Cortex: Content Analysis Support Software, a Resource for Qualitative Research

1 Introduction

• Content analysis systematically describes and interprets textual content to support understanding of core messages.The paper situates this method within qualitative research focused on meanings and contexts.Source: §1 · p. 1
• Growing volumes of dense interviews and open-ended responses increase the time and cognitive effort needed to organize, read, and analyze content.Coder fatigue, deadlines, and complex manual coding rules can also compromise precision and introduce subjective biases.Source: §1 · p. 2
• Because qualitative analysis can delay the wider research process, tools should speed this phase without compromising analytical quality.The passage frames this as a potential support to scientific activity.Source: §1 · p. 2
• Cortex is a Bardin-based web application that uses an LLM to suggest indices, indicators, and thematic categories from a research corpus and theoretical framework.Suggestions retain verification and traceability to original source text.Source: §1 · p. 2

2 Methodology and System Design

• The study combined bibliographic research with four semi-structured interviews of researchers experienced in content analysis or qualitative methods.The interviews explored needs, difficulties, and expectations for digital tools, and were analyzed qualitatively to identify requirements.Source: §2 · p. 2
• Cortex’s data model reflects Bardin’s methodology, with entities for analyses, documents, indices, indicators, categories, and register units.The software was organized into Authentication, Configuration, Pre-Analysis, and Exploration modules.Source: §2 · pp. 2–3
• The application uses a decoupled client-server design, with a C#/.NET backend API and a React/TypeScript frontend.Figure 1 depicts the client interface, backend API, hybrid persistence layers, and LLM RAG mechanism.Source: §2 · p. 3, Figure 1 · p. 3
• During corpus configuration, document files are stored in Google Cloud Storage while PostgreSQL retains metadata and logical URIs.This hybrid strategy separates physical files from their database references.Source: §2 · p. 3
• For intelligent processing, Cortex uses Gemini through Vertex AI’s RAG mechanism, while its response is schema-validated and persisted for frontend display.The described processing path includes generated structured responses, deserialization into C# domain classes, and relational database storage.Source: §2 · pp. 3–4

3.1 Qualitative Research and Content Analysis

• Qualitative research seeks deep understanding of social phenomena, while content analysis interprets textual material; interpretation inevitably involves the researcher’s personal lens.A completely neutral, objective, exhaustive reading is described as unattainable.Source: §3.1 · p. 4

3.1.1 Pre-analysis

• Pre-analysis organizes the corpus and ideas, identifies documents, and prepares indicators by relating concrete textual indices to rules for assessing them.Bardin’s therapeutic-interview example links speech disturbances such as stuttering to an indicator such as their frequency; Table 1 is identified as an example of hypothesis, indices, and indicators.Source: §3.1.1 · pp. 4–5, Table 1 · p. 5

3.1.2 Exploration of the Material

• Material exploration applies pre-analysis decisions through coding, segmentation, categorization, and frequency counting.Bardin identifies computers as valuable for complex analyses involving multiple variables and occurrence tracking, which can accelerate otherwise manual work.Source: §3.1.2 · p. 5

3.1.3 Treatment of Results, Inference, and Interpretation

• Treatment of results condenses and represents findings, then uses inference to connect descriptive results with theory before interpretation.Content-analysis validity is grounded in the researched material, congruence with the theoretical framework, and alignment with the research objective—not a single true reading.Source: §3.1.3 · pp. 5–6

3.2 Natural Language Processing and Artificial Intelligence

• Natural Language Processing is an AI subfield focused on enabling computers to understand and process human languages.The passage notes that recent NLP advances have been driven particularly by machine learning and deep learning.Source: §3.2 · p. 6

3.3 Retrieval-Augmented Generation (RAG)

• RAG combines retrieval and generation, using chunking, embeddings, and similarity search to retrieve relevant document context for a query.The background presents this architecture in response to LLM limitations such as hallucinations and knowledge cutoffs.Source: §3.3 · p. 6

3.3.1 Vertex AI

• Vertex AI manages document ingestion, parsing, chunking, and indexing, making an indexed corpus available to Gemini through RAG.The platform is described as avoiding the need to manually build and maintain each pipeline component.Source: §3.3.1 · p. 6

4.1 ATLAS.ti

• ATLAS.ti supports document coding and code-relationship visualization, but its core workflow remains a manual workbench despite newer LLM features.Its structure is influenced by Grounded Theory.Source: §4.1 · pp. 6–7

4.2 NVivo

• NVivo organizes coded excerpts as Nodes and offers query and coding-matrix tools, while still requiring researchers to read, interpret, and code material manually.It is described as ATLAS.ti’s main competitor and has introduced LLM modules.Source: §4.2 · p. 7

4.3 Elicit

• Elicit uses RAG and natural-language queries to find articles and extract information, but targets literature review rather than primary data analyzed under a specific qualitative methodology..Source: §4.3 · p. 7

4.4 Comparative Analysis and Gap Identification

• ATLAS.ti and NVivo emphasize manual methodological rigor, whereas Elicit automates literature synthesis; none of the analyzed tools targets LLM assistance within Content Analysis phases.Cortex is proposed as a collaborator that uses Gemini to suggest preliminary indices, indicators, and categories from the corpus and theoretical framework without replacing the researcher.Source: §4.4 · p. 7

