Gen AI-based Legal Research Assistant using Multi-Agent Collaboration, Retrieval Augmented Generation, and Chain of Thought Prompting

Abstract

The present invention is related to an AI-driven legal research assistant system (100) which optimizes legal research and analysis using large language models equipped with multi-agent collaboration and hierarchical reasoning capabilities. It leverages retrieval augmented generation for accurate information retrieval, with a dynamic knowledge repository. Real-time coordination is managed by specialized agents, while chain of thought prompting ensures transparent reasoning. The system comprises multiple integrated components including multi-agent communication protocol (130), fine-tuned model (112), retrieval augmented generation pipeline (140), hierarchical chain of thought mechanism (110), and Agents: Citation agent (137), Validation agent (136), QA agent (135), Drafting agent (131),Neutralizer agent (132) and Analyzer agent (133).

Specification

Description:FIELD OF THE INVENTION:
The current invention relates to the legal field particularly, a generative AI system to help in legal research for legal professionals.
BACKGROUND OF THE INVENTION:
Traditional legal research tools primarily rely on basic keyword search functions to help professionals find relevant cases and statutes. These tools, while useful for locating specific terms, lack the deeper, conceptual understanding required for handling complex legal queries. Older systems do not adapt or improve based on user input, resulting in static search capabilities that miss alternative interpretations, nuanced perspectives, or legal arguments. Without adaptive learning features, these systems provide limited functionality, often producing search results that fail to capture the full context of legal inquiries or help users derive insights from related cases.
AI technologies and machine learning advancements have introduced opportunities to enhance traditional legal research by incorporating NLP and predictive analytics. Large language models (LLMs) and generative AI represent a promising shift, with the potential to automate routine research tasks, understand the context of legal texts, recognize patterns across case law, and even suggest likely case outcomes based on historical data. Unlike static systems, LLMs can continually learn from data, improving their relevance and accuracy over time. These advancements could transform legal research by delivering insights beyond mere keyword matches and providing more adaptive, personalised support to legal professionals.
Research has explored various aspects of the challenges and solutions in legal language processing. The syntactic complexity and specialized terminology used in legal texts, emphasizing the need for advanced NLP techniques to help professionals understand dense, jargon-heavy language [1]. Investigation on the information retrieval challenges legal professionals face, highlighting how essential information is often buried within lengthy documents, complicating the search for relevant case precedents and legal principles [2]. Studies on tools like LexRank and Bi-LSTM models have demonstrated some progress in document summarization and classification, yet these technologies still struggle with the complexity of legal texts and their variability across different legal contexts [3][4].
Despite these advancements,research falls short in addressing the full scope of legal research challenges. Many tools remain limited by their narrow focus on keyword matching, lacking the adaptability needed for complex, multi-faceted legal documents. The datasets used in developing these tools are often small or domain-specific, reducing the tools' effectiveness in handling diverse cases or jurisdictions. These gaps limit the usability and impact of traditional and current NLP tools, which struggle with real-world applications due to inconsistencies in performance and the absence of deeper contextual understanding.
Our approach leverages generative AI and a custom LLM with built-in metacognitive reasoning to address these limitations. By understanding context and learning from interactions, our model continuously improves its relevance and precision in legal research. It can synthesize vast amounts of information, analyze previous cases, and provide coherent arguments based on past judgments and patterns in case law. Our approach also includes advanced data security protocols, such as blockchain, to protect client confidentiality and legal integrity, making the research process not only more efficient and insightful but also secure and trustworthy.
OBJECT OF THE INVENTION:
The primary object of the invention is to develop a generative AI-based legal research assistant that efficiently analyzes and synthesizes information from diverse legal databases, enabling comprehensive legal research capabilities.
Another object of the invention is to implement a multi-agent collaboration system that enhances the reliability and accuracy of AI-generated legal analysis, drafting and answers through simultaneous working of agents.
Yet Another object of the invention is to integrate retrieval augmented generation (RAG) pipeline within the system to ensure AI outputs are consistently grounded in verifiable legal information, templates, precedents, and authoritative sources.
Another object of the invention is to utilize chain of thought (COT) prompting strategies to enhance the transparency and explainability of AI reasoning processes in legal analysis, providing clear logical pathways for conclusions reached in document processing.
