Quantum-RAG: Hybrid Retrieval for Low-Resource NLP

• Quantum-RAG is a hybrid retrieval framework that combines sparse BM25, dense FAISS, and a quantum-inspired amplitude similarity layer to enhance contextual relevance in low-resource NLP.
• It employs real-valued, L2-normalized embeddings with a Born-rule-style kernel to capture robust semantic overlaps, mitigating limitations of conventional retrieval techniques.
• Empirical results in Punjabi systems show improved perplexity, ROUGE-L, and cultural fidelity, demonstrating practical gains over traditional multilingual baselines.

Quantum-RAG denotes the incorporation of quantum or quantum-inspired mechanisms into retrieval-centered machine learning pipelines. In its principal usage in low-resource NLP, it refers to a hybrid retrieval system introduced alongside PunGPT2 and Pun-RAG for Punjabi, where sparse retrieval, dense retrieval, and a quantum-inspired semantic similarity layer are fused to improve factual grounding, contextual relevance, and cultural fidelity in generation (Singh et al., 3 Aug 2025). The label is not fully univocal, however: in high-energy physics, "Quantum-RAG" is also used as shorthand for "Quantum Rationale-Aware Graph Contrastive Learning," where "RAG" denotes rationale-aware graph learning rather than retrieval-augmented generation (Jahin et al., 2024). Across these usages, the term most often indicates quantum-inspired modeling on classical hardware rather than deployment on a full QRAM-backed quantum computing stack (Phalak et al., 2023).

1. Terminology and scope

The Punjabi-language formulation of Quantum-RAG emerged from a specific low-resource retrieval problem. Punjabi, despite approximately 100M speakers, has historically been underserved in NLP, and multilingual models such as mBERT and MuRIL were described as underperforming because of tokenization inefficiencies and vocabulary dilution. Within that setting, the quantum component is not a hardware quantum memory or a fault-tolerant quantum circuit; it is an amplitude-based similarity layer operating on normalized real-valued vectors and combined with BM25 and FAISS retrieval (Singh et al., 3 Aug 2025).

A recurring source of confusion is that the same acronym has been reused in a different domain. In jet discrimination, "Quantum-RAG" refers to a graph-contrastive learning framework with a quantum rationale generator, not to retrieval-augmented generation. The two usages share a reliance on quantum-inspired or variational quantum ideas, but they address different tasks, data structures, and evaluation protocols (Jahin et al., 2024).

Usage of "Quantum-RAG"	Domain	Core mechanism
Quantum-RAG in Pun-RAG	Low-resource Punjabi NLP	BM25 + dense FAISS + quantum kernel similarity
Quantum Rationale-Aware Graph Contrastive Learning	Jet discrimination	Quantum rationale generator for graph augmentations

In the Punjabi setting, a further misconception is that "quantum" implies complex-valued state evolution or specialized quantum hardware. The formulation provided for Quantum-RAG explicitly uses amplitude-only, real-valued normalization, emphasizes minimal memory overhead, and states that the system runs on classical hardware. This makes it a quantum-inspired retrieval method rather than a QRAM implementation (Singh et al., 3 Aug 2025).

2. Punjabi Quantum-RAG as a retrieval architecture

Quantum-RAG was introduced inside a broader Punjabi LLM suite comprising PunGPT2, Pun-RAG, and Pun-Instruct. PunGPT2 is described as the first fully open-source suite of Punjabi LLMs, trained from scratch on a 35GB domain-diverse corpus; the detailed experimental description specifies a 35.5GB Punjabi corpus spanning news, folk literature, social media comments, religious texts, and archives, totaling 4.8M documents. Retrieval is performed over a curated Punjabi knowledge base drawn from that corpus, with tokenization based on a Punjabi BPE tokenizer optimized on the corpus and reported as using 45M subword units (Singh et al., 3 Aug 2025).

The motivation for the hybrid design is the complementarity of sparse and dense retrieval failure modes in low-resource corpora. Pure BM25 retrieval is lexically precise but brittle under morphological variation, synonymy, and orthographic or register variation, including Gurmukhi/Shahmukhi and formal versus colloquial usage. Pure dense retrieval improves semantic matching but may overfit to surface distributions in low-resource settings and can increase memory footprint when large vector stores are used. Quantum-RAG was proposed to insert a third scoring signal that captures contextual overlap more robustly for subtle, culture-bound queries (Singh et al., 3 Aug 2025).

