Disco-RAG: Discourse-Aware RAG Framework

• Disco-RAG is a framework that integrates intra-chunk discourse trees and inter-chunk rhetorical graphs to transform flat evidence into a coherent argumentative flow.
• It leverages RST-based parsing to identify core versus supporting content, enabling precise mapping of rhetorical relations across retrieved passages.
• The system uses a discourse-driven planning blueprint to condition LLM-based generation, resulting in improved correctness, coherence, and factuality.

Disco-RAG is a discourse-aware retrieval-augmented generation framework that explicitly models rhetorical structure to bridge the gap between what retrievers deliver—“a bag of relevant chunks”—and what generators need—“a coherent argumentative flow that synthesizes dispersed evidence” (Liu et al., 7 Jan 2026). It was introduced to address a limitation of standard RAG pipelines, which typically segment documents into chunks, retrieve Top-$k$ passages, and prepend them to the query for generation, thereby treating evidence in a flat and unstructured way. Disco-RAG augments this process by constructing intra-chunk discourse trees, building inter-chunk rhetorical graphs, and deriving a discourse-driven planning blueprint that conditions generation. The resulting system is an inference-time, training-free pipeline that aims to improve correctness, coherence, and factuality on knowledge-intensive question answering and long-document summarization (Liu et al., 7 Jan 2026).

1. Conceptual basis and motivation

Disco-RAG is motivated by the observation that standard RAG pipelines ignore both intra-chunk structure and inter-chunk coherence. Intra-chunk structure concerns which sentences are core versus supporting and which rhetorical relations hold within a passage. Inter-chunk coherence concerns how claims connect, contrast, or entail across documents. According to the paper, flat concatenation encourages overgeneralization, poor conflict resolution, and weak synthesis, especially when evidence is conditional or distributed (Liu et al., 7 Jan 2026).

The framework therefore introduces discourse structure as a prior over retrieval context. It parses local rhetorical hierarchies inside each retrieved chunk via Rhetorical Structure Theory (RST), infers a global rhetorical graph across retrieved chunks, and uses both structures to produce a planning blueprint that guides answer generation. This organization changes the role of retrieval context from a flat set of passages into a structured scaffold for synthesis (Liu et al., 7 Jan 2026).

A central claim of the method is that rhetorical function matters in addition to topical relevance. Rather than linking entities or facts alone, Disco-RAG models relations such as support, contradiction, elaboration, background, and cause/consequence. This permits the generation model to prioritize nuclei, incorporate satellites as evidence, order arguments, and surface conditions or caveats when evidence is dispersed or conflicting (Liu et al., 7 Jan 2026).

2. End-to-end architecture and formalization

At inference, Disco-RAG follows five stages: retrieval, intra-chunk discourse tree parsing, inter-chunk rhetorical graph construction, discourse-driven planning, and structure-guided generation (Liu et al., 7 Jan 2026).

Stage	Function	Stated implementation
Retrieval	Select evidence chunks	Fixed-size chunks of 256 tokens; Qwen3-Embedding-8B; cosine similarity; Top-10
Intra-chunk parsing	Build local discourse trees	LLM-based RST parser with EDU segmentation and relation assignment
Inter-chunk graph	Model cross-passage coherence	Directed rhetorical relations over retrieved chunks in a single listwise pass
Planning	Produce discourse-aware blueprint	Prompted planner over $(q, C, T, G)$
Generation	Produce final answer	LLM conditioned on $C, T, G, B$; beam search width 3

Retrieval uses fixed-size chunking with a default chunk size of 256 tokens and no sliding window. A dense retriever, Qwen3-Embedding-8B, computes semantic embeddings, and retrieval uses cosine similarity to return Top-10 chunks for a query in open-domain settings such as Loong and ASQA, or is restricted to the source article in closed-domain summarization such as SciNews. The paper states that retrieval settings are kept fixed across baselines (Liu et al., 7 Jan 2026).

The standard RAG objective is written as

$y=\arg\max_{y'} P(y' \mid q, C(q;D)).$

Disco-RAG augments this with structure:

$y=\arg\max_{y'} P(y' \mid q, C, T, G, B).$

With autoregressive decoding, the generation distribution becomes

$$P(y \mid q, C, T, G, B)=\prod_t P(y_t \mid y_{<t}, q, C, T, G, B).$$

This formulation makes explicit that output generation is conditioned not only on retrieved chunks $C$, but also on local discourse trees $T$, the global graph $G$, and the plan $B$ (Liu et al., 7 Jan 2026).