Yet Another object of the invention is to facilitate real-time validation of legal case citations and references through automated cross-referencing with authenticated legal databases, ensuring the accuracy of legal information provided.
Another object of the invention is to provide customizable legal research workflows that adapt to different jurisdictions, practice areas, and user preferences while maintaining consistency in legal analysis and interpretation.
Yet Another object of the invention is to provide a feedback integration system allowing legal professionals to provide input on system outputs through adaptive fine-tuning utilizing user feedback.
SUMMARY OF THE INVENTION:
In accordance with the different aspects of the present invention, an AI-driven legal research assistant is presented. The system assists legal professionals in significantly reducing case resolution time by delivering accurate and quick legal analysis and drafts. The system integrates a finetuned LLM (112), utilizing multi-agent collaboration (130) to enhance reasoning capabilities and retrieval augmented generation (RAG) (140) to improve accuracy.
The invention employs a fine-tuned large language model (LLM) (112), specifically adapted for Indian laws and procedures, along with metacognitive reasoning, enabling multi-agent collaboration (130) for comprehensive legal document analysis and automated generation of verified outputs. The system shows substantial potential for enhancing case resolution efficiency and improving decision-making quality within the Indian judicial system.
The present invention utilizes the capabilities of finetuned LLM (112) to create a sophisticated system capable of processing, analyzing, and synthesizing complex legal information. The implementation of a multi-agent collaborative framework (130) enhances the reliability and accuracy of AI-generated legal insights while substantially reducing the risk of hallucinations or misinterpretations ensuring consistent and dependable legal analysis across diverse case scenarios.
The invention's integration of retrieval augmented generation (RAG) technology (140) facilitates efficient access and utilization of extensive legal knowledge repositories (142). This approach ensures all AI outputs are firmly grounded in factual information and relevant legal precedents. Furthermore, the implementation of RAG pipeline (140) provides access to multiple legal sources (140-143), addressing a crucial requirement in the legal domain where decision rationale holds equal importance to the final conclusions.
Additional aspects, advantages, features and objects of the present disclosure would be made apparent from the drawings and the detailed description of the illustrative embodiments constructed in conjunction with the appended claims that follow.
BRIEF DESCRIPTION OF THE DRAWINGS
A preferred embodiment of the invention is now described in detail. Referring to the drawings, numbers indicate parts throughout the views. Unless explicitly stated in the following disclosure, the drawings are not necessarily drawn to scale. As used in the description herein and throughout the claims, the following terms take the meanings explicitly associated herein unless the context dictates otherwise: the meaning of "a," "an." And "the" includes plural reference, the meaning of "in" includes "in" and "on.". The summary of the present invention, as well as the detailed description, are better understood when read in conjunction with the accompanying drawings that illustrate one or more possible embodiments of the present invention, which:
Fig. 1: A system architecture diagram of an AI-based legal research assistant
Fig. 2: A flow diagram depicting the data flow over the AI Legal Research Assistant
Fig. 3: Detailed system process flow diagram
The present disclosure describes a comprehensive legal research assistant system (100) designed to transform legal research and analysis in Legal System. The following detailed description provides a thorough understanding of the various embodiments and components of the invention
Referring to the core architecture of the invention as shown in Fig. 1, the system employs a transformer-based architecture as the foundational model (112). The system undergoes specialized domain adaptation through QLoRA (Quantized Low-Rank Adaptation) technique, processing a carefully curated collection of Indian commercial law documents, including statutes, case law, and authoritative legal commentaries
In one embodiment, the invention implements a sophisticated multi-agent ecosystem comprising specialized agents (131-137), each designed for specific legal tasks. These agents communicate and exchange information through a feedback mechanism, which provides performance assessment to the agents and enhances the quality of their outputs.The inter-agent communication system employs a message-passing protocol (130), enabling collaborative refinement of outputs through shared insights and coordinated analysis
The Document Analysis Agent extracts essential details from legal documents and processes information by applying legal principles and precedents to cases, supporting case assessment through reasoning aligned with legal standards. The Citations Agent retrieves relevant citations from legal data sources and the internet, verifying that referenced cases and statutes are current and applicable. The Draft Agent generates reports or template drafts based on user queries using the provided data. The Bias Detection Agent evaluates the outputs of other agents for potential biases, ensuring objectivity and fairness in the legal analytical process. The QA Generation Agent creates concise, comprehensible summaries of complex legal arguments, enhancing understanding by converting intricate points into accessible insights
