Grahak-Nyay: Advancing Consumer Justice in India

• Grahak-Nyay is a dual concept that integrates a fast, transparent consumer grievance redressal mechanism with a domain-specific legal chatbot under Indian law.
• It employs retrieval-augmented generation and expert-validated datasets to simulate legal reasoning, streamline case triage, and optimize judicial workflows.
• Designed as a decision-support tool, Grahak-Nyay enhances timeliness and fairness for millions of pending consumer disputes while addressing procedural complexities.

Grahak-Nyay denotes an emerging technical and legal conception of consumer justice in India centered on faster, more intelligible, and more reliable grievance redressal for ordinary consumers. In the recent literature, the term appears in two closely related senses: as a normative theme of “fast and fair justice delivery to consumers” that can be operationalized through case prioritization, transparent scheduling, and notification workflows; and as the title of a domain-specialized legal chatbot that uses open-source LLMs and retrieval-augmented generation to guide consumers through rights, remedies, and filing pathways under Indian consumer law (Varma et al., 16 Oct 2025, Ganatra et al., 7 Jul 2025). Around this core, a broader research cluster links Grahak-Nyay to complaint triage, unfair-terms detection, precedent-grounded adjudication, multi-agent judicial simulation, and graph-based outcome prediction.

1. Concept and institutional setting

The Grahak-Nyay research agenda is motivated by a structural access-to-justice problem. One strand emphasizes that India’s judicial system carries about 52 million pending cases across all courts, with delays affecting criminal, civil, family, and commercial disputes and eroding trust among “ordinary consumers” whose claims may be small in monetary value yet critical to livelihoods, such as refunds, defective goods, or denied insurance claims (Varma et al., 16 Oct 2025). Another strand focuses specifically on consumer grievance redressal and identifies procedural complexity, legal jargon, and jurisdictional challenges as major barriers even where legal representation is not mandatory and institutional mechanisms such as the National Consumer Helpline (NCH) and e-Daakhil already exist (Ganatra et al., 7 Jul 2025).

The consumer-law literature describes a two-level grievance architecture. At Level 1, a consumer is expected to first use the service provider’s internal grievance mechanism. At Level 2, unresolved matters may proceed to government-backed or quasi-judicial channels, including the NCH, sectoral ombudsmen or regulators, and the Consumer Disputes Redressal Commissions through e-Daakhil (Ganatra et al., 7 Jul 2025). The scale of demand is explicit: the NCH handles, on average, about 107,966 complaints per month in FY 2023–2024 (Ganatra et al., 7 Jul 2025).

Within this setting, Grahak-Nyay is not merely a slogan for consumer protection. It is framed as an attempt to improve timeliness, fairness, transparency, and scalability in grievance handling and adjudication. A recurring design objective is to prevent “ordinary” cases from being indefinitely sidelined while also preserving judicial control and procedural intelligibility (Varma et al., 16 Oct 2025).

2. Grahak-Nyay as a domain-specialized legal assistant

As a concrete system, “Grahak-Nyay: Consumer Grievance Redressal through LLMs” implements a legal-domain chatbot tailored to Indian consumer law. Its high-level pipeline is a conversational RAG architecture: user query and chat history are first sent to a query rewriting LLM; the rewritten query is embedded and used to retrieve legal context from a curated knowledge base; and an answer generation LLM then produces the response using the retrieved context plus the conversation history (Ganatra et al., 7 Jul 2025).

Both the query rewriter and the answer generator use Llama-3.1-8B-Instruct. Retrieval uses the embedding model mixedbread-ai/mxbai-embed-large-v1, with top-$k$ retrieval fixed at $k=4$ and document ranking by cosine similarity,

$$s(q, d_i) = \cos\bigl( \text{emb}(q), \text{emb}(d_i) \bigr).$$

The answer-generation model is served via Text Generation Inference (TGI) on a single A100 40GB GPU (Ganatra et al., 7 Jul 2025).

A notable design decision concerns chunking. Fixed-length chunking was found problematic because long answers were split across chunks and multiple short answers could be merged into noisy retrieval units. The final system therefore represents the knowledge base as Question-Answer–style documents, with each chunk equal to one QA pair (Ganatra et al., 7 Jul 2025). This makes the retrieval unit jurisprudentially narrow and semantically coherent.

