1. An automated retrieval-augmented educational question–answer generation system (100) configured to generate, evaluate, classify, and export pedagogically structured question–answer pairs from academic documents, the system comprising: a preprocessing Module (102) configured to: i. receive one or more input documents; ii. extract textual content using a PDF parsing sub-module (104); and iii. segment the extracted textual content into overlapping contextual chunks using an Overlapping Chunk Segmentation sub-module (106); a retrieval Module comprising: i. a Sentence Embedding component (108) configured to convert each contextual chunk into a dense vector representation; and ii. a FAISS nearest-neighbor retrieval module (110) configured to index the dense vector representations and retrieve top-K semantically relevant content chunks based on similarity search; a dual Prompt Generation Module (112) operatively coupled to the Retrieval Module, the Dual Prompt Generation Module (112) comprising i. a Grounded QA Generation sub-module (114) configured to generate one or more question–answer pairs using the retrieved top-K content chunks; and ii. a Bloom’s Cognitive Diversity Enforcement sub-module (116) configured to generate question–answer pairs aligned with multiple cognitive levels of Bloom’s taxonomy; a rubric Evaluation Module (118) configured to evaluate each generated question–answer pair based on predefined multi-criteria quality parameters and generate an evaluation score; a bloom’s Classification component comprising: i. a Heuristic Keyword Mapping module (120) configured to assign a cognitive level label to each generated question–answer pair;and ii. a dataset Export Module (124) configured to compile and export the evaluated and classified question–answer pairs into a structured, labeled dataset including question content, answer content, evaluation scores, cognitive classification labels, and format labels; wherein the system is configured to operate as an integrated end-to- end pipeline for generating retrieval-grounded, cognitively classified, and quality-evaluated educational question–answer datasets.

2. The system as claimed in claim 1, wherein the PDF parsing sub-module (104) is configured to extract textual content from digitally native PDF files and scanned PDF documents using optical character recognition techniques.

3. The system as claimed in claim 1, wherein the Overlapping Chunk Segmentation sub-module (106) segments textual content using a sliding window mechanism with a predefined overlap ratio to preserve contextual continuity between adjacent chunks.

4. The system as claimed in claim 1, wherein the Sentence Embedding component (108) utilizes a transformer-based language model to generate semantically meaningful dense vector representations of each contextual chunk.

5. The system as claimed in claim 1, wherein the FAISS nearest-neighbor retrieval module (110) performs similarity search using cosine similarity or inner product distance metrics to retrieve top-K relevant content chunks.

6. The system as claimed in claim 1, wherein the Grounded QA Generation sub-module (114) is configured to constrain generation of question–answer pairs strictly to the retrieved top-K content chunks to ensure factual grounding.

7. The system as claimed in claim 1, wherein the Bloom’s Cognitive Diversity Enforcement sub-module (116) incorporates predefined instructional verbs and cognitive directives corresponding to Bloom’s taxonomy levels.

8. The system as claimed in claim 1, wherein the Rubric Evaluation Module (118) evaluates generated question–answer pairs using multiple quality parameters including accuracy, clarity, completeness, reasoning depth, and contextual faithfulness.

9. The system as claimed in claim 1, wherein the Rubric Evaluation Module (118) applies weighted aggregation to compute a composite quality score for each generated question–answer pair.

10. The system as claimed in claim 1, wherein the Heuristic Keyword Mapping module (120) assigns cognitive level labels based on rule-based detection of action verbs, syntactic structures, and semantic cues within the generated question.

6. DATE AND SIGNATURE Dated this on 07th March, 2026 Signature Mr. Srinivas Maddipati (IN/PA 3124) Agent for applicant

Description:DESCRIPTION
Technical Field of the Invention

The present invention relates generally to the field of artificial intelligence, educational technology, and natural language processing. More particularly, the invention relates to a retrieval-augmented framework for generating pedagogically aligned and evidence-grounded question answer (QA) pairs from academic documents using large language models integrated with Bloom’s Taxonomy classification and rubric-based evaluation mechanisms.