Within the Pun-RAG framework, the pipeline is specified as follows:

1. Query preprocessing and encoding: the query is tokenized with the Punjabi BPE tokenizer; BM25 term weights are derived; a dense query vector is produced for FAISS retrieval; and an amplitude-based embedding is derived for quantum-inspired scoring.
2. Candidate retrieval: BM25 returns top $k_s$ sparse candidates with scores $S_{\text{BM25}}(d)$, and FAISS returns top $k_d$ dense candidates with scores $S_{\text{dense}}(d)$.
3. Candidate pooling: the candidate set $C$ is formed as the union of the sparse and dense results.
4. Quantum kernel scoring: amplitude-based embeddings are computed for the query and for each $d \in C$, and a quantum kernel similarity is evaluated.
5. Score normalization and fusion: each score type is normalized per query across $C$ to the same range, and a weighted late-fusion score is computed.
6. Reranking and generation: the candidate pool is reranked by the fused score, top passages are selected, and the query plus retrieved Punjabi passages are concatenated into the PunGPT2 or Pun-Instruct input within the approximately 1024-token context window used in pretraining.

Pun-Instruct is a parameter-efficient instruction-tuned variant using QLoRA, and the same retrieval pipeline can be used for instruction-following tasks. The reported intent of the retrieval stage is to improve factuality and reduce hallucinations while preserving cultural fidelity in grounded Punjabi generation (Singh et al., 3 Aug 2025).

3. Amplitude embeddings, quantum kernel similarity, and late fusion

The mathematical core of the Punjabi Quantum-RAG formulation is an amplitude-based feature map over real-valued embeddings. For a text $x$, let $\phi(x) \in \mathbb{R}^m$ be a real-valued embedding, such as a pooled passage vector. The normalized amplitude state is defined as

$\psi(x) = \frac{\phi(x)}{\|\phi(x)\|_2},$

so that $S_{\text{BM25}}(d)$0 (Singh et al., 3 Aug 2025).

The quantum kernel is then the squared magnitude of the inner product,

$S_{\text{BM25}}(d)$1

The paper also gives the equivalent pure-state density-matrix form, with $S_{\text{BM25}}(d)$2 and

$S_{\text{BM25}}(d)$3

This is the "quantum-inspired" step: it does not introduce complex phases, but it interprets normalized amplitudes through a Born-rule-style overlap score. The accompanying explanation states that the squared inner product emphasizes high-confidence overlaps and can act like an "interference-aware" similarity signal, boosting passages that share multiple aligned semantic components with the query while moderating superficially similar but misaligned passages (Singh et al., 3 Aug 2025).

Classical retrieval scores are fused by query-wise normalization and weighted late fusion:

$S_{\text{BM25}}(d)$4

$S_{\text{BM25}}(d)$5

and

$S_{\text{BM25}}(d)$6

with $S_{\text{BM25}}(d)$7 and $S_{\text{BM25}}(d)$8. The weights are tuned on validation data, but no numeric values are reported (Singh et al., 3 Aug 2025).

The intended division of labor among the terms is explicit: the $S_{\text{BM25}}(d)$9 term provides lexical precision via BM25, the $k_d$0 term provides semantic matching via dense retrieval, and the $k_d$1 term provides quantum-inspired amplitude alignment to capture richer contextual overlap and to attenuate brittle cases where sparse or dense retrieval alone mis-ranks candidates. Because the amplitude vectors are simple $k_d$2-normalized passage-level vectors, the paper emphasizes that the quantum component does not increase vector dimensionality and does not require complex-valued storage (Singh et al., 3 Aug 2025).

4. Implementation characteristics and empirical results

The implementation combines a dense FAISS retriever, a standard BM25 inverted index, and quantum-inspired embeddings derived by $k_d$3-normalizing passage-level vectors. FAISS indexes the curated Punjabi knowledge base and supports approximate nearest-neighbor search; BM25 supplies sparse lexical matching. The paper does not specify the FAISS index type, quantization settings, BM25 $k_d$4 parameters, or the encoder used to produce the underlying $k_d$5 vectors. Quantum scoring is applied only to the candidate pool formed by the union of BM25 and FAISS top-$k_d$6 results, which is presented as the reason the additional compute remains modest (Singh et al., 3 Aug 2025).