The system prompt imposes strong domain and safety constraints. The chatbot is instructed to handle only consumer-law-related issues across sectors such as banking, telecom, insurance, e-commerce, real estate, education, automobiles, airlines, and food safety; to ask one question at a time; to rely only on retrieved context for factual details such as phone numbers and organizational information; and to reject out-of-domain requests with the explicit response: “I can't answer that. I can help you with consumer-related issues.” It is also instructed to guide users through grievance narration, limitation-period checking, opposing-party identification, desired relief, legal notice drafting, and recourse to the National Consumer Helpline (1915, formerly 14404) and UMANG (Ganatra et al., 7 Jul 2025).

This architecture positions Grahak-Nyay as a pre-filing and parallel-support instrument rather than a substitute for formal adjudication. Its role is to compress approximately 1,500+ pages of legal materials into concise, consumer-facing guidance while keeping the legal substrate editable through the knowledge base rather than frozen into model weights (Ganatra et al., 7 Jul 2025).

3. Datasets, judgment integration, and HAB evaluation

The Grahak-Nyay chatbot is supported by a set of new domain resources. GeneralQA provides foundational consumer-law coverage under the Consumer Protection Act, 2019. The main text reports 53 QA pairs, while the appendix reports 52 QA pairs (Ganatra et al., 7 Jul 2025). SectoralQA extends the coverage to sector-specific grievances; the main paper reports 23 sectors and 889 QA pairs, while the appendix describes 27 sector documents yielding about 835 QA pairs (Ganatra et al., 7 Jul 2025). These sectoral entries cover domains such as banking, insurance, telecom, e-commerce, medical negligence, real estate, railways, and airlines.

A third corpus, SyntheticQA, was created specifically for RAG evaluation. It contains 4,734 questions generated from existing FAQs and answers using Llama-3.1-70B-Instruct, with consistency-oriented answer generation then performed by Llama-3.1-8B-Instruct (Ganatra et al., 7 Jul 2025). Its purpose is to stress-test retrieval and grounding where general and sector-specific material may both appear relevant.

For dialogue evaluation, the paper introduces NyayChat, a dataset of 303 conversations—the abstract says 300 annotated chatbot conversations—with an average of 32.01 turns and about 3,475 tokens per conversation (Ganatra et al., 7 Jul 2025). These dialogues were constructed by law students from a database of 1,200 District Consumer Disputes Redressal Commission judgments and then reviewed by at least two legal experts. The system also incorporates a Judgments corpus of 570 judgments from Indian Consumer Courts across 23 sectors, manually summarized and categorized to support citation, retrieval, and user trust (Ganatra et al., 7 Jul 2025).

Evaluation is organized around both standard generation metrics and a domain-specific rubric called HAB: Helpfulness, Accuracy, Brevity. Each dimension is scored on a 5-point Likert scale. Helpfulness measures whether the chatbot gives actionable and relevant guidance; Accuracy measures legal and factual correctness, including phone numbers and procedural details; Brevity measures conciseness and avoidance of unnecessary questioning (Ganatra et al., 7 Jul 2025). The paper does not define a single aggregate formula, but the three-dimensional structure is intended to replace purely lexical overlap as the central criterion for legal chatbot quality.

The reported results are strong but differentiated. On a 65-conversation subset with expert reference responses, Grahak-Nyay achieved ROUGE-1 66.9, ROUGE-2 41.1, ROUGE-L 33.2, BERTScore 90.9, METEOR 41.9, and BLEU 37.4 (Ganatra et al., 7 Jul 2025). On LLM-based HAB evaluation using gpt-4o-mini as the evaluator, the mean scores were 4.65 for Helpfulness, 3.61 for Accuracy, and 3.12 for Brevity (Ganatra et al., 7 Jul 2025). For RAG quality, Faithfulness scores were approximately 8.58 on SectoralQA, 9.02 on GeneralQA, and 9.30 on SyntheticQA, with Answer Relevance around 7.35–7.48 (Ganatra et al., 7 Jul 2025). Human evaluation further found that Grahak-Nyay outperformed ChatGPT-4.0, Claude-3.5, Llama-3.1-405B-128k, and Llama-3.1-8B-128k in Helpfulness and Brevity, while performing similarly to larger models in Accuracy and better than the bare Llama-3.1-8B base (Ganatra et al., 7 Jul 2025).