5 Results: The Cortex Software

• Cortex guides researchers through chronological Content Analysis phases in a web application designed to make LLM complexity less prominent and keep the workflow method-focused.The interface and underlying logic are described as focused on methodological rigor.Source: §5 · p. 7

5.1 Authentication and Workspace Initialization

• The authenticated workspace uses JWT for data isolation and provides a dashboard of ongoing analyses; continuous persistence lets users pause and resume without data loss.To start a project, the researcher defines its title.Source: §5.1 · pp. 7–8, Table 2 · p. 8

5.2 Configuration Stage: Corpus Constitution

• Corpus setup captures the research question, primary textual documents, and optional reference files, which RAG uses for context rather than treating them as primary data.The system validates these parameters and enforces linear progression to prevent backward changes.Source: §5.2 · p. 8

5.3 Pre-Analysis: LLM-Driven Index Generation

• In Pre-Analysis, Gemini receives the research question, corpus metadata, Bardin’s index-selection and indicator criteria, and document URIs through a RAG-based process.Researchers see suggested indices, descriptions, indicators, and exact source extracts with page numbers beside an integrated viewer for checking context.Source: §5.3 · p. 8, Figure 2 · p. 9

5.4 Human-in-the-Loop: Concept Appropriation

• Researchers may edit, remove, or add indices, but cannot alter extracted references or page numbers, keeping interpretation under human control and source evidence immutable..Source: §5.4 · p. 9

5.5 Material Exploration and Automated Categorization

• Material Exploration groups register units into categories using researcher-validated indices and frequency rules; charts, a category-by-index matrix, and expandable cards show frequencies and source-linked justifications.Final categories are non-editable because the system treats them as the systematic result of applying validated indices to the corpus.Source: §5.5 · pp. 9–10, Figure 3 · p. 10

5.6 Results Export and Methodological Traceability

• Cortex exports a PDF containing project metadata, the research question, frequency graphs, the categorization matrix, and category details with source excerpts and justifications.The report is intended to let readers and peer reviewers trace the path from corpus to conceptual inferences.Source: §5.6 · p. 10

6 Final Considerations

• Cortex automates index identification and counting while preserving human control over interpretative decisions and guiding researchers through the workflow.The development used .NET, React, and Vertex AI.Source: §6 · p. 10
• The authors report that Cortex successfully guides researchers through the content-analysis workflow..Source: §6 · p. 10
• The minimum viable product exports only PDF; planned extensions include other export formats, audio transcription, and advanced visualizations such as word clouds..Source: §6 · p. 10
• The project’s source code and repository are publicly available..Source: §6 · p. 10

A.2 Prompt: Exploration Stage

• The exploration prompt asks the model to identify register units containing supplied indices, group them into thematic categories, and count category frequencies under explicit criteria.It specifies distinct text segments, semantic fit, non-repetition in context, and literal meaning.Source: §A.2 · pp. 17–18
• The prompt requires use of supplied indices and indicators, exact excerpts, and justified classifications, while cautioning that categories and indices need not correspond one-to-one.Its fictional examples group technical difficulties, fatigue, and isolation into a broader category, and combine learning tools and schedule flexibility as gains.Source: §A.2 · pp. 17–20
• The prompt treats coding as rule-based transformation and identifies segmentation, enumeration, and classification or aggregation as coding choices, with category selection as the current task.It frames exploration as applying decisions established during pre-analysis.Source: §A.2 · pp. 20–21
• A recording unit is the coded unit of meaning, while a context unit supplies the larger segment needed to interpret it; unit selection should fit the material and analytic objectives.The prompt notes that context-unit size affects results and entails a trade-off between interpretive relevance and review effort.Source: §A.2 · pp. 21–22
• Simple frequency counts treat occurrences equally and rest on the assumption that repeated items become more significant, so the choice of measure should not be automatic.The appendix distinguishes the recording unit being counted from the rule used to count it.Source: §A.2 · p. 23
• Categories group elements by shared characteristics using a chosen criterion; categorization involves isolating elements and then classifying them, and different criteria can alter groupings.The appendix describes semantic, syntactic, lexical, and expressive criteria.Source: §A.2 · pp. 23–24
• The appendix illustrates categorization with prior studies, including value-analysis grids and categories used to analyze school speeches and educational values.These examples show that category sets vary with the study’s analytic focus.Source: §A.2 · pp. 24–26
• The school-speech example divides 1860–1965 into seven periods, with themes ranging from elite integration to a 1961–1965 “crisis of objectives.”The excerpt reports thematic analysis of changes in values advocated by the school system.Source: §A.2 · p. 26

• Indices should be pertinent, observable, manifest, and representative; indicators should be precise, reliable, systematic, and fit for purpose.Source: §A.1 · p. 12
• The prompt calls for theoretical contextualization, comparable coding units, and pre-testing indicators on document passages.Source: §A.1 · pp. 12, 15
• The requested JSON records indices with names and descriptions, precise indicators, and references to source documents and quoted excerpts.Source: §A.1 · p. 13
• Indices are observable traces suggesting phenomena of interest, while indicators specify how to measure or identify those traces to support inference.Source: §A.1 · pp. 12, 16
• An anxiety hypothesis can be operationalized through speech-disturbance indices, with their frequency serving as an indicator.The examples include interrupted phrases, repetition, stuttering, and incoherent sounds.Source: §A.1 · p. 16