Disco-RAG is described as training-free. All components use structured prompts; the generator employs beam search with width 3; and the system uses only task instructions without in-context examples (Liu et al., 7 Jan 2026).

3. Intra-chunk discourse trees and inter-chunk rhetorical graphs

For each chunk $(q, C, T, G)$0, Disco-RAG employs an LLM-based RST parser to perform Elementary Discourse Unit (EDU) segmentation and discourse parsing, producing a local hierarchical tree $(q, C, T, G)$1 that differentiates core content (nucleus) from supporting content (satellite) and labels relations such as Elaboration, Contrast, and Cause (Liu et al., 7 Jan 2026). The framework adopts Rhetorical Structure Theory to model local hierarchies: each chunk is segmented into EDUs and then organized into a binary tree whose internal nodes connect a nucleus and a satellite, or in some cases two nuclei.

The relation inventory for intra-chunk parsing is explicitly controlled and includes Elaboration, Explanation, Evidence, Example, Contrast, Comparison, Concession, Antithesis, Cause, Result, Consequence, Purpose, Condition, Temporal, Sequence, Background, Circumstance, Summary, Restatement, Evaluation, Interpretation, Attribution, Definition, and Classification (Liu et al., 7 Jan 2026). Parsing uses an LLM-based zero-shot parser via a structured prompt and jointly performs segmentation and relation assignment.

Formally, for chunk $(q, C, T, G)$2, the parser outputs a tree

$(q, C, T, G)$3

where $(q, C, T, G)$4 are EDU nodes and $(q, C, T, G)$5 are directed edges labeled with the relation set $(q, C, T, G)$6. The parsing model is abstracted as

$(q, C, T, G)$7

where the first term models EDU boundary predictions and the second term models relation type probabilities between EDUs, including nuclearity orientation (Liu et al., 7 Jan 2026).

The operational algorithm is prompt-guided and bottom-up: the parser segments $(q, C, T, G)$8 into EDUs, identifies the most central EDU as the root nucleus, decides nucleus/satellite roles for other EDUs, assigns relation labels from the allowed list, and composes a binary tree bottom-up. The resulting tree provides fine-grained local salience and rhetorical intent signals (Liu et al., 7 Jan 2026).

Across retrieved chunks $(q, C, T, G)$9, Disco-RAG constructs a directed graph

$C, T, G, B$0

where each edge $C, T, G, B$1 marks the rhetorical contribution of source chunk $C, T, G, B$2 to target chunk $C, T, G, B$3 via relation $C, T, G, B$4, or UNRELATED if no meaningful relation exists (Liu et al., 7 Jan 2026). The inter-chunk relation set is tailored to cross-passage interactions and includes SUPPORTS, CONTRADICTS, ELABORATES, EXEMPLIFIES, CAUSES, RESULTS_FROM, ENABLES, PREVENTS, PRECEDES, FOLLOWS, SIMULTANEOUS, BACKGROUND_FOR, GENERALIZES, SPECIFIES, COMPARES_WITH, CONTRASTS_WITH, SUPPLEMENTS, REPLACES, MOTIVATES, JUSTIFIES, and UNRELATED.

Relations are predicted listwise by the same parser in a single pass over all retrieved chunks:

$C, T, G, B$5

The paper states that this listwise strategy lets the parser consider global context, increasing consistency of relation directions and labels across chunk pairs (Liu et al., 7 Jan 2026). Disco-RAG does not perform numerical message passing over $C, T, G, B$6; instead, the graph is injected via prompt conditioning to guide selection and ordering of evidence during generation.

4. Planning blueprint and generation behavior

Conditioned on the query $C, T, G, B$7, the intra-chunk trees $C, T, G, B$8, and the inter-chunk graph $C, T, G, B$9, a planner module produces a discourse-aware blueprint $y=\arg\max_{y'} P(y' \mid q, C(q;D)).$0 (Liu et al., 7 Jan 2026). The paper defines the planner mapping as

$y=\arg\max_{y'} P(y' \mid q, C(q;D)).$1

$y=\arg\max_{y'} P(y' \mid q, C(q;D)).$2 is a natural-language blueprint that describes the rhetorical plan: which nuclei to prioritize, which satellites to include as evidence, how to order contrasting claims, how to surface conditions and scope, and how to resolve conflicts, for example via Contrast and Background relations (Liu et al., 7 Jan 2026).