The invention implements a hierarchical chain of thought (HCoT) mechanism used by Document Analysis Agent featuring multi-level reasoning capabilities that decompose complex legal issues into manageable components, and layered explainability features providing detailed reasoning at each hierarchical level, from case strategy to specific interpretations
In a further embodiment, the invention integrates neuro-symbolic components including a symbolic layer implemented in Python for encoding explicit legal rules and procedures, and a custom neural-symbolic interface enabling logical inference queries between system components
The invention incorporates continuous learning mechanisms comprising a feedback integration system allowing legal professionals to provide input on system outputs, and adaptive fine-tuning protocols utilizing new legal data and user feedback while maintaining system stability through elastic weight consolidation
By implementing the above techniques the system provides transparent step-by-step reasoning processes, enhancing both analysis quality and decision transparency. The invention's integration of retrieval augmented generation with domain-specific optimizations and multi-agent architecture significantly improves the efficiency and accuracy of legal research processes
DETAILED DESCRIPTION OF THE INVENTION
The detailed disclosure presents a legal research assistant system (100) engineered to support legal professionals through advanced research methodologies. The system facilitates comprehensive legal research by analyzing law points from specific state legislations and examining relevant historical case records (142) to enhance legal planning. Understanding precedents, which is fundamental to effective legal strategy, is achieved through the system's deep analysis of past judgments pertinent to the case at hand, identifying significant patterns and outcomes that inform strategic decision-making. This analytical capability enables the anticipation of potential challenges and formulation of responses based on historical judicial data (230).
The system performs meticulous analysis through multiple case-critical factors, encompassing historical judgments and opposing party backgrounds. This exhaustive examination ensures comprehensive coverage of all relevant aspects, establishing a robust foundation for legal strategy development. The Multi-Agent System (130, 210) extends beyond analytical capabilities to generate precise legal brief documents adhering to established best practices. These briefs are systematically constructed to meet superior legal documentation standards, presenting structured case outlines supported by relevant data and precedential references.
At the core of the system lies a pre-trained transformer model with 8 billion parameters, meticulously fine-tuned (112) on our proprietary dataset encompassing over one million judgments from courts and tribunals across the world, along with comprehensive state-wise legal statutes. This extensive corpus ensures comprehensive coverage of legal precedents and decisions, facilitating efficient retrieval of pertinent case laws for various legal applications, from case preparation to academic research and legal development tracking.
The fine-tuning process leverages QLoRA (Quantized Low-Rank Adaptation) methodology, which strategically balances memory efficiency with model performance (220). This sophisticated approach begins by implementing 4-bit quantization of the base model weights, creating a compressed representation that maintains critical information while reducing memory footprint. The adaptation process then incorporates low-rank matrices through LoRA decomposition, establishing an efficient parameter space that captures essential model adjustments while minimizing computational overhead. The training process implements a specialized legal domain loss function that carefully balances standard text generation capabilities with specific legal context preservation requirements.
The model undergoes systematic fine-tuning across multiple epochs using a meticulously curated legal corpus (142), encompassing structured legal documents, case precedents, and statutory interpretations. The training architecture implements gradient accumulation techniques with an effective batch size of 128, utilizing mixed-precision training combined with a carefully calibrated learning rate schedule. The process incorporates domain-specific optimizations including specialized legal token embeddings and attention mechanisms specifically tailored for legal document structure. Throughout the training process, the model's performance undergoes continuous evaluation using domain-specific metrics, including legal accuracy scores, precedent relevance measurements, and citation accuracy verification, ensuring optimal performance in legal document analysis and generation tasks.
The system architecture implements a sophisticated multi-agent ecosystem (130, 210) built on principles of distributed cognitive systems, where specialized agents operate within a hierarchical framework optimized for legal domain tasks. This ecosystem employs a dynamic agent orchestration protocol that enables both parallel and sequential processing patterns, facilitating complex task decomposition and collaborative problem-solving through a distributed computation model. The inter-agent communication protocol implements a sophisticated message-passing architecture based on an enhanced version of the Actor Model, where each agent functions as an independent computational unit with dedicated message queues and processing logic.