4. Case management, prioritization, and consumer-court extension

A second major line of Grahak-Nyay research addresses not legal advice but judicial workflow. “A Comprehensive Framework for Efficient Court Case Management and Prioritization” proposes a cloud-based court case management and prioritization framework that is described and simulated mainly for Criminal, Civil, and Family cases but is explicitly designed to be generic and extendable to consumer courts, consumer dispute redressal commissions/forums, and consumer disputes in e-commerce, telecom, insurance, and banking (Varma et al., 16 Oct 2025).

The framework has a Data Entry Layer, a Core Processing / Scheduling Engine, dedicated databases for pending, disposed, and configuration data, a Notification System, and a Judicial Interface. Case metadata include case ID, case type, filing date, severity, input priority level, and legal sections; data are stored in a MySQL database in the cloud (Varma et al., 16 Oct 2025). The scheduling engine computes case age from filing date and previous hearing dates, classifies cases by type, and assigns a priority weight through a regression-based ML model with expert-validated section-specific weights.

The priority computation is summarized as

$$w_i = f\big( \text{Age}_i,\, \text{Severity}_i,\, \text{PriorityLevel}_i,\, \text{Sections}_i,\, \text{CaseType}_i,\, \text{HearingsCount}_i \big),$$

with cases then sorted in descending order of $w_i$ (Varma et al., 16 Oct 2025). Scheduling respects a fixed daily capacity—for example, 100 cases/day, split 50 fresh and 50 old—while accounting for holidays and judge leave. A pendency-based uplift periodically increases the weight of low-priority cases that have remained pending beyond a threshold, so that they are not postponed indefinitely (Varma et al., 16 Oct 2025).

For Grahak-Nyay specifically, the paper states that the architecture can be instantiated as a Consumer Court Case Management System by adding “Consumer” as a case category, encoding relevant consumer laws and time limits, and adjusting the priority model to handle consumer-specific parameters such as monetary value, perishable goods, healthcare negligence, or vulnerability-related factors (Varma et al., 16 Oct 2025). The system also includes SMS/email notifications to judges, advocates, and litigants, event triggers for hearing assignment and disposal, and dashboards showing prioritized daily cause lists.

The evaluation uses 10,000 synthetic cases and reports that the system can efficiently process large volumes, that notification delivery is reliable, and that judges and registrars give positive usability feedback (Varma et al., 16 Oct 2025). The regression-based ML model is reported to achieve F1 score 99.7%, Precision 99.8%, and Recall 99.4% on simulated data (Varma et al., 16 Oct 2025). Since these numbers are simulation-based and the consumer-court component is an explicit extension rather than the primary evaluated domain, a careful reading is that the paper supplies an administrative substrate for Grahak-Nyay rather than a validated consumer-court deployment.

5. Deliberative, precedent-aware, and appellate AI

A third research strand models Grahak-Nyay as an explainable decision-support or simulation problem. SAMVAD is a “sandbox court”: a multi-agent, LLM-driven simulation of Indian judicial deliberation grounded in a RAG knowledge base of the Indian Penal Code, the Constitution of India, and the Code of Criminal Procedure (Devadiga et al., 4 Sep 2025). Its architecture includes a Judge agent, Prosecution Counsel, Defense Counsel, multiple Adjudicators, and an Orchestrator. The paper states that the architecture is domain-agnostic and directly reusable for consumer justice by replacing the knowledge base with the Consumer Protection Act, sectoral regulations, and consumer-forum case law (Devadiga et al., 4 Sep 2025).

