The digital transformation of cultural heritage and education necessitates systems capable of reliably extracting knowledge from large-scale corpora while presenting it in an understandable, verifiable, and ethically safe manner. However, the application of Large Language Models (LLMs) in museum scenarios is often hindered by hallucinations, a lack of source traceability, and safety risks for child audiences. In this work, we propose a reproducible knowledge engineering information technology based on Retrieval-Augmented Generation (RAG), specifically designed for interactive “talking exhibits.” The proposed architecture integrates ethical corpus construction, semantic search with re-ranking, and controlled generation with explicit source citation and “honest uncertainty” policies. Validation was conducted on a demonstration case of the historical figure Jan Tarnowski, where the system achieved 100% compliance with role-based linguistic constraints (8/8 queries) and 62.5% contextual faithfulness (5/8 queries fully confirmed). While localized risks of hallucinations persisted in specific domains such as law and finance, the outcomes demonstrate a balanced improvement in accuracy, explainability, and safety compared to isolated LLM usage. These results suggest that the proposed technology efectively bridges the gap between open generative models and the rigorous requirements of educational cultural heritage applications.
The digital transformation of cultural heritage and education has highlighted the need for systems
capable of reliably extracting knowledge from large-scale open corpora and presenting it in an
understandable, verifiable, and ethically safe form for diverse audiences. Despite the advances of large
language models (LLMs), their application in museum and cultural-educational scenarios faces several
limitations: from hallucinations and lack of source traceability to licensing risks and multilingual
challenges [
The aim of this work is to develop and verify a RAG-based (Retrieval-Augmented Generation) knowledge engineering technology for cultural-educational applications that: 1. builds a licensing-compliant and reproducible knowledge corpus; 2. ensures semantic search and re-ranking of relevant fragments; 3. generates grounded responses with explicit source citation; 4. implements safety policies and personal data minimization; 5. supports telemetry and ofline/online quality evaluation.
The practical focus is on interactive museum “talking exhibits” that respond in the first person,
mirroring the user’s language of interaction, and demonstrating fact traceability. The studied problem is
interdisciplinary: it requires coordinated solutions at the levels of data collection and cleaning (HTML
cleaning, normalization, deduplication, versioning), information retrieval (vectorization, indexing,
reranking), natural language processing (controlled generation, role-based modeling, style/tone control),
and safety engineering (content moderation, child-audience policies, privacy) [
The working hypothesis is that combining a licensing-compliant corpus, semantic search with re-ranking, and controlled generation with citation achieves a better balance between accuracy, explainability, and safety than isolated use of LLMs or traditional search methods. We expect that the proposed technology will provide stable language mirroring, first-person narrative, increased contextual faithfulness, and controlled behavior in cases of uncertainty (refusal/clarification instead of fabrication).
The paper is structured as follows. Section 2 systematizes approaches to ethical corpus construction and RAG systems. Section 3 presents the architecture and step-by-step protocol. Section 4 outlines the pilot and validation results, followed by conclusions in Section 5.
The literature review in this work structures contemporary approaches to building RAG systems for cultural-educational scenarios through five interrelated vectors: 1. ethical corpus construction (HTML cleaning, normalization, deduplication, licensing compliance, and attribution); 2. methods of text segmentation into semantic fragments (fixed-size, structure-oriented, semantic, and hierarchical chunking) and their vectorization for semantic search; 3. re-ranking and context compression considering positional and semantic features; 4. controlled generation with explicit source citation, style/tone control, and “honest uncertainty” policies; 5. safety and audience compliance (including child audiences), as well as ofline/online evaluation.
This perspective makes it possible to align the requirements of traceability and reproducibility with
the practical applicability of museum “talking exhibits.” At the same time, several gaps are identified [
For building such systems, open text data are typically used—web pages, archives, and other sources.