The Document Analysis Agent (DAA) does the document processing pipeline, implementing a hierarchical chain of thought model (110) that processes legal documents through a structured, multi-layered reasoning mechanism. The multi-layered reasoning is a reasoning mechanism which, . The DAA breaks down complex legal documents into manageable abstractions, enabling deep understanding of content, context, and structure. This multi-level approach utilizes recursive chain of prompting designed to cascade outputs from each level as inputs to subsequent levels, creating a recursive data flow that enhances depth of analysis. The agent processes documents through multiple hierarchical levels, with each level focusing on increasingly complex aspects of document understanding, from basic text analysis to sophisticated legal reasoning and contextual interpretation.
The document structuring component employs a graphs (113) that models relationships between different sections of legal documents. This graphs uses edge features to represent different types of legal relationships (e.g., definitions, exceptions, conditions) and uses parallel memory manager (122) to propagate information across the document graph. The graph structure is dynamically updated based on detected relationships and cross-references within the document.
The Citations Agent (137, 317) implements a sophisticated retrieval-augmented generation architecture integrated with a temporal relevance scoring mechanism. This agent maintains and analyzes citation relationships through a dynamic citation graph that tracks both direct references and implicit connections between legal documents. The system continuously updates citation relationships while considering temporal relevance, precedential weight, and jurisdictional applicability. This enables the agent to not only identify relevant citations but also understand their evolving significance within the legal landscape.
The system maintains a dynamic knowledge base (113) using an efficient index structure with similarity search using semantic search manager (123) that enables fast retrieval of relevant citations. The citation relationships are stored in a specialized graph database that maintains temporal ordering and allows for efficient updates and queries. The retrieval process uses approximate nearest neighbor search with specialized legal domain metrics that consider both semantic similarity and citation patterns.
The Drafting Agent (131, 311) utilizes a prompting with legal domain-specific attention mechanisms. This agent excels in generating structured legal documents by implementing sophisticated template adaptation mechanisms that dynamically adjust to specific document requirements. The agent maintains awareness of document structure, legal formatting requirements, and context-specific content generation needs. It ensures that generated documents maintain consistency in style, tone, and legal accuracy while adhering to jurisdiction-specific requirements and formatting standards.
The template adaptation mechanism utilizes templates from database (143) that combines pre-defined structural templates with dynamic content generation. This is implemented through a hybrid architecture that combines retrieval-based template selection with generative refinement. The template selection component uses a prompting with knowledge graphs (113) to match document requirements with appropriate templates, while the refinement component employs a conditional transformer that generates content while maintaining consistency with the selected template.
The Neutralizer Agent (132, 312) implements a multi-dimensional bias detection that analyzes textual content across various bias categories. This agent employs a fairness-aware neural architecture with specialized attention mechanisms designed to identify and flag potential biases in legal content. The bias detection process utilizes counterfactual testing, where the system generates alternative versions of text with protected attributes modified, allowing for direct comparison and bias measurement. This is implemented through a conditional generation framework that maintains semantic consistency while varying potentially biased elements.
This bias detection system considers general bias i.e not limited to procedural fairness, substantive fairness, and equal treatment under the law. The agent's sophisticated analysis helps ensure that generated legal content maintains objectivity and fairness while adhering to legal principles of equity and justice.
Then Neutralizer Agent (132, 312) implements a sophisticated bias mitigation system neutralizes identified biases while preserving the legal substance and intent of the content. At its core, the Neutralizer Agent employs a modified sequence-to-sequence prompting with specialized attention mechanisms for bias-aware content transformation. The system implements a two-stage processing pipeline: first, a bias understanding phase that processes the output from the BDA, and second, a content reformation phase that generates neutralized text while maintaining legal accuracy.
The bias understanding component utilizes a specialized encoder that processes both the original text and the bias analysis metadata. This encoder implements cross-attention recognition that allows it to focus on specific regions and aspects of bias identified. The encoder creates a rich representational mapping that captures both the content semantics and bias characteristics of the input text.
The Validator Agent (136, 316) implements a comprehensive validation framework that ensures the accuracy, reliability, and consistency of legal documents through a multi-stage verification process. The system maintains a dynamic validation state that tracks the confidence level of different aspects of the document. This is implemented using retrieval augmented generation (RAG) (140) using citations through multiple validation to produce an overall reliability score. The validation process is iterative, with the agent capable of suggesting and verifying corrections in real-time.