Consensus in SAMVAD is procedural. If $N$ adjudicators produce leanings in $$\mathcal{L} = \{\text{Guilty}, \text{Not Guilty}, \text{Undecided}\}$$, the stopping condition is based on

$$\text{AgreementRatio} = \frac{\max_{\ell \in \mathcal{L}} \text{count}(\ell)}{N},$$

with a threshold qualitatively described as $>80\%$ (Devadiga et al., 4 Sep 2025). In reported experiments, five sample case runs with five adjudicators achieved final agreement ratios of 0.90–1.00, participation of 1.00, and average argument grounding scores of 0.30–0.45. Ablation studies show that RAG increases agreement ratios, roughly doubles grounding scores in one example from 0.21 to 0.42 with Qwen-2.5-7B, increases meaningful statements, and improves verdict consistency (Devadiga et al., 4 Sep 2025).

Where SAMVAD remains a legal deliberation simulator, CyberJurors moves directly into platform consumer disputes. It introduces the E-commerce Dispute Verdicts (EDV) task and VerdictBench, a multimodal benchmark of 6,000 real-world cases derived from a platform in which 17 randomly assigned volunteer jurors decide transaction disputes by majority vote (Sun et al., 27 May 2026). Each case preserves transaction metadata, chat history, text evidence, images, and videos; overall verdicts show buyers win 37.4% of cases and sellers win 62.6% (Sun et al., 27 May 2026).

CyberJurors combines Individual Verdict Chain-of-Thought with Jury Consensus Verdict and a Precedent Base. At the collective level, consensus is computed by

$c_t = \frac{1}{17}\max(\hat{y}_t^b,\hat{y}_t^s),$

with early stopping when $k=4$0, where $k=4$1 (Sun et al., 27 May 2026). On VerdictBench, CyberJurors reports Accuracy 0.7292, Weighted F1 0.7258, Macro F1 0.7037, Recall 0.6999, Precision 0.7100, MAE 4.73, and RMSE 6.37, outperforming both single-model baselines and court-simulator baselines (Sun et al., 27 May 2026). The paper emphasizes that generic multimodal models struggle with long, noisy evidence and that structured clue selection, deliberation, and precedents are critical.

A related but distinct approach appears in “Reviewing the Reviewer: Graph-Enhanced LLMs for E-commerce Appeal Adjudication”, which models hierarchical appellate correction through the Evidence-Action-Factor-Decision (EAFD) schema (Du et al., 27 Feb 2026). Cases are represented as typed graphs whose reasoning paths take the form

$k=4$2

The central innovation is the Request More Information (RMI) outcome, returned when critical verification actions remain unexecuted rather than allowing unconstrained text generation to hallucinate missing checks (Du et al., 27 Feb 2026). The quantitative gains are large: LLM (Query-Only) attains 0.708 accuracy, LLM + RMI reaches 0.875, and the full graph-enhanced EAFD system reaches 0.958 offline accuracy with Macro-F1 0.867; in production it achieves 96.3% alignment, with 100% precision on approval in online deployment (Du et al., 27 Feb 2026). For Grahak-Nyay, the relevance lies in appeals, documentary conflicts, and procedurally grounded requests for missing evidence.

6. Adjacent computational components: triage, contract scrutiny, and outcome prediction

Not all Grahak-Nyay work is explicitly framed as courts or chatbots. Upstream complaint handling is addressed by RE-GrievanceAssist, an end-to-end pipeline for real-estate complaint management in which complaint volumes exceed 1000 per day and tickets are aggregated via Freshdesk (C et al., 2024). The system uses a response/no-response classifier with TF-IDF + XGBoost, a user type classifier with fastText, and an issue/sub-issue hierarchy with TF-IDF + XGBoost, deployed as a Databricks batch job every 20 minutes (C et al., 2024). Reported test-set F1 scores are 86.39% for response/no-response, 90% for user type, 72.95% for issue, and 62.23% for sub-issue, with a 40% reduction in overall manual effort and monthly cost reduction of Rs 1,50,000 since August 2023 (C et al., 2024). In Grahak-Nyay terms, this is a triage and routing layer that precedes formal redressal.

