Radio-RAG: Radio Domain Retrieval & Control

• Radio-RAG is a family of domain-specific systems that use retrieval-augmented generation to interpret ITU radio regulations, ORAN specifications, and network performance data.
• The methodology involves segmenting regulatory texts, embedding diverse data (text, graphs, history), and retrieving evidence with specialized pipelines for high accuracy.
• Radio-RAG frameworks integrate evidence retrieval with control logic to optimize radio resource management and regulatory compliance while minimizing computational overhead.

Radio-RAG most directly denotes a domain-specific retrieval-augmented generation system for interpreting the ITU Radio Regulations, and more broadly refers to a small but technically diverse set of radio-domain RAG systems spanning radio regulations, ORAN specification reasoning, and RAG-empowered radio resource management. In these systems, retrieved evidence is not restricted to prose documents: it can be paragraph-scale regulatory clauses, ORAN specification chunks, knowledge-graph context, or historical network states paired with prior resource-allocation outcomes. This suggests that Radio-RAG is better understood as a family of radio-domain evidence-grounded generation and control pipelines than as a single canonical architecture (Kassimi et al., 11 Sep 2025, Salan et al., 28 Nov 2025, Ahmad et al., 4 Jul 2025).

1. Scope and terminology

In the most specific current usage, “Radio-RAG” is the name of a telecom-specific RAG pipeline for reliable interpretation of radio regulations. That system centers the ITU Radio Regulations (Edition of 2024), uses retrieval over segmented regulatory text, and evaluates question answering in a legally sensitive, high-stakes setting where correct grounding in authoritative clauses is essential (Kassimi et al., 11 Sep 2025). In a broader systems sense, the label also covers RAG-empowered radio control in O-RAN-inspired 6G slicing and ORAN knowledge assistants that reason over standards, APIs, and RIC-related specifications (Salan et al., 28 Nov 2025, Ahmad et al., 4 Jul 2025).

A concise way to distinguish the main technical strands is to separate them by task and evidence type.

System	Radio-domain task	Retrieved evidence
Radio-RAG (Kassimi et al., 11 Sep 2025)	Radio-regulation QA	ITU regulatory chunks
ReLLM (Salan et al., 28 Nov 2025)	Dynamic radio resource management	Historical traffic/performance records
ORAN RAG benchmark (Ahmad et al., 4 Jul 2025)	ORAN standards QA	Vector chunks, graph context, or both

The term is also homographic. “RadioRAG” appears in radiology as an online retrieval-augmented question-answering system that retrieves Radiopaedia content in real time for radiology QA. That usage is distinct from radio-telecom applications, even though both share the same retrieve-then-generate pattern (Arasteh et al., 2024).

2. Retrieval substrates and evidence representations

The radio-regulations instantiation is a dense retrieval pipeline over the ITU Radio Regulations. Text is extracted from telecom-regulation PDFs, segmented into paragraphs, embedded with Sentence-Transformers all-MiniLM-L6-v2, and indexed in FAISS with FAISS-GPU. The reported chunk sizes are 150, 300, 500, 700, and 1000, with overlaps such as 10, 30, and 50, and the best retrieval-only result is about 97% retrieval accuracy at chunk size 700, overlap 50, and top-7 retrieval under a ROUGE-L-F1-based support-overlap criterion (Kassimi et al., 11 Sep 2025). The paper defines retrieval correctness by comparing retrieved concatenated context $R_i$ with gold supporting context $C_i$ using

$F_1^{(i)} = \mathrm{ROUGE\mbox{-}L\mbox{-}F1}(R_i, C_i),$

and counts retrieval as correct if

$$F_1^{(i)} \ge \alpha, \qquad \alpha \triangleq \gamma F_{1,\max}, \qquad F_{1,\max} \triangleq \frac{2\min(R,C)}{R+C}.$$