The Validator Agent (136, 316) also implements a specialized attention mechanism that allows it to focus on different aspects of validation depending on the document type and jurisdictional requirements. This is achieved through a context-aware attention router that directs validation resources based on document characteristics and validation priorities.
The QA Generation Agent (135, 315) utilizes a prompt optimized for legal document summarization and question-answer pair generation. This agent specializes in creating comprehensive summaries of legal documents while identifying key points that might require clarification or elaboration. The agent generates relevant questions and corresponding answers that help users better understand complex legal documents, ensuring that critical legal concepts and implications are properly communicated and understood.
The QA Generation Agent (135, 315) is used for processing documents at multiple levels (sentence, paragraph, section) to generate coherent summaries. The question generation process uses a modified seq2seq architecture with copy mechanisms that allow direct copying of legal terms and citations when appropriate.
All these agents operate within a shared computational framework that enables efficient information sharing and coordination. The system implements a distributed computation where agents can be updated independently while maintaining consistency through shared embeddings and cross-agent attention mechanisms. The entire system is optimized using a combination of feedback learning on multi agent framework (130, 210).
The inter-agent coordination framework implements a sophisticated memory sharing with parallel memory manager (122) that enables collaborative learning across agents through a shared memory buffer through knowledge graphs (113). This coordination system ensures smooth information flow between agents while maintaining data consistency and accuracy. The system implements a collective knowledge integration mechanism where agent outputs are combined through carefully calibrated weighting mechanisms, ensuring that the final output represents a coherent synthesis of multiple specialized analyses.
Through this sophisticated integration of specialized agents (130, 210), the system achieves a synergistic processing capability where the collective output exceeds the sum of individual agent contributions. This enables comprehensive legal document analysis and generation with high accuracy and reliability, while maintaining the flexibility to adapt to varying legal requirements and document types. The system's architecture ensures continuous learning and improvement through feedback loops and performance monitoring, making it increasingly effective at handling complex legal documentation tasks over time.
The invention's process commences when a user submits a legal query or document through the frontend interface, initiating a multi-stage processing pipeline. Upon submission, the system immediately generates a unique request identifier and performs preliminary validation of the input format. The validated request is then transmitted through an API gateway, which implements comprehensive request sanitization protocols before forwarding to the central controller class (301). This controller (300) serves as the orchestration hub for the entire processing workflow, generating distinct AI session identifiers for each agent instance and establishing a shared memory space that facilitates efficient inter-agent communication throughout the processing lifecycle.
The controller (300) implements a dynamic resource allocation mechanism that evaluates request complexity and initializes appropriate computational resources. This allocation process considers factors such as document length, query complexity, and required processing depth. The controller maintains a state machine that tracks the progression of each request through various processing stages, ensuring proper sequencing and synchronization of agent operations. This orchestration layer establishes monitoring channels for real-time agent processing and maintains a processing queue that optimizes multi-agent operations based on priorities.
The document analysis phase begins with the Document Analysis Agent (DAA) implementing a hierarchical decomposition process (310). This agent processes the input through multiple abstraction levels, beginning with sentence-level tokenization and analysis, proceeding to paragraph-level contextual grouping, and culminating in section-level relationship mapping. The agent employs parallel memory managers that facilitate both forward propagation of definitions and backward propagation of dependencies, ensuring comprehensive context preservation across the document structure.
Simultaneously, the Citation Agent (317) executes a finding citations sequence. This agent implements semantic search for explicit citation identification while simultaneously conducting semantic analysis to uncover implicit references. The CA calculates temporal relevance scores for each identified citation, considering factors such as citation date, precedential weight, and jurisdictional applicability. The agent maintains a dynamic citation graph that undergoes continuous updates as new relationships are discovered and obsolete citations are pruned.
The drafting process is executed by the Drafting Agent (311), which implements a hybrid architecture combining template-based and generative approaches. The DA first performs document type classification and identifies applicable jurisdictional requirements and formatting constraints. It then selects appropriate templates from a dynamic template database, implementing real-time adaptations based on specific content requirements. The content generation process maintains consistency in legal terminology while ensuring compliance with jurisdictional formatting standards. The DA employs specialized attention mechanisms that focus on legal domain-specific aspects during content generation, ensuring accurate representation of legal concepts and principles.