Contract formation and ex ante unfairness are addressed by “Unfair TOS: An Automated Approach using Customized BERT”, which treats terms of service as a consumer-justice problem because users often accept dense contractual terms without reading them (Akash et al., 2024). Using the CLAUDETTE dataset of 9,414 statements annotated across eight unfairness categories plus a binary unfairness label, the paper combines Legal-BERT, a Fine-tune-base strategy, SMOTE, and SVC (Akash et al., 2024). It reports macro F1 = 0.922 for unfair clause detection and strong per-category results, including 1.0 F1 for arbitration, choice of law, and jurisdiction in the reported setup (Akash et al., 2024). This strand pushes Grahak-Nyay upstream from dispute resolution to automated detection of contractual asymmetry.

At the level of precedent-sensitive outcome modeling, “Exploring Graph Neural Networks for Indian Legal Judgment Prediction” reframes legal judgment prediction as binary node classification over a heterogeneous legal graph of cases, acts, and time nodes (Khatri et al., 2023). Its best-performing model, using XLNet pre-trained embeddings, achieves macro F1 score of 75% for the judgment prediction task, while link prediction achieves ROC of more than 80% (Khatri et al., 2023). The paper also performs fairness-related redaction experiments on names and gendered pronouns and reports “almost the same output” after redaction (Khatri et al., 2023). The text explicitly notes that, if a corpus of consumer disputes were available, the same GNN approach could underpin a Grahak-Nyay system that predicts outcomes and suggests relevant past consumer decisions (Khatri et al., 2023).

7. Limits, misconceptions, and research trajectory

Across the literature, a common misconception is that Grahak-Nyay research seeks autonomous replacement of lawyers, judges, or commissions. The papers instead repeatedly frame their systems as decision-support, simulation, workflow, or assistance tools. The Grahak-Nyay chatbot is intended to lower barriers for self-representation, not to replace lawyers or official bodies; SAMVAD is explicitly positioned as a simulation and research tool; CyberJurors is described as a decision-support tool rather than an unregulated replacement for human oversight; and the case-management framework emphasizes that judges retain final control and that the system is decision-support, not a decision-maker (Ganatra et al., 7 Jul 2025, Devadiga et al., 4 Sep 2025, Sun et al., 27 May 2026, Varma et al., 16 Oct 2025).

The limitations are also recurrent. For the chatbot, the paper acknowledges residual hallucinations, jurisdictional limitations, temporal drift, and the risk of user over-reliance (Ganatra et al., 7 Jul 2025). For case-management systems, challenges include data quality, the digital divide, institutional resistance, algorithmic bias, and privacy concerns (Varma et al., 16 Oct 2025). For SAMVAD, the stated limitations include no explicit doctrinal correctness evaluation, a simplified deliberation model, and limited RAG coverage (Devadiga et al., 4 Sep 2025). For CyberJurors, the authors point to bias, limited persona realism, unresolved jury size–cost trade-offs, and strong platform-specificity (Sun et al., 27 May 2026). For EAFD-based appeal adjudication, limitations include dependence on documentation richness, a domain- and policy-specific factor ontology, and incomplete visibility into non-textual or internal tools (Du et al., 27 Feb 2026).

Future work follows two broad directions. One direction deepens automation within formal institutions: the case-management paper proposes Deep Learning / Transformers for Legal Documents, Reinforcement Learning (RL) for Scheduling, and Predictive Analytics for time-to-disposal and resource allocation (Varma et al., 16 Oct 2025). The other direction expands consumer-facing and platform-facing AI: the Grahak-Nyay chatbot suggests fine-tuning, more advanced retrieval, and controlled user studies; CyberJurors identifies better multimodal evidence perception and compromise or mediated outcomes as next steps; and EAFD-based systems suggest active learning via RMI and policy-aware schema adaptation (Ganatra et al., 7 Jul 2025, Sun et al., 27 May 2026, Du et al., 27 Feb 2026).

Taken together, these works indicate that Grahak-Nyay is best understood as a layered socio-technical program for consumer justice in India: intake assistance and legal explanation at the front end, triage and prioritization in administrative pipelines, precedent-aware and evidence-grounded reasoning in adjudication, and auditable requests for missing information where automation reaches its epistemic boundary.