-executed semantic reasoning aligns structural (graph) and topical (query) dimensions. Notable limitations include increased latency and compute due to multiple LLM invocations, sensitivity to prompt wording, and inflexibility in dynamic graph updates. The requirement to iterate and tune prompt engineering remains a major factor, with future directions emphasizing adaptive prompt scheduling and reinforcement learning for end-to-end optimization (Sarnaik et al., 3 Nov 2025).

7. Position Within Contemporary GraphRAG Landscape

Prompt-driven GraphRAG frameworks ("PROPEX-RAG", "CORE-KG", and modular designs such as "LEGO-GraphRAG") illustrate a paradigm in which prompts are structured, sequential, and context-rich, directly influencing both symbolic and neural modules. This stands in contrast to purely embedding-based retrieval and maximally fusion-centric RAG approaches. The embedding of prompt logic within graph construction, traversal, and answer generation underpins superior empirical results in complex reasoning (multi-hop QA, legal networks, biomedical document analysis) (Sarnaik et al., 3 Nov 2025, Meher et al., 20 Jun 2025, Cao et al., 6 Nov 2024).

Prompt-aware design emerges as the decisive mechanism for high-fidelity knowledge extraction, robust multi-hop reasoning, and rapid domain adaptation, setting a benchmark for the next generation of RAG and GraphRAG research.