The reported backbone and training setup are also concrete. The PunGPT2 backbone has 124M parameters and was trained on a single A100 GPU (MIG 3g.20gb) over 48 hours. The training stack uses Hugging Face Transformers, Accelerate, Datasets, and PEFT for QLoRA. The suite is presented as fully open-source, although the provided text does not include URLs (Singh et al., 3 Aug 2025).

The principal intrinsic and downstream metrics are perplexity, training loss, ROUGE-L, and human-rated cultural fidelity.

System	Perplexity	ROUGE-L
mBERT	45.2	28.7
MuRIL	42.1	30.9
PunGPT	2.24	37.4
Pun-RAG	2.10	38.5
Pun-Instruct	2.15	39.2
Quantum-RAG	2.05	40.1

System	Training loss	Cultural fidelity
mBERT	3.92	3.4
MuRIL	3.85	3.7
PunGPT	0.85	4.4
Pun-RAG	0.80	4.6
Pun-Instruct	0.82	4.7
Quantum-RAG	0.78	4.8

These results place Quantum-RAG ahead of the other reported Punjabi-system variants on all four reported metrics. The abstract additionally states that the models significantly outperform strong multilingual baselines including mBERT, mT5, and MuRIL in perplexity, factuality, and fluency. An ablation is also reported: removing the quantum-inspired retrieval component reduces human preferences for fluency and accuracy, although no numeric deltas are provided (Singh et al., 3 Aug 2025).

Several practical details remain unspecified. Latency and throughput are not quantified; retrieval-focused metrics such as nDCG, MRR, and recall@k are not reported; and the precise encoder behind the amplitude vectors is omitted. Those omissions matter methodologically because the quality of $k_d$7 constrains the quality of $k_d$8 and therefore the utility of the kernel score. The available evidence nevertheless indicates that, within this Punjabi evaluation setup, the additional kernel term improves contextual relevance and factual grounding without substantial memory expansion (Singh et al., 3 Aug 2025).

5. Relation to QRAM and quantum memory architectures

Punjabi Quantum-RAG is a quantum-inspired retrieval method implemented on classical hardware, but it naturally invites comparison with QRAM-centered notions of quantum retrieval. QRAM is defined as a memory that maps quantum addresses to quantum or classical data without collapsing superpositions. With $k_d$9 address qubits indexing $S_{\text{dense}}(d)$0 memory cells, a standard QRAM memory operation is written as

$S_{\text{dense}}(d)$1

or, in the survey notation for classical data $S_{\text{dense}}(d)$2 stored at address $S_{\text{dense}}(d)$3,

$S_{\text{dense}}(d)$4

This superposition access is what would enable genuinely quantum retrieval or similarity oracles in a future Quantum-RAG pipeline (Phalak et al., 2023).

The survey distinguishes several QRAM architectures with different operational trade-offs. Bucket-brigade QRAM uses a binary-tree routing mechanism with qutrit switches and activates only $S_{\text{dense}}(d)$5 switches for a single-address query, although its worst-case or full-superposition circuit width and depth are reported as $S_{\text{dense}}(d)$6. Fanout QRAM distributes address control globally and activates $S_{\text{dense}}(d)$7 switches even for a single query, making it comparatively noise-prone. FF-QRAM provides an explicit circuit method for loading classical data into amplitudes, but with circuit depth $S_{\text{dense}}(d)$8; the rotation angle for a normalized datum $S_{\text{dense}}(d)$9 is $C$0. PQC-based variants such as EQGAN reduce depth by storing data approximately, and qudits-based memory compresses multiple qubits into fewer higher-dimensional elements (Phalak et al., 2023).

The survey’s relevance to Quantum-RAG is prospective rather than direct. It frames a possible retrieval layer in which embeddings or key-value pairs are stored in QRAM, a query state $C$1 is prepared, a similarity oracle is constructed using QRAM-supplied data, and amplitude amplification or estimation is used to boost high-similarity items before readout and post-selection. The total time for such a pipeline is decomposed schematically as

$C$2

with loading often identified as the dominant bottleneck for data-heavy applications (Phalak et al., 2023).