The planner therefore serves as an explicit intermediate representation between retrieval and decoding. Rather than relying on the generator to infer rhetorical organization from a flat prompt, Disco-RAG supplies a plan that specifies order, emphasis, supporting evidence, and conflict handling. This suggests that the method attempts to externalize argumentative structure before generation, rather than leaving coherence formation entirely to the decoder.

In the final stage, the answer is generated by an LLM conditioned on the retrieved chunks $y=\arg\max_{y'} P(y' \mid q, C(q;D)).$3, the local trees $y=\arg\max_{y'} P(y' \mid q, C(q;D)).$4, the global graph $y=\arg\max_{y'} P(y' \mid q, C(q;D)).$5, and the plan $y=\arg\max_{y'} P(y' \mid q, C(q;D)).$6 (Liu et al., 7 Jan 2026). The intended effect is that the decoder attends to nuclei and salient cross-chunk links, follows the planned argumentative flow, and integrates supporting satellites when they strengthen claims or provide explanations.

Disco-RAG does not introduce a reranker. The paper instead states that discourse structures and the plan are used downstream to organize and filter evidence during generation (Liu et al., 7 Jan 2026). Robustness analyses are reported to show stable performance across chunk sizes and larger $y=\arg\max_{y'} P(y' \mid q, C(q;D)).$7, indicating that structural guidance helps integrate and dampen noise without explicit reranking.

5. Empirical evaluation and ablation findings

Disco-RAG is evaluated on Loong, ASQA, and SciNews (Liu et al., 7 Jan 2026). Loong is described as long-context multi-document QA, ASQA as ambiguous long-form QA, and SciNews as long-document lay summarization. Metrics follow official protocols: Loong uses Exact Match and an LLM-based LLM Score; ASQA uses Exact Match, ROUGE-L, and Disambiguation Recall; SciNews uses ROUGE-L, BERTScore, SARI, and SummaC.

On Loong, with Llama-3.3-70B, Disco-RAG achieves LLM Scores of 71.00 on Set 1 (10K–50K tokens), 63.61 on Set 2 (50K–100K), 58.86 on Set 3 (100K–200K), and 54.62 on Set 4 (200K–250K), compared with standard RAG scores of 62.78, 53.77, 45.77, and 35.61 respectively (Liu et al., 7 Jan 2026). Averaged overall, Disco-RAG obtains 62.07 and surpasses training-based StructRAG, reported at 60.38 under comparable reporting. The paper states that gains increase with input length, highlighting that discourse structure is crucial for long-context synthesis.

On ASQA, Disco-RAG yields consistent improvements over standard RAG. With Llama-3.3-70B, Exact Match rises from 38.2 to 42.0, ROUGE-L from 37.2 to 42.3, and Disambiguation Recall from 24.1 to 32.8 (Liu et al., 7 Jan 2026). The system is reported to beat strong prompting-based systems including MAIN-RAG, with ROUGE-L 42.0, and Tree of Clarifications, with ROUGE-L 39.7, while achieving the best Disambiguation Recall with Qwen2.5-72B at 33.2.

On SciNews, with Llama-3.3-70B, Disco-RAG achieves ROUGE-L 21.11, BERTScore 65.67, SARI 44.37, and SummaC 69.48 (Liu et al., 7 Jan 2026). These results surpass standard RAG and often prior state of the art, including RSTformer at ROUGE-L 20.12 and SARI 41.56, while approaching the best factual consistency result from a prior plan-conditioned baseline, SummaC 72.40.

Ablation results on Loong with Llama-3.3-70B quantify the contribution of each component. Removing the RST tree lowers overall LLM Score from 62.07 to 56.22; removing the inter-chunk graph yields 57.10; removing planning yields 59.75 (Liu et al., 7 Jan 2026). Two generic planning baselines built on standard RAG—retrieve-and-plan and plan-and-retrieve—improve only modestly over standard RAG, with LLM Scores 50.64 and 51.38 versus 49.33. The paper interprets this as evidence that hierarchical and cross-passage discourse modeling drives much of the gain.