Background of the Invention

The rapid advancement of large language models (LLMs) has significantly influenced automated content generation in educational environments. Such models are increasingly employed for generating question answer (QA) pairs from academic documents to support tutoring systems, automated assessments, and curriculum-aligned instructional platforms. However, despite their capability to produce linguistically fluent outputs, conventional standalone LLM-based systems frequently generate responses that are not fully grounded in the source material, which may lead to factual inconsistencies and the generation of hallucinated content.

To mitigate such issues, retrieval-augmented generation (RAG) architectures have been introduced, wherein semantic retrieval mechanisms are used to retrieve relevant document segments prior to generating responses. Such systems typically employ vector similarity search techniques to identify contextually relevant information from source documents. Although RAG-based approaches improve factual grounding of generated responses, existing implementations are primarily optimized for open-domain question answering and are not specifically designed for structured pedagogical question–answer generation. Consequently, such systems do not inherently ensure cognitive diversity or systematic coverage aligned with established educational taxonomies.

Educational assessment frameworks such as Bloom's Taxonomy emphasize hierarchical cognitive levels ranging from remembering and understanding to analyzing, evaluating, and creating. Current automated QA generation systems tend to produce predominantly lower-order recall-based questions, lacking balanced representation across higher-order cognitive skills. Moreover, Bloom’s classification models typically operate independently of the generation pipeline, resulting in disjointed workflows without integrated pedagogical control.

The evaluation of LLM-generated QA pairs commonly relies on shallow metrics or single-score “LLM-as-judge” approaches. These methods lack structured, multi-dimensional rubric-based evaluation mechanisms that assess factual grounding, cognitive level alignment, clarity, completeness, and pedagogical suitability. Furthermore, most systems do not incorporate back-retrieval validation or evidence verification mechanisms to systematically detect hallucinations, nor do they produce benchmark-ready, annotated datasets suitable for research and curriculum-aligned deployment.

Despite the availability of individual components such as semantic retrieval, Bloom’s classifiers, and rubric-based evaluators, there is currently no unified framework that integrates retrieval grounding, cognitive-level-aware prompting, structured rubric scoring, classification, hallucination validation, and benchmark dataset creation into a single end-to-end pipeline.

To address these limitations, the present invention provides an integrated retrieval-augmented, Bloom-classified framework for educational question–answer generation that combines semantic chunk retrieval, dual prompting strategies (baseline and Bloom-aware), structured JSON output enforcement, a multi-criterion pedagogical rubric evaluation engine, heuristic Bloom’s and format classification, and hallucination validation via back-retrieval, thereby producing evidence-based, cognitively diverse, and benchmark-ready QA datasets within a unified system architecture.

Objects of the Invention

The principal object of the present invention is to provide an integrated retrieval-augmented educational question answer (QA) generation framework that combines semantic document retrieval with large language model-based generation to ensure factual grounding and contextual relevance.

Another object of the present invention is to generate QA pairs systematically aligned with Bloom's Taxonomy, thereby promoting balanced cognitive diversity across multiple levels including remembering, understanding, applying, analyzing, evaluating, and creating.

Another object of the present invention is to provide the multi-criterion rubric-based evaluation engine for assessing the pedagogical quality, factual accuracy, clarity, completeness, and educational suitability of generated QA pairs.

Another object of the present invention is to provide an automated hallucination validation mechanism using back-retrieval and evidence verification to significantly reduce unsupported or factually incorrect answers in generated content.

Another object of the present invention is to provide a curated, annotated, benchmark-ready dataset of evidence-based QA pairs with cognitive and pedagogical labels, suitable for educational assessment systems, tutoring platforms, and AI benchmarking research.

