ChronoRAG: Chronology-Aware RAG

• ChronoRAG is a retrieval-augmented generation framework that prioritizes narrative chronology and local continuity by assembling contextually coherent passages.
• It employs a two-layer graph and offline indexing to extract relation descriptions from fixed-length chunks and group summaries to guide precise retrieval.
• Empirical evaluations on NarrativeQA demonstrate improved ROUGE-L scores, underscoring its effectiveness in mitigating fragmentation issues of standard RAG methods.

ChronoRAG is a retrieval-augmented generation framework for long-form narrative question answering that treats chronology and local continuity as first-class retrieval concerns rather than incidental context. Introduced in “Chronological Passage Assembling in RAG framework for Temporal Question Answering” (Kim et al., 26 Aug 2025), it is designed for settings in which correct answers depend not merely on locating a fact-bearing sentence, but on reconstructing a coherent sequence of events within a constrained context window. Its central mechanism, “chronological passage assembling,” retrieves fine-grained relation descriptions as keys and returns temporally ordered, context-enriched passages as values. The method is presented as an indexing and inference-time assembly architecture rather than a jointly trained parametric model.

1. Problem formulation in narrative question answering

ChronoRAG is motivated by a failure mode of standard RAG on narrative texts such as stories and scripts: semantically relevant snippets are often insufficient because their meaning depends on adjacent events and on the order in which those events occur. The paper’s central claim is that narratives are not well modeled as bags of independently retrievable snippets. In the motivating example, retrieved snippets may mention both “London” and “Paris,” yet without adjacent narrative context the system cannot determine which city corresponds to reunion and which to farewell (Kim et al., 26 Aug 2025).

The target task is long-context question answering over narrative documents, evaluated on NarrativeQA. Within that benchmark, the authors define a temporal subset called “Time Questions,” consisting of questions containing temporal cues such as “When,” “While,” “During,” “After,” and “Before.” These questions require reasoning across multiple events and their ordering. The dataset used in the experiments contains 355 stories and scripts with 10,557 question-answer pairs, and the keyword-based temporal subset contains 1,111 questions (Kim et al., 26 Aug 2025).

This framing distinguishes ChronoRAG from conventional chunk-based retrieval. Standard RAG retrieves semantically similar sentences or fixed chunks, but this can fragment the timeline and sever the sequential relations needed for narrative understanding. A plausible implication is that, for narrative QA, retrieval quality depends not only on semantic relevance but also on preserving event sequence and contextual continuity.

2. Two-layer graph and offline index construction

ChronoRAG uses a two-stage pipeline: an offline graph construction phase and an online retrieval-and-generation phase. The offline phase converts each raw document into a simple two-layer hierarchical graph. Layer 0 contains fixed-length original chunks from the document. Layer 1 contains relation-description sentences extracted from summaries of chunk groups. These Layer 1 relation descriptions serve as retrieval keys, while links back to original chunks and neighboring relation descriptions are preserved through document-order indices (Kim et al., 26 Aug 2025).

The graph construction process begins with fixed-length chunking. The paper gives 100 tokens per chunk as an example. The authors explicitly reject semantic chunking with variable-length boundaries because they want consistent token lengths for retrieval. Chunks are then grouped into larger units, with 10 chunks per group given as an example, and an LLM summarizes each group. This summarization step is intended to filter less important information and clarify the broader flow of events.

From these summaries, the system performs entity-relation extraction. The LLM is instructed to extract entities and generate relational descriptions between entities, but only the relation descriptions are retained for indexing and retrieval. The paper’s rationale is that standalone entity descriptions overlap and can disrupt narrative flow, whereas relation descriptions act like compact “atomic facts” that remain retrieval-friendly while still being narratively meaningful.

ChronoRAG then assigns narrative-order indices. Both Layer 1 relation sentences and Layer 0 chunks receive indices based on document order. Each relation sentence also stores the index of the original chunk it came from as a child index. Chronology is therefore not inferred through explicit temporal parsing or timestamp extraction; it is approximated by preserving original document order as a proxy for narrative time. This is the system’s temporal backbone.