Perturbation analyses show that randomly corrupting relation labels, flipping edge directions, or removing steps reduces performance substantially (Liu et al., 7 Jan 2026). This is consistent with the claim that Disco-RAG is sensitive to the quality of the induced discourse structure.

6. Qualitative behavior, efficiency, limitations, and related terminology

The paper provides three case studies to illustrate how discourse structure changes synthesis behavior (Liu et al., 7 Jan 2026). In a question about The Lion King on Broadway, Disco-RAG integrates preview and official opening dates and a later relocation by leveraging sequence and background relations, whereas standard RAG returns only “1997.” In a question about Italy’s 2006 World Cup venue, Disco-RAG identifies Berlin’s Olympiastadion via support and background relations across chunks, while standard RAG answers Munich. In a question about Pluto’s planetary status, Disco-RAG uses Contrast and Background to distinguish historic from current definitions and avoid an outdated categorical answer.

Human evaluation on SciNews, using 15 samples and three raters, shows that Disco-RAG improves Relevance (2.40), Simplicity (2.43), Conciseness (2.27), and Faithfulness (2.53) versus full-context and standard RAG systems; Fleiss’ $y=\arg\max_{y'} P(y' \mid q, C(q;D)).$8 indicates substantial agreement (Liu et al., 7 Jan 2026). Attention analysis is reported to suggest that discourse conditioning reduces attention entropy and raises factual consistency, measured by SummaC.

In efficiency terms, Disco-RAG introduces two additional LLM calls at inference: one for the inter-chunk graph and one for the plan (Liu et al., 7 Jan 2026). RST trees are constructed offline for the corpus. On Loong, token cost is approximately $y=\arg\max_{y'} P(y' \mid q, C(q;D)).$9 that of standard RAG and end-to-end latency increases by approximately 2 seconds per query, averaged across $y=\arg\max_{y'} P(y' \mid q, C, T, G, B).$0. Experiments use a cluster with $y=\arg\max_{y'} P(y' \mid q, C, T, G, B).$1A100 80GB GPUs. The paper states that the method is stable across chunk sizes, remains strong up to Top-50 retrieval, and degrades less under injected noisy passages.

Several limitations are identified. Disco-RAG depends on a zero-shot LLM parser for RST; on RST-DT it attains competitive nuclearity (63.1 F1) and relation (58.6 F1) relative to a supervised baseline, but span F1 is lower (70.4 versus 79.8) (Liu et al., 7 Jan 2026). Acceptability studies report values of 0.72 for trees, 0.80 for graphs, and 0.93 for plans, with $y=\arg\max_{y'} P(y' \mid q, C, T, G, B).$2 ranging from 0.709 to 0.862. The paper notes that errors in segmentation or labels can propagate, that domain transfer beyond Loong, ASQA, and SciNews remains to be studied, and that LLM-based judgment on Loong and automatic metrics have known limitations.

Disco-RAG is situated alongside GraphRAG, KG-guided RAG, GRAG, HybGRAG, StructRAG, RAS, ConTReGen, MMAIN-RAG, and FLARE (Liu et al., 7 Jan 2026). The stated distinction is that Disco-RAG models rhetorical structure at two levels—hierarchical intra-chunk trees and inter-chunk rhetorical graphs—and integrates both into a planning blueprint that directly conditions decoding. Rather than emphasizing topological connectivity, entity graphs, iterative retrieval, or flat active retrieval, it focuses on how pieces of evidence function rhetorically and how they should be orchestrated.

A terminological ambiguity appears in DF-RAG, which states: “If by ‘Disco-RAG’ you mean a diversity-conscious retrieval-augmented generation system, DF-RAG is exactly that” (Khan et al., 23 Jan 2026). The same source further notes that if “Disco-RAG” refers to a specific prior method, it is not defined in that paper. In the arXiv paper titled “Disco-RAG: Discourse-Aware Retrieval-Augmented Generation,” however, the term unambiguously denotes the discourse-aware framework described above (Liu et al., 7 Jan 2026). The overlap is therefore lexical rather than methodological: DF-RAG concerns query-aware diversity in retrieval, whereas Disco-RAG, in the specific sense of (Liu et al., 7 Jan 2026), concerns discourse-aware structuring of retrieved evidence.