Brief Summary of the Invention

The present invention provides a retrieval-augmented framework for educational question answer (QA) generation is provided, wherein academic documents are parsed, segmented into overlapping semantic chunks, and indexed using a vector-based retrieval mechanism to enable contextually grounded content generation. The system ensures that all generated QA pairs are derived from retrieved evidence, thereby significantly reducing hallucinations and improving factual reliability.

In accordance with the aspect of the present invention, a dual prompting architecture is implemented comprising a baseline grounded prompt and a Bloom's Taxonomy-aware prompt, which together enforce cognitive diversity across multiple hierarchical learning levels. This structured prompting mechanism enables systematic generation of questions spanning lower-order and higher-order cognitive skills aligned with pedagogical objectives.

In accordance with the aspect of the present invention, a structured JSON-based output module and a multi-dimensional rubric evaluation engine are integrated into the pipeline to assess generated QA pairs using a defined set of quality criteria, including factual alignment, clarity, completeness, and cognitive appropriateness. The rubric scoring mechanism enables weighted aggregation and standardized quality control of generated outputs.

In accordance with the aspect of the present invention, an automated hallucination validation module utilizing back-retrieval verification is incorporated to cross-check generated answers against source document evidence, thereby enhancing reliability and auditability. The framework further includes heuristic cognitive-level and pedagogical format classification modules to annotate each QA pair with structured metadata.

In accordance with the aspect of the present invention, the system produces a curated, benchmark-ready dataset comprising evidence-grounded, cognitively classified, and rubric-scored QA pairs, thereby enabling scalable curriculum-aligned assessment generation, intelligent tutoring applications, and educational AI research within a unified end-to-end architecture.

Brief Description of Drawings

The invention will be further understood from the following detailed description of a preferred embodiment taken in conjunction with an appended drawing, in which:

Figure 1 illustrates the overall system(100) architecture and operational pipeline of the EDURAG-QA retrieval-augmented, Bloom-classified educational question–answer generation framework.

REFERENCE NUMERALS
100 – Overall EDURAG-QA system
102 – Preprocessing Module
104– PDF parsing
106 – Overlapping chunk segmentation
108 – Sentence embedding
110 – FAISS nearest-neighbor retrieval
112 – Dual Prompt Generation Module
114 – Grounded QA generation
116 – Bloom’s cognitive diversity enforcement
118 – Rubric Evaluation Module
120 – Heuristic keyword mapping
122 – Dataset Export Module

Detailed Description of the Invention

The present invention will now be described in detail with reference to the accompanying drawing. The following description is provided to illustrate the invention and should not be construed as limiting the scope of the invention.

The manner in which the present invention is formulated is described in greater detail below with reference to the accompanying drawing. It is to be understood that the drawing illustrates a typical embodiment of the invention and is not intended to limit the scope of the invention. The system may admit modifications and alternative configurations without departing from the spirit and scope of the invention.

According to the exemplary embodiment of the present invention, the invention relates to an automated educational question–answer generation framework designated as the overall EDURAG-QA system (100), configured to generate, evaluate, classify, and export pedagogically structured question–answer pairs from academic source documents through an integrated retrieval-augmented architecture.

According to the exemplary embodiment of the present invention, The system comprises a preprocessing module (102) configured to receive one or more academic documents in digital format. The preprocessing module (102) includes a PDF parsing sub-module (104) configured to extract machine-readable textual content from the input documents. The extracted content is further processed by an overlapping chunk segmentation sub-module (106) configured to divide the text into sequential segments using a sliding window mechanism with a predefined overlap so as to preserve contextual continuity between adjacent segments and enhance semantic coherence during the retrieval process.

According to the exemplary embodiment of the present invention, the segmented text chunks are processed by a Sentence Embedding component (108), which converts each chunk into a dense vector representation using a transformer-based embedding model. The generated embeddings are stored and indexed through a FAISS nearest-neighbor retrieval module (110), configured to perform similarity-based search operations using cosine similarity or inner-product distance metrics. Upon receiving a query or prompt instruction, the FAISS nearest-neighbor retrieval module (110) retrieves top-K semantically relevant chunks to ground subsequent question–answer generation in verified source material.