The final offline step is neighborhood assembling. The paper distinguishes the retrieved item’s role as a key from its role as a value. The key is the individual Layer 1 relation-description sentence used for similarity retrieval. The value is a passage assembled from neighboring Layer 1 passages concatenated in index order. This allows retrieval to remain fine-grained while the evidence passed to generation remains locally coherent.

3. Inference-time retrieval and chronological passage assembling

At inference time, ChronoRAG performs hierarchical retrieval. A question first retrieves high-precision Layer 1 relation descriptions using embedding similarity, then uses child indices to recover the corresponding Layer 0 chunks, and finally assembles the selected material into a context for the generator. The implementation uses the arctic-Snowflake-embed-l model for retrieval embeddings, and the paper states that retrieval scores are computed using embedding similarity only; BM25 is not used because the authors want to avoid “distortion of the original text during generation” (Kim et al., 26 Aug 2025).

The system’s chronology handling is explicitly local. Once a Layer 1 relation is retrieved, ChronoRAG expands it with adjacent relation descriptions according to neighboring narrative-order indices, thereby creating a local passage that preserves storyline continuity. The linked Layer 0 chunks are then added to recover finer-grained evidence omitted by summaries and relation descriptions, such as dialogue or other factual detail. The paper describes this combination of Layer 1 abstraction and Layer 0 fidelity as leveraging “hierarchical granularity” to obtain a “comprehensive and balanced context.”

In the final answer-generation stage, the original query is concatenated with the retrieved context and passed to a LLM. Retrieved passages are separated by double line breaks and are ordered by relevance. The paper is explicit that final ordering is by relevance, not by a globally reconstructed chronology across all retrieved passages. Chronology is preserved locally within each assembled passage through index-order concatenation, but the system does not include a separate global temporal reranker or temporal consistency module beyond the indexing and local assembly mechanism.

ChronoRAG is therefore primarily algorithmic. The paper provides no formal equations for retrieval scoring, temporal ordering, loss functions, or generation probabilities, and it introduces no learned end-to-end retriever. It uses existing embedding models for retrieval and LLM prompting for summarization and relation extraction. This suggests that ChronoRAG is best understood as a retrieval/indexing architecture plus an inference-time passage assembly procedure.

4. Empirical performance and ablation evidence

The evaluation is centered on NarrativeQA, with ROUGE-L as the metric. The paper motivates ROUGE-L by noting that NarrativeQA answers are short and pronoun-heavy, so longest-common-subsequence overlap captures wording agreement without overly penalizing paraphrase. All compared systems use top-$k=20$ retrieval, a 1,500-token context limit, and the same generation sampling strategy. For preprocessing, the authors use meta-llama-3-8B-Instruct; for final answer generation, they use unifiedqa-v2-t5-3b-1363200 (Kim et al., 26 Aug 2025).

ChronoRAG achieves the best reported ROUGE-L on both the full NarrativeQA set and the Time Questions subset. The next best baseline is RAPTOR-CT. The ablation study isolates the contributions of passage assembling, chunk summarization, and relation extraction, and the trade-off study shows that expanding the number of adjacent sentences included in each assembled passage to two reduces performance under the fixed 1,500-token budget.

Setting	Whole Data	Time Questions
ChronoRAG	0.308	0.268
Next best baseline (RAPTOR-CT)	0.297	0.261
w/o passage assembling	0.295	0.252
w/o chunk summarization	0.272	0.233
w/o relation extraction	0.255	0.227

These results imply two distinct effects. First, relation extraction is central to retrieval precision: removing it reduces performance to the level of NaiveRAG. Second, passage assembling matters especially for temporal reasoning: removing passage assembling costs 0.016 ROUGE-L on Time Questions, which is larger than the drop on the full dataset. The linking-window study further reports that when the number of adjacent sentences included in each passage is increased to two, performance declines to 0.300 on the full dataset and 0.258 on Time Questions, suggesting that wider local expansion improves coherence only up to the point where it begins to crowd out overall coverage.