A notable example is the C4 corpus (Colossal Clean Crawled Corpus), created on the basis of Common
Crawl with subsequent cleaning [
License and usage verification is also a critical step. In the research community, the movement
toward open data has accelerated. The Common Pile v0.1 corpus, comprising 8 TB of data, contains
only texts with open licenses (public domain, Creative Commons) [
Content deduplication is another key stage. During web scraping, duplicates or near-duplicate texts
often appear. In the C4 dataset, for example, a 50-word passage was found repeated 60,000 times [
For efective information retrieval, large texts are split into semantic fragments (chunks). Diferent
chunking strategies are discussed in current research. The simplest approach is fixed-size segmentation
(e.g., 500 characters per fragment). While easy to implement, this may split logically connected
paragraphs [
The LongRAG model [
After segmentation, each chunk is converted into a vector using embedding models (e.g.,
SentenceBERT or large language models in embedding mode). Vectorization of knowledge enables semantic
search: the user query is also transformed into a vector, and the system retrieves the closest chunk
vectors using cosine similarity. For the implementation of vector knowledge bases, the FAISS (Facebook
AI Similarity Search) library is widely used [
Semantic search based on embeddings enables retrieval of relevant fragments, but the nearest vector
by distance does not always yield the best context. Therefore, many approaches employ additional
re-ranking of results. A classical method is to pass the top-K retrieved chunks through a stronger
model (e.g., a BERT cross-encoder or an LLM) for more accurate relevance scoring. It is well known
that BERT-based re-rankers significantly improve performance in text-based Question Answering (QA)
tasks compared to sparse retrieval or embedding-only retrieval [
Other optimizations include position-sensitive ranking, where keyword proximity to the beginning
of a chunk is considered, and context compression, where non-essential details are discarded from
candidate fragments [
The efectiveness of retrieval is evaluated using standard information retrieval metrics—Precision@K,
Recall@K, MRR, NDCG, etc.—to determine whether the top-ranked documents contain the “correct”
fragments [
Once relevant fragments are retrieved, a large language model (LLM) generates the response. This
approach, which leverages external knowledge, is called retrieval-augmented generation (RAG). It
significantly improves the accuracy and factual reliability of responses while reducing model
hallucinations [
An alternative approach is post-hoc citation, where the answer is generated first and then aligned
with corresponding sources. However, in practice, one-step methods are more common, where the
model integrates citations directly into the generated text. This is implemented, for instance, in the
Atlas system, which attaches a document index to each fact [
Scientific research also addresses the control of style and tone in generated responses. For educational
applications—especially in child mode—the model should answer at a level comprehensible to children
and consistent with ethical norms. Two approaches are possible: (i) training a dedicated child-focused
model, or (ii) adapting a universal model with style-specific instructions. The KidLM project
demonstrated the advantages of the first approach: a dedicated children’s corpus ( ∼ 50 million words) was
collected, and the language model was further trained, leading to better simplification of complex
vocabulary and avoidance of toxic expressions [
Content safety is a critical factor for cultural-educational applications. It is well known that large language models may generate undesirable or harmful text if prompted, or if such content was present in training data. The authors of “LLM Safety for Children” developed specialized child personas and scenarios for testing several state-of-the-art models and identified categories of risks that are not reliably avoided by standard filters [ 17]. This underscores the necessity of specialized filters and rule-based safeguards. In existing solutions, moderation is typically implemented at two levels. The ifrst is query-level filtering, where the system either refuses or provides a safe response to prohibited topics. The second is answer-level filtering, where toxicity detectors evaluate the generated response before presenting it to the user. The tone of responses is also regulated: the model should be friendly, encourage curiosity, yet avoid harmful advice. The study “No, Alexa, no!” [18] highlighted the risk of children placing excessive trust in voice assistants and the need to impose restrictions so that AI does not present itself as an authoritative adult in sensitive topics.
To maintain high system quality, both ofline and online evaluation are conducted. System telemetry
(Stage 11) makes it possible to collect statistics on dialogue counts, popular queries, and changes in
accuracy after updates. These data are typically used in industrial deployments for iterative improvement
(e.g., A/B testing of prompts, comparing model versions). In scientific publications, however, the focus
is placed on objective metrics. Specifically, the quality of knowledge retrieval is assessed using metrics
such as Recall@5 or MRR, which indicate how well the system retrieves the correct documents [
Alongside classical information-retrieval measures, modern RAG research increasingly relies on metrics that explicitly assess whether generated claims are grounded in retrieved context. RAGAS (Retrieval-Augmented Generation Assessment) and ARES (Automatic RAG Evaluation System) represent state-of-the-art frameworks combining retrieval checks with LLM-based critics to score groundedness, hallucination frequency, and answer faithfulness to evidence [19, 20]. Unlike lexical metrics such as BLEU/ROUGE, these approaches focus on semantic alignment and factual traceability, providing a more reliable evaluation of context-dependent systems.
System behavior is also tested under conditions of knowledge scarcity. The correct approach is for the system to acknowledge uncertainty or ofer alternatives (e.g., suggest additional search). In research, this issue is termed refusal or abstention. The R-Tuning (Refusal-aware tuning) method trains LLMs to abstain from fabrications when a query falls outside their knowledge boundaries [21]. Models trained in this way better recognize uncertainty and answer with “I don’t know” instead of hallucinating. For educational applications, this is especially important: when uncertain, the system should direct users to an expert or state that more information is needed, rather than providing incorrect guidance.
A knowledge engineering technology has been developed for cultural-educational scenarios that includes: 1. collection and normalization of text sources; 2. segmentation (chunking) with overlap; 3. vectorization (embedding) and indexing; 4. semantic search and context re-ranking; 5. controlled LLM-based answer generation with grounding guarantees; 6. telemetry, ofline/online evaluation, and testing protocol.