The ORAN benchmarking study evaluates three retrieval substrates over 74 ORAN Alliance specification documents: Vector RAG, GraphRAG, and Hybrid GraphRAG. In the vector pipeline, PDFs are loaded with PyPDFLoader, split with LangChain Recursive Text Splitters into 1024-token chunks with no overlap, embedded with Google models/embedding-001, stored in Chroma, and retrieved by cosine similarity with top-4 chunk selection. In the graph pipeline, the same ORAN text is transformed into a knowledge graph with Neo4j LLM Knowledge Graph Builder and LLMGraphTransformer, stored in Neo4j AuraDB, and queried through entity extraction and Cypher traversal. The hybrid pipeline concatenates vector-derived context first and graph-derived context second, with prompt instructions to prioritize vector context for broad coverage and graph context for relational supplementation (Ahmad et al., 4 Jul 2025).

The radio-resource-management formulation is structurally different. ReLLM does not retrieve natural-language reference passages from a standards corpus; it retrieves examples with the best historical performance and the closest traffic arrival rates from a database of previous resource allocation decisions and their corresponding network performance outcomes, using Euclidean distance as the similarity criterion. The retrieved items act as contextual exemplars in a meta-prompt for slice-aware PRB or RB reallocation, so the retrieval substrate is closer to structured case memory than to document QA-style RAG (Salan et al., 28 Nov 2025).

3. Architectural patterns and control logic

A recurrent architectural theme in Radio-RAG is that retrieval is tightly coupled to control or decision routing rather than treated as a passive context-expansion step. In ReLLM, the control loop is explicitly O-RAN-shaped: KPIs are collected through a KPM Monitoring xApp in the near-real-time RIC, forwarded to the non-real-time RIC over the A1 interface, stored in a database, and used by two agents. The first is an SLA violation detection agent that computes a sigmoid-based risk factor

$\rho_k[n]=\frac{1}{1+e^{-a_k(\epsilon_k[n]-b_k)}},$

and the second is a dynamic PRB or resource reallocation agent that is invoked only when the risk exceeds a predefined threshold $\theta$. The system-wide SLA compliance index is

$$\sigma[n] = -\sum_{\forall k \in S} w_k \rho_k[n]^2,$$

and the event-triggered design is intended to minimize unnecessary LLM inference calls while satisfying slice SLAs (Salan et al., 28 Nov 2025).

The ORAN benchmark paper is less about online control and more about architectural retrieval choice, but it reaches an analogous conclusion: retrieval strategy should be matched to the structure of radio-domain knowledge. Vector RAG is passage-centric, GraphRAG is relation-centric, and Hybrid GraphRAG uses serial dual retrieval plus prompt-level context fusion. The evaluated question set from ORAN-Bench-13K is explicitly stratified into Easy, Intermediate, and Hard, corresponding to simple QA, complex reasoning QA, and multi-hop reasoning, which operationalizes the intuition that graph-aware retrieval becomes more useful as ORAN reasoning becomes more cross-document and relational (Ahmad et al., 4 Jul 2025).

An edge/network-aware systems analogue appears in EACO-RAG, which is not a radio paper per se but is directly relevant at the architectural level. It proposes a three-tier local/edge/cloud RAG system in which a gate selects retrieval location and generation location. The decision variables are

$d^t := [r^t, g^t] \in \mathcal{D},$

and the context includes delay to cloud, delay to edge, similarity between the query and summaries in the edge database, and query complexity. The operational objective is to minimize a weighted sum of resource cost and delay cost,

$u^t(c^t, d^t) = \delta_1 \cdot u_{\text{r}^t(c^t, d^t) + \delta_2 \cdot u_{\text{d}^t(c^t, d^t),$

subject to minimum accuracy and maximum delay constraints. The paper solves this with Safe Online Bayesian Optimization over a contextual multi-armed bandit formulation, and the authors explicitly note that the tiers map naturally onto user equipment, base-station MEC or RAN edge, and regional cloud (Li et al., 2024). For Radio-RAG, that makes EACO-RAG a control skeleton for latency- and cost-aware retrieval/generation placement even though it does not model wireless PHY or MAC effects.