The qualitative case study in the paper reinforces this interpretation. RAPTOR retrieves broad summaries that may omit details through compression; LightRAG and GraphRAG extract entities and relations directly from original text, but GraphRAG in particular appears to over-explain many elements, producing noisy retrieval; Propositionizer and LightRAG remain too general for questions involving changes in relationships over time. ChronoRAG is described as retrieving “the minimal set of chronologically adjacent passages while suppressing unrelated narrative details,” which the authors interpret as evidence of better temporal coherence and lower retrieval noise.

5. Relation to adjacent RAG research and nomenclature

Within its own paper, ChronoRAG is positioned against several RAG families. Relative to NaiveRAG, it inserts a structured abstraction layer between raw chunks and generation. Relative to proposition-based retrieval such as Propositionizer, it uses atomic-like relation sentences as retrieval keys but does not treat them as sufficient evidence; the novelty lies in returning chronologically assembled passages rather than isolated facts. Relative to summary-tree approaches such as RAPTOR, it uses a single summary layer rather than recursive summarization trees. Relative to LightRAG and GraphRAG, it is not primarily entity-centric; the paper argues that entity-relation graphs struggle to capture the sequential relations that matter in stories (Kim et al., 26 Aug 2025).

Broader temporal RAG work highlights where ChronoRAG is specific and where it is limited. TA-RAG decomposes a query into semantic content and temporal constraints, performs interval-based temporal filtering, and orders retrieved chunks chronologically for diachronic question answering (Lau et al., 21 Jul 2025). E$^2$RAG preserves distinct entity mentions and event subgraphs to maintain temporal, causal, and character consistency in narratives (Zhang et al., 6 Jun 2025). TG-RAG models corpora as a temporal knowledge graph plus a hierarchical time graph with timestamped relations and incremental updates (Han et al., 15 Oct 2025). Compared with those systems, ChronoRAG does not extract explicit temporal intervals or timestamps; it uses document order as a proxy for narrative time. This suggests that it occupies a narrower but operationally simpler point in the design space: chronology-aware passage assembly for narratives rather than general time-conditioned retrieval over evolving corpora.

The term “ChronoRAG” also appears in a different sense in the time-series forecasting literature. In QuiZSF, the “ChronoRAG Base (CRB)” denotes a hierarchical tree-structured temporal database for storing and retrieving auxiliary time series, rather than a narrative QA framework (Ma et al., 9 Aug 2025). That usage is conceptually related only at a high level: both prioritize temporally structured retrieval, but the stored objects, retrieval objectives, and downstream models differ substantially.

6. Limitations, scope, and significance

ChronoRAG has several explicit limitations. Its chronology signal is approximated by document order, which is reasonable for many stories and scripts but does not directly address flashbacks, nonlinear storytelling, or other mismatches between textual order and event time. It does not use an explicit temporal parser, temporal graph with typed temporal edges, or learned temporal ordering model. It relies on LLM-based summarization and relation extraction, so preprocessing errors can propagate. Its empirical evaluation is limited to NarrativeQA and a keyword-defined temporal subset. Finally, because final passages are ordered by relevance rather than by a reconstructed global event chain, the method preserves local chronology within passages more strongly than it constructs a document-wide temporal ordering for generation (Kim et al., 26 Aug 2025).

These constraints clarify the significance of the method. ChronoRAG does not propose a new embedding model, a new generator, or a formal temporal reasoning objective. Its contribution is a retrieval strategy for narratives: summarize local regions, extract relation descriptions as precise keys, maintain document-order and source links, assemble adjacent passages in chronological order, and recover original chunks for detail. The paper presents this as evidence for a broader principle in narrative QA: the right retrieval unit is not necessarily the smallest unit, but the smallest unit plus enough temporal neighborhood to preserve meaning.

In that sense, ChronoRAG marks a specific shift in RAG design for narratives. It treats temporal order and adjacent-event continuity as structural properties of retrieval rather than as issues to be left entirely to the generator. The reported improvements on both full NarrativeQA and Time Questions are presented as evidence that reasoning over temporal order is crucial in resolving ambiguity and supporting narrative comprehension (Kim et al., 26 Aug 2025).