Temporal GraphRAG: Time-Aware QA Framework

• Temporal GraphRAG is a framework that augments retrieval-augmented generation with explicit temporal modeling by representing evolving corpora as bi-level temporal graphs.
• It employs multi-granularity temporal summarization and a personalized PageRank-based retrieval pipeline to ensure semantically and temporally aligned evidence extraction.
• An incremental graph update protocol minimizes re-computation, enabling efficient, time-sensitive question answering even on dynamic, continuously evolving datasets.

Temporal GraphRAG (TG-RAG) is a Retrieval-Augmented Generation (RAG) framework that incorporates explicit temporal modeling into knowledge-intensive question answering tasks. By representing evolving corpora as temporally-disambiguated graphs and designing update-efficient, temporally-sensitive retrieval workflows, TG-RAG directly addresses challenges in grounding LLMs on dynamic, time-sensitive factual knowledge. The framework improves answer accuracy, supports efficient incremental updates, and achieves robust performance even in the presence of complex temporal reasoning queries (Han et al., 15 Oct 2025).

1. Bi-Level Temporal Graph Representation

TG-RAG encodes external corpora as a bi-level temporal graph to capture both fine-grained timestamped facts and their hierarchical temporal organization. The lower level is a temporal knowledge graph (TKG) $G^{\text{TKG}} = (\mathcal{V}, \mathcal{E}, T)$, where each edge $\varepsilon = (v_i, r, v_j, t)$ denotes an entity-relation-entity quadruple at timestamp $t$. Identical facts at different times are kept as distinct edges (e.g., “Company X revenue $Y_1$ in $t_1$” vs. $Y_2$ in $t_2$), preserving the full evolutionary record.

Above this, the temporal hierarchy graph $G^{\text{Time}} = (T, H)$ organizes all timestamps $T$ into a parent-child lattice (e.g., day → month → year), with each time node $\tau$ linked to its covered corpus intervals. Cross-layer edges connect facts to the relevant hierarchical time intervals, enabling both local (fine-grained) and global (coarse-grained) temporal access (Han et al., 15 Oct 2025).

2. Multi-Granularity Temporal Summarization

TG-RAG maintains a summary $\varepsilon = (v_i, r, v_j, t)$0 for each node $\varepsilon = (v_i, r, v_j, t)$1 in the temporal hierarchy. For leaves, $\varepsilon = (v_i, r, v_j, t)$2 is an LLM-guided condensation of all facts with $\varepsilon = (v_i, r, v_j, t)$3. For internal nodes, $\varepsilon = (v_i, r, v_j, t)$4 recursively summarizes its own facts combined with its children’s summaries:

$\varepsilon = (v_i, r, v_j, t)$5

where $\varepsilon = (v_i, r, v_j, t)$6 is an LLM-prompted summarization function, $\varepsilon = (v_i, r, v_j, t)$7 is the set of edges timestamped $\varepsilon = (v_i, r, v_j, t)$8, and $\varepsilon = (v_i, r, v_j, t)$9 are the children of $t$0.

This structure supports both specific (“What did X report on Q1 2022?”) and abstract (“How have revenues trended since 2020?”) QA, as summaries at multiple granularities are available on demand (Han et al., 15 Oct 2025).

3. Incremental Graph Update Protocol

TG-RAG supports efficient ingestion of new data via a protocol designed to minimize re-computation:

1. Temporal Quadruple Extraction: New corpus segments are processed through an LLM to extract $t$1 tuples.
2. Graph Merging: Entities, edges, and time nodes are extended as needed; only genuinely new time nodes are inserted into the temporal hierarchy.
3. Summary Propagation: Fresh summaries are generated for new leaf time nodes. For each ancestor of a newly created node, summaries are regenerated to incorporate underlying changes, but unaffected hierarchy remains untouched.

This process ensures that updates are restricted to the impacted segments of the graph, vastly reducing the recomputation and prompt token overhead relative to approaches that require (re-)summarization of the entire corpus (Han et al., 15 Oct 2025).