4. Empirical behavior and measured trade-offs

The radio-regulations system provides the clearest direct accuracy result for a telecom-specific Radio-RAG benchmark. On its multiple-choice evaluation, GPT-4o scores $59.0\% \pm 0.5$ without retrieval, $C_i$0 with naive full-document insertion, and $C_i$1 with the proposed structured RAG pipeline. The same paper reports a +23% absolute gain for DeepSeek-R1-Distill-Qwen-14B and a +3% absolute gain for DeepSeek-R1-Distill-Qwen-1.5B under RAG, while the optional reranker gives only about +1% absolute accuracy at approximately $C_i$2 end-to-end compute cost (Kassimi et al., 11 Sep 2025). These numbers make two points simultaneously: targeted grounding matters, and naive context stuffing does not.

For radio resource management, the central empirical claims come from an end-to-end O-RAN-compliant OAI testbed with two slices. In Scenario 2, with load pairs sampled from $C_i$3 Mbps and the experiment repeated 70 times, fixed allocations $C_i$4, $C_i$5, and $C_i$6 are compared against ReLLM. The fixed $C_i$7 split gives the lowest latency for the high-priority slice but produces packet drop ratios up to 50% for the low-priority slice. By contrast, the ReLLM-based controller keeps the latency of $C_i$8 under the 10 ms SLA target while maintaining a very small drop ratio for $C_i$9, reported as about 0.5% and elsewhere as “near-zero.” The paper also reports that the gated version significantly reduces token consumption over time relative to calling the resource allocation agent at every step (Salan et al., 28 Nov 2025).

For ORAN standards QA, the aggregate benchmark shows a more nuanced pattern. Across 600 selected ORAN-Bench-13K questions, GraphRAG and Hybrid GraphRAG both outperform vector-only retrieval on faithfulness, with scores 0.59, 0.59, and 0.55 for Graph, Hybrid, and Vector, respectively. Hybrid GraphRAG has the best factual correctness at 0.58 versus 0.50 for GraphRAG and 0.48 for Vector RAG, while GraphRAG has the best context relevance at 0.11 versus 0.10 for Vector and 0.04 for Hybrid. The paper also reports that Vector RAG performs best on easy questions for factual correctness, reaching 0.61, but declines on medium and hard questions, whereas graph-based methods are better on medium and hard cases requiring cross-document or relational reasoning (Ahmad et al., 4 Jul 2025).

At the edge/cloud systems level, EACO-RAG offers a different kind of trade-off. Relative to KGRAG-3B, it reduces cost by $F_1^{(i)} = \mathrm{ROUGE\mbox{-}L\mbox{-}F1}(R_i, C_i),$0 and delay by $F_1^{(i)} = \mathrm{ROUGE\mbox{-}L\mbox{-}F1}(R_i, C_i),$1 at only an $F_1^{(i)} = \mathrm{ROUGE\mbox{-}L\mbox{-}F1}(R_i, C_i),$2 sacrifice in accuracy; relative to RAG-3B, it improves accuracy by $F_1^{(i)} = \mathrm{ROUGE\mbox{-}L\mbox{-}F1}(R_i, C_i),$3 with comparable cost and delay. The abstract additionally reports cost reductions up to $F_1^{(i)} = \mathrm{ROUGE\mbox{-}L\mbox{-}F1}(R_i, C_i),$4 and $F_1^{(i)} = \mathrm{ROUGE\mbox{-}L\mbox{-}F1}(R_i, C_i),$5 under different delay constraints (Li et al., 2024). Those results are not radio-specific, but they quantify the sort of latency-cost-accuracy Pareto frontier that a radio-aware RAG deployment would have to manage.