Bias detection and neutralization occur through a sophisticated two-phase process. The Neutralizer Agent (312) first implements a multi-dimensional analysis that identifies potential biases across various protected attributes. This process generates counterfactual versions of the content to measure bias impact quantitatively, then executes a two-stage processing sequence: first, a bias understanding phase that creates detailed context preservation mappings and identifies core legal intent; second, a content reformation phase that implements bias-aware text transformations while maintaining legal accuracy and semantic consistency.
The validation process implements a comprehensive quality assurance through the Validator Agent (316). This agent executes multiple validation cycles, verifying citation accuracy, legal principle compliance, and jurisdictional requirement adherence. The validation process maintains a dynamic validation state that tracks confidence levels across different aspects of the document. When discrepancies are detected, the agent initiates an iterative refinement sequence, generating specific corrections and revalidating modified content until all quality thresholds are met.
The system's integration layer implements sophisticated output aggregation mechanisms through a shared computational framework. This framework maintains consistency through shared embeddings and cross-agent attention mechanisms, enabling efficient information sharing while preserving the independence of individual agent operations. The final output assembly process incorporates proper formatting and citations, with comprehensive quality scoring ensuring adherence to predefined quality thresholds.
Throughout the entire process, the system maintains continuous feedback loops that enable dynamic optimization of agent operations. This includes collecting detailed response quality metrics, tracking agent performance parameters, and analyzing error patterns. The system implements cross-request learning mechanisms that enable progressive improvement in processing accuracy and efficiency. The methodology's modular design allows for independent optimization of individual components while maintaining system coherence through well-defined interfaces and communication protocols.
The process concludes with a final quality assurance check that verifies the completeness and accuracy of the generated output. This includes validation of all citations, verification of formatting compliance, and confirmation of response relevance to the original query. The system then generates the appropriate response format based on the initial request type and transmits it back to the user through the API interface. This comprehensive methodology ensures the generation of high-quality legal content while maintaining strict compliance with legal standards and requirements.
, Claims:1. A Gen AI-based Legal Research Assistant using Multi-Agent Collaboration, Retrieval Augmented Generation, and Chain of Thought Prompting, the architecture comprising: a plurality of collaborative agents (130) with metacognitive reasoning, wherein said agents employ an AI engine (110) configured with a fine-tuned large language model (LLM) (112), contextual memory (111), and a knowledge graph (113); and a controller class (121) comprising a parallel memory manager (122) and a semantic search manager (123) adapted to interface with designated legal data sources (140).
2. The system as claimed in claim 1, wherein the collaborative agents (130) include a Citation Agent (137), Drafting Agent (131), Validator Agent (136), Analyzer Agent (133), Neutralizer Agent (132), and QA Agent (135), each configured to collaboratively analyze legal information, interface with various data sources (140), and deliver accurate legal insights.
3. The system as claimed in claim 1, wherein the metacognitive reasoning enables each collaborative agent (130) to execute multi-step reasoning, providing agents with the ability to leverage data sources (140) and the knowledge graph (113) to collaborate effectively with the reasoning capabilities of other agents, thereby producing enhanced collective output.
4. The system as claimed in claim 1, wherein the collaborative approach and continuous refinement process mitigates potential bias within AI-generated legal analysis and insights, the architecture integrating collaborative agents (130) with metacognitive reasoning employing the AI engine (110) comprising the LLM (112), contextual memory (111), knowledge graph (113), and controller class (121) with a parallel memory manager (122) and semantic search manager (123) to interface with legal data sources (140).
5. The system as claimed in claim 2, wherein a Document Analysis Agent (DAA) implements a document processing pipeline utilizing a hierarchical chain-of-thought model (110) to analyze legal documents, wherein the DAA deconstructs complex legal documents into manageable abstractions, facilitating a comprehensive understanding of content, context, and structure.
6. The system as claimed in claim 1, wherein the Neutralizer Agent (132, 312) performs multi-dimensional bias detection by analyzing textual content across various bias categories, employing fairness-aware reasoning with an attention mechanism focused on relevant context, designed to identify and flag potential biases within legal content.
7. The system as claimed in claim 1, wherein the Validator Agent (136, 316) maintains a dynamic, iterative, and real-time validation state, tracking the confidence levels of various document aspects through retrieval-augmented generation (RAG) (140) across multiple validation steps to produce a comprehensive reliability score.

Documents