4. Temporally-Aware Evidence Retrieval for Question Answering

Inference in TG-RAG comprises a dynamic, temporally-filtered retrieval pipeline:

1. Query Time Identification: Temporal expressions are extracted from the input query using LLMs, determining the relevant time nodes $t$2.
2. Query Subgraph Construction: The query embedding $t$3 is compared (cosine similarity) to candidate edge embeddings $t$4; the top-K relation edges form the dynamic subgraph $t$5, filtered to $t$6.
3. Personalized Graph Propagation: Personalized PageRank is run on $t$7, with the seed set $t$8 consisting of entities from temporally-matching edges. Output entity scores guide fine-grained evidence selection.
4. Context Construction: Edge and chunk scores derived from entity centrality and semantic match (cosine) guide assembly of a local evidence set $t$9. Summaries $Y_1$0 for $Y_1$1 are included for contextual breadth.
5. Answer Synthesis: An LLM integrates atomic points from both raw facts and summaries, applying a self-assessment filter for confidence and synthesizing the final answer.

This retrieval mechanism ensures that all retrieved evidence is both semantically and temporally aligned with the query, addressing the central challenge of time-consistent factual grounding (Han et al., 15 Oct 2025).

5. Evaluation: ECT-QA Benchmark and Temporal QA Performance

TG-RAG is benchmarked on ECT-QA, a time-sensitive question answering dataset curated from 480 quarterly earnings call transcripts (2020–2024), encompassing 1,105 specific questions and 100 abstract/trend questions. The protocol assesses both pre- and post-incremental update performance, including adaptation to queries on newly-added data and the efficiency of update workflows.

Experimental results demonstrate that TG-RAG surpasses prior baselines in temporal QA performance and update efficiency:

• Specific QA, Historical (2020–2023): TG-RAG achieves Correct = 0.599 vs. GraphRAG’s 0.405, at 6.3M prompt tokens (GraphRAG: 37.1M).
• Post-Incremental Update: Minimal correct rate drop (0.599 → 0.587); update cost is 1.6M tokens, compared to GraphRAG’s 30M tokens.
• New Queries on New Data: Correct = 0.617 for TG-RAG; GraphRAG: 0.398.
• Abstract/Trend QA: Pairwise wins in temporal coverage (0.889), comprehensiveness (0.833), and diversity and overall metrics against GraphRAG (Han et al., 15 Oct 2025).

Ablation studies confirm the critical roles of temporal filtering, personalized graph propagation, and timestamp-resolved indexing, as their removal leads to substantial performance degradation.

6. Distinctions Against Related Temporal Retrieval Frameworks

TG-RAG distinguishes itself from other temporal GraphRAG variants by several design decisions:

• Explicit modeling of timestamped facts as distinct graph edges (no aggregation of facts across times), reducing temporal ambiguity.
• Hierarchical time graph and multi-granular summaries, enabling both precise and trend-level responses.
• Incremental summary updates guided by temporal hierarchy, avoiding global graph reprocessing.
• Personalized PageRank (PPR)-based, time-matched retrieval at inference, providing temporally scoped and semantically relevant evidence.

Alternative frameworks such as T-GRAG (Li et al., 3 Aug 2025) and STAR-RAG (Zhu et al., 19 Oct 2025) also use temporal knowledge graphs and multi-stage retrieval. However, T-GRAG emphasizes temporal subgraph filtering and query decomposition for multi-time queries; STAR-RAG introduces MDL-summarized rule graphs and enforces temporal proximity via seeded personalized PageRank but does not focus on incremental update protocols or hierarchical summaries in the manner of TG-RAG.

7. Practical Limitations and Generalization

TG-RAG introduces additional complexity in graph construction, multi-level indexing, and staged update propagation as compared to classic RAG pipelines. While its evaluation demonstrates strong results on financial QA tasks, domain transfer and automated fine-tuning of all retrieval and generation components remain open research areas. The modular, update-efficient temporal graph backbone suggests promising applicability to broader evolving-corpus QA tasks and other temporally-conditioned knowledge reasoning domains (Han et al., 15 Oct 2025).

A plausible implication is that such explicit, temporally-structured RAG frameworks will be increasingly critical as LLMs are deployed in settings with continuous data evolution, high granularity of time, and the need for up-to-date, time-resolved answers.