5. Recurring design tensions

A central misconception is that more external context automatically yields better radio-domain behavior. The radio-regulations paper directly contradicts that view: full-document insertion for GPT-4o yields only a marginal gain of 0.6 percentage points, whereas structured retrieval yields an 11.9-point absolute improvement (Kassimi et al., 11 Sep 2025). The ORAN benchmark reaches a similar conclusion from another angle: Hybrid GraphRAG has the best factual correctness, but its context relevance is the lowest of the three pipelines because the dual retrieval stream can become verbose and redundant (Ahmad et al., 4 Jul 2025). ReLLM’s event-triggered design is motivated by exactly the same principle in control form: LLM reasoning should not be invoked at every interval, because unnecessary calls waste tokens and compute without improving SLA management (Salan et al., 28 Nov 2025).

A second misconception is that Radio-RAG is necessarily document-centric. The literature already spans at least three evidence models: regulation clauses segmented from PDFs, vector and graph contexts extracted from ORAN specifications, and nearest-neighbor retrieval over historical network states and allocation outcomes. This suggests that in radio-domain settings, “retrieval” is better understood as access to a task-specific non-parametric evidence store, which may be textual, structured, or operational.

A third misconception is that network awareness in edge/cloud RAG is already equivalent to radio awareness. EACO-RAG is explicit about the limits of that analogy. It uses delay-to-cloud, delay-to-edge, semantic locality, and query difficulty, but it does not model wireless PHY or MAC effects, spectrum scheduling, fading, handover, packet loss, or radio resource management. The paper therefore functions as a strong architectural blueprint for hierarchical deployment and collaborative routing, but only indirectly as a radio-specific design (Li et al., 2024).

6. Open problems and likely directions

Several unresolved problems appear consistently across the current Radio-RAG literature. In radio regulations, benchmark construction is still partially automated: questions are generated with google/flan-t5-xxl, filtered with AliMaatouk/Llama-3-8B-Tele, and then subjected only to a light human pass to remove obviously illogical questions. The authors explicitly note the lack of a formal expert-verification stage, so dataset quality and coverage remain open issues for any deployment that aspires to compliance-grade reliability (Kassimi et al., 11 Sep 2025).

In ORAN knowledge systems, evaluation is still largely answer-centric. The benchmarking paper calls for future work on latency and compute overhead, multimodal context integration, and deployment within orchestration frameworks such as SMO or RIC. It also omits several details that matter operationally, including a full graph schema, significance testing, detailed prompt disclosure, and cross-version generalization across evolving specifications (Ahmad et al., 4 Jul 2025). These gaps matter because ORAN assistants are intended for tasks such as xApp and rApp generation, root-cause analysis, and intent-driven management, where latency, traceability, and version control are first-order requirements.

In radio control, the main open issues are scalability and radio realism. ReLLM is demonstrated in a two-slice scenario on an O-RAN-compliant testbed, but the paper does not provide a broader scalability study over larger numbers of slices or users, and its retrieval mechanism is described only at a high level as nearest-neighbor access over historical examples. The first agent is mathematically grounded by the sigmoid risk score, but the paper does not specify a separate forecasting model beyond that formalism (Salan et al., 28 Nov 2025). Substantial additional modeling would be needed for larger heterogeneous traffic mixes, stronger prompt transparency, or deeper coordination between RIC layers.

At the systems-architecture level, EACO-RAG makes the missing radio variables explicit by omission. It treats delay as an input context variable, but it does not derive that delay from a wireless access model, expose bandwidth-sensitive transfer size, model contention among many users over the same bearer, or track mobility and handovers. The paper itself suggests that natural extensions would include channel quality indicator, uplink scheduling delay, backhaul utilization, handover probability, battery state, edge-server queue occupancy, energy consumption, radio resource usage, and service-level reliability constraints (Li et al., 2024). Taken together, these omissions indicate that a fully developed Radio-RAG remains an open systems problem at the intersection of retrieval, control, networking, and domain-grounded evaluation.