According to the exemplary embodiment of the present invention, the retrieved contextual chunks are provided to a dual prompt generation module (112) including a grounded QA generation sub-module (114) configured to generate question–answer pairs based on retrieved contextual evidence, and a Bloom’s cognitive diversity enforcement sub-module (116) configured to generate questions across multiple cognitive levels aligned with Bloom’s taxonomy.

According to the exemplary embodiment of the present invention, the generated question answer pairs are evaluated by a Rubric Evaluation Module (118), which assesses each output based on predefined quality criteria including accuracy, clarity, completeness, reasoning depth, and contextual faithfulness. The evaluation mechanism may include weighted aggregation of multiple scoring parameters to produce an overall quality score for each generated pair.

According to the exemplary embodiment of the present invention, the system further comprises a Heuristic Keyword Mapping module (120), configured to classify each generated question into an appropriate Bloom’s cognitive level based on rule-based detection of instructional verbs, syntactic patterns, and semantic cues.
According to the exemplary embodiment of the present invention, the classified and evaluated question answer pairs are compiled by a Dataset Export Module (124), which generates a structured, labeled benchmark dataset in a machine-readable format such as JSON or CSV. The exported dataset includes question content, answer content, cognitive classification labels, format labels, evaluation scores, and associated metadata, thereby enabling deployment in learning management systems, assessment engines, adaptive learning platforms, and digital question banks.

Accordingly, the present invention provides an integrated, retrieval-grounded, cognitively classified, and quality-evaluated framework for scalable educational question–answer generation, while allowing modifications and equivalent implementations without departing from the scope of the invention.

Figure 1 illustrates the overall system(100) architecture and operational pipeline of the EDURAG-QA retrieval-augmented, Bloom-classified educational question–answer generation framework.
As shown in Figure 1, the system is divided into three major operational phases:
• Phase A – Ingestion and Retrieval,(102)
• Phase B – Multi-Prompt Question–Answer Generation, and(112)
• Phase C – Evaluation and Classification.(116)
Phase A(102) includes document ingestion(104), text extraction, chunk segmentation(106), embedding generation(108), and similarity-based retrieval to obtain top-K(110) relevant content chunks from academic documents.
Phase B includes dual prompt design incorporating baseline grounding prompts and Bloom’s taxonomy-aware prompts, followed by retrieval-augmented language model-based question–answer generation and structured output formatting (114).
Phase C includes rubric-based multi-criteria evaluation (118), heuristic Bloom’s cognitive classification (120), pedagogical format labeling, and structured dataset(122) export including scores and metadata.
The figure collectively represents the end-to-end integrated pipeline for generating a classified and validated educational QA dataset.

Advantages of the Invention
The present invention provides improved factual reliability in educational question answer generation by incorporating a retrieval-augmented mechanism that grounds generated outputs in verified source documents. By retrieving top-K relevant content chunks before generation, the system significantly reduces hallucination and enhances traceability, thereby ensuring that the produced questions and answers remain contextually accurate and academically trustworthy.

The invention enhances pedagogical effectiveness by integrating Bloom’s taxonomy-aware prompting and automated cognitive classification. This enables the generation of diverse question types across multiple learning levels, including knowledge, comprehension, application, analysis, evaluation, and creation. Additionally, the inclusion of rubric-based multi-criteria evaluation ensures that each generated question answer pair meets predefined quality standards in terms of clarity, completeness, reasoning depth, and relevance.

