AI for Creating Standardized Reports

Introduction: This thesis focuses on the intelligent
generation of standardized reports - specifically the
8D report, a widely adopted framework in quality
management used to systematically identify root
causes, implement corrective actions, and ensure
continuous improvement in response to customer
complaints or internal quality issues.

Approach: Rather than fine-tuning a dedicated model
for each specific report template, this thesis
investigates the use of general-purpose AI models -
specifically the large language model (LLM) GPT-4o-
mini - to interpret report structures and task
specifications through prompt engineering and
Retrieval-Augmented Generation (RAG). The system
aims to retrieve relevant information from
heterogeneous data sources and accurately fill the
required report fields, while maintaining constraints
such as report structure, data type consistency, object
integrity, and completeness. Key challenges include:
- Secure handling of sensitive and business-critical
data
- Understanding and processing unstructured
information from diverse sources
- Supporting flexible, customizable report formats
- Minimizing LLM hallucinations and ensuring factual
correctness
- Integrating human-in-the-loop validation for final
quality assurance

To address these challenges, an LLM-powered
pipeline is developed that includes input document
preprocessing, data vectorization, and structured
report content generation. The pipeline also features
an automated self-reflection step that guides the user
in addressing missing or suspicious field values
before intelligently updating the report based on user
instructions. Various strategies - such as few-shot
prompting, static and dynamic query generation,
multi-turn interactions, and ReAct-style agents - are
evaluated to identify optimal configurations. Several
open-source frameworks are assessed, with
particular focus on the toolkit Docling, valued for its
ability to handle multiple document formats, perform
advanced PDF parsing and Optical Character
Recognition (OCR), and support integrated RAG
workflows for automated report generation.

Result: The proposed generative AI-based pipeline is
evaluated for reliability, efficiency, cost, and
scalability. Results show that high-quality
standardized report template creation and content
generation can be achieved without relying on rigid,
code-based applications. Based on the observations
and research, the thesis proposes different
combinations of tools and engineering strategies to
achieve an optimal cost-performance balance,
depending on task complexity and input structure.
The final prototype demonstrates that, with thoughtful
system design, AI-driven report generation is both
practical and adaptable for real-world business use.