This comparison clarifies a central point: the Punjabi system does not bear the loading-depth, coherence, routing-loss, and fault-tolerance burdens analyzed in the QRAM survey. A plausible implication is that its practical gains derive from a lightweight kernelized scoring layer rather than from any quantum speedup. Conversely, the survey suggests that a future fully quantum retrieval stack for RAG would face significant constraints from error accumulation, address-width scaling, and data-loading cost, especially on NISQ hardware (Phalak et al., 2023).

6. Alternative usage: quantum rationale-aware graph learning

In a distinct literature, "Quantum-RAG" denotes "Quantum Rationale-Aware Graph Contrastive Learning" for quark-versus-gluon jet tagging. Here the task is not retrieval-augmented generation but graph representation learning on particle-level collider data. The framework introduces a quantum rationale generator that produces node-level salience probabilities, constructs rationale-preserving and rationale-perturbing views of a jet graph, and trains a ParticleNet-based encoder with a multi-term contrastive loss (Jahin et al., 2024).

The model operates on a filtered version of the Pythia8 Quark/Gluon Jets for Energy Flow dataset. Jets are selected with $C$3 in $C$4 GeV and $C$5, balanced between quark and gluon labels, then truncated to the top 7 particles by $C$6, yielding $C$7 total examples with 10,000 train, 1,250 validation, and 1,250 test examples. Each jet is represented as a graph with node features $C$8 and edge attributes $C$9 (Jahin et al., 2024).

The quantum rationale generator uses 7 qubits, one per retained particle. It initializes with Hadamard gates, encodes node features through $d \in C$0 rotations and edge relations through controlled-$d \in C$1 gates, and applies $d \in C$2 entangling blocks with trainable single-qubit rotations. The implemented variational circuit has 45 trainable parameters. Node-level rationale probabilities are defined through Hamming-weight-1 measurement outcomes,

$d \in C$3

and are normalized to form a discrete distribution over nodes. These probabilities determine masks or weights for rationale-aware augmentations and contrastive learning (Jahin et al., 2024).

The reported best result is a test AUC of $d \in C$4, with accuracy $d \in C$5 and F1 $d \in C$6. The paper further states that QRGCL uses only 45 QRG parameters, whereas the classical rationale generator baseline has approximately 1,073 parameters, and that QRGCL outperforms classical, quantum, and hybrid GCL and GNN benchmarks in this setting (Jahin et al., 2024).

The importance of this alternative usage is terminological. In this literature, "RAG" means rationale-aware graph learning, and the method is explicitly distinguished from retrieval-augmented generation. For readers encountering "Quantum-RAG" across disciplines, the ambiguity is substantive rather than stylistic (Jahin et al., 2024).

7. Limitations and open directions

The Punjabi Quantum-RAG formulation has several explicit limitations. The paper does not specify the encoder used to compute the passage-level vectors $d \in C$7, so the source of the amplitude representation remains underdetermined. The design is amplitude-only and omits complex phases, which the authors identify as a possible avenue for richer interference modeling. BM25 remains sensitive to orthographic noise and code-mixing, dense and quantum-inspired modules cannot fully compensate for poorly curated passages, and retrieval-specific metrics such as nDCG, MRR, and recall@k are absent. The framework is motivated for Punjabi but is stated to require validation across other low-resource languages with different scripts and morphology (Singh et al., 3 Aug 2025).

The future directions named for the Punjabi system are correspondingly concrete: alternative quantum kernels such as density-matrix mixtures and phase-augmented embeddings, end-to-end training of the quantum kernel with retrieval supervision, and integration with other retrievers such as HNSW-only and ColBERT-style token-level matching. These suggestions indicate that the present formulation should be understood as a first practical integration of quantum representations in low-resource language generation rather than a finalized retrieval architecture (Singh et al., 3 Aug 2025).

The QRAM survey broadens the frontier but also sharpens the constraints. It highlights error-resilient routing, efficient state preparation, fault-tolerant QRAM resource estimation, and hybrid architectures with classical pre-filtering plus QRAM-assisted subroutines as open problems. For near-term quantum retrieval, the survey judges hybrid designs operating on tens to hundreds of items after classical pre-filtering to be the most viable. This suggests that fully quantum versions of retrieval-augmented generation remain contingent on advances in memory loading, coherence preservation, and oracle design, whereas current "Quantum-RAG" systems derive their practicality from quantum-inspired scoring and hybridization rather than from end-to-end quantum computation (Phalak et al., 2023).