The invention provides an automated end-to-end pipeline that integrates document ingestion, semantic retrieval, multi-prompt generation, evaluation, classification, and structured dataset export. The modular architecture enables scalability for large academic corpora and produces deployment-ready outputs in machine-readable formats, thereby reducing manual effort, saving time and cost, and facilitating integration with learning management systems and digital assessment platforms.
, Claims:CLAIMS
We Claim:
1. An automated retrieval-augmented educational question–answer generation system (100) configured to generate, evaluate, classify, and export pedagogically structured question–answer pairs from academic documents, the system comprising:
a preprocessing Module (102) configured to:
i. receive one or more input documents;
ii. extract textual content using a PDF parsing sub-module (104); and
iii. segment the extracted textual content into overlapping contextual chunks using an Overlapping Chunk Segmentation sub-module (106);
a retrieval Module comprising:
i. a Sentence Embedding component (108) configured to convert each contextual chunk into a dense vector representation; and
ii. a FAISS nearest-neighbor retrieval module (110) configured to index the dense vector representations and retrieve top-K semantically relevant content chunks based on similarity search;
a dual Prompt Generation Module (112) operatively coupled to the
Retrieval Module, the Dual Prompt Generation Module (112) comprising
i. a Grounded QA Generation sub-module (114) configured to generate one or more question–answer pairs using the retrieved top-K content chunks; and
ii. a Bloom’s Cognitive Diversity Enforcement sub-module (116) configured to generate question–answer pairs aligned with multiple cognitive levels of Bloom’s taxonomy;
a rubric Evaluation Module (118) configured to evaluate each generated
question–answer pair based on predefined multi-criteria quality
parameters and generate an evaluation score;
a bloom’s Classification component comprising:
i. a Heuristic Keyword Mapping module (120) configured to assign a cognitive level label to each generated question–answer pair;and
ii. a dataset Export Module (124) configured to compile and export the evaluated and classified question–answer pairs into a structured, labeled dataset including question content, answer content, evaluation scores, cognitive classification labels, and format labels;
wherein the system is configured to operate as an integrated end-to-
end pipeline for generating retrieval-grounded, cognitively
classified, and quality-evaluated educational question–answer
datasets.
2. The system as claimed in claim 1, wherein the PDF parsing sub-module (104) is configured to extract textual content from digitally native PDF files and scanned PDF documents using optical character recognition techniques.

3. The system as claimed in claim 1, wherein the Overlapping Chunk Segmentation sub-module (106) segments textual content using a sliding window mechanism with a predefined overlap ratio to preserve contextual continuity between adjacent chunks.

4. The system as claimed in claim 1, wherein the Sentence Embedding component (108) utilizes a transformer-based language model to generate semantically meaningful dense vector representations of each contextual chunk.

5. The system as claimed in claim 1, wherein the FAISS nearest-neighbor retrieval module (110) performs similarity search using cosine similarity or inner product distance metrics to retrieve top-K relevant content chunks.

6. The system as claimed in claim 1, wherein the Grounded QA Generation sub-module (114) is configured to constrain generation of question–answer pairs strictly to the retrieved top-K content chunks to ensure factual grounding.

7. The system as claimed in claim 1, wherein the Bloom’s Cognitive Diversity Enforcement sub-module (116) incorporates predefined instructional verbs and cognitive directives corresponding to Bloom’s taxonomy levels.

8. The system as claimed in claim 1, wherein the Rubric Evaluation Module (118) evaluates generated question–answer pairs using multiple quality parameters including accuracy, clarity, completeness, reasoning depth, and contextual faithfulness.

9. The system as claimed in claim 1, wherein the Rubric Evaluation Module (118) applies weighted aggregation to compute a composite quality score for each generated question–answer pair.

10. The system as claimed in claim 1, wherein the Heuristic Keyword Mapping module (120) assigns cognitive level labels based on rule-based detection of action verbs, syntactic structures, and semantic cues within the generated question.

6. DATE AND SIGNATURE

Dated this on 07th March, 2026
Signature

Mr. Srinivas Maddipati
(IN/PA 3124)
Agent for applicant