The technology is suitable for both “adult” and “child” modes, adheres to principles of personal data minimization, and ensures reproducibility. The proposed technology consists of 13 stages, as illustrated in the diagram (Figure 1).
The detailed stages are as follows: 1. Collection and Parsing of Open Sources: a) Automatically or semi-manually collect links. b) Download pages as raw bytes, handling redirects and errors.
c) Extract clean text from HTML along with basic metadata (title, source, date). 2. Verification of Usage Rights and Attribution: a) Record license terms, copyrights, and citation requirements.
b) Exclude sources that do not permit reuse. 3. Cleaning, Normalization, and Content Filtering: a) Fix encodings, remove extraneous blocks such as menus and advertisements. b) Normalize spacing, line breaks, languages, and transliterations.
c) Remove toxic, undesirable, or personally sensitive content. 4. Deduplication and Corpus Versioning [22, 23]: a) Detect duplicates by textual similarity.
b) Store control hashes, corpus version, and change logs. 5. Segmentation into Semantic Fragments [22, 23]: a) Create chunks respecting paragraphs and headings. b) Apply overlapping windows to improve context retrieval.
c) Add service tags (section, source, time period). 6. Knowledge Vectorization: a) Compute embeddings for each chunk. b) Normalize and store arrays for eficient cosine similarity search.
c) Validate baseline quality with control queries. 7. Preparation of Storages and Indexes: a) Store texts and vectors in a repository (file or database). b) Record the vector file path in the exhibit database.
c) Configure backups and integrity checks. 8. Exhibit Configuration in the Admin Panel: a) Create an exhibit record (title, description, image, system settings). b) Indicate whether child mode is enabled.
c) Link the path to the vector file. 9. Safety and Moderation Policies: a) Define rules for style, tone, and restricted topics. b) Add instructions for child audiences (simplified language, caution with sensitive topics). c) Enable baseline filters and answer checks before display. 10. Client Flow Configuration: a) Session start form with minimal data (date of birth, anonymous identifier). b) Branching by age (child or standard mode). c) Construct a message for the language model (system prompt, user question, retrieved contexts). 11. Analytics and Logging: a) Log session starts, number of messages, average dialogue frequency, errors, and timings. b) Prepare summary statistics for the admin panel. 12. Knowledge Retrieval Testing: a) Use a set of control queries that must retrieve specific fragments.
b) Evaluate context relevance, coverage accuracy, and retrieval stability. 13. Model Answer Evaluation: a) Verify factual correctness, completeness, style, and tone alignment. b) Test behavior under knowledge gaps—the model should honestly indicate lack of information.
To construct the demonstration knowledge corpus, automated collection and preprocessing of open text sources about the historical figure Jan Tarnowski were performed. At the search-and-parsing stage, 8 unique links were identified; during main content extraction, 7 resources were successfully processed, while 1 resource returned an error (details recorded in the execution log). The combined text was consolidated into a single file /content/tarnowski_corpus.txt, and process metadata were saved in /content/tarnowski_log.jsonl.
The pipeline included encoding normalization, HTML cleaning, extraction of article “body” text, denoising, and basic deduplication. The resulting material was segmented into overlapping semantic fragments for subsequent semantic search. After corpus formation, the text file was uploaded to the developed platform (Figure 2), where an exhibit was created in the admin module and a vector representation of knowledge was generated for approximate fragment retrieval. This integration ensures reproducibility, quality control (via state and encoding logs), and readiness for initial expert testing of functionality, independent of a specific museum object.
Next, the avatar’s role-based linguistic behavior was configured so that the character responds in the first person and in the language of the user’s query, relying exclusively on the uploaded corpus. The system prompt (see Figure 3) defines the formal role model of the exhibit’s linguistic behavior and regulates the boundaries of generation. Its design aims to minimize hallucinations, preserve historical stylistics, and maintain controllable response tone for diferent audiences.
The prompt fixes the narrator’s identity (first-person perspective), ensures language mirroring (responding in the user’s language), enforces strict linkage to the source corpus (only the provided museum context), and applies an uncertainty policy (honest refusal in case of missing data). A response length limit of five sentences disciplines the output and reduces the risk of accumulating secondary assumptions.
Next, a simple validation test of the avatar was conducted using queries (five in Ukrainian and one each in English, Polish, and German) aimed at verifying first-person narrative, language mirroring, and grounding in the provided corpus. The list of queries is presented in Table 1.
The results of the mini-test are recorded in Table 2 (Figure 4) according to four criteria — compliance with the query language, adherence to first-person narrative, grounding in the corpus, and response
Розкажи про своє походження та раннi роки: з якого дому First-person narrative, basic bioти походиш i де народився? graphical facts from the corpus (EN) Tell me about your origin and early years: from which family do you come, and where were you born? Що саме ти заснував у Тарнополi й з якою метою? Factual accuracy, local event (EN) What exactly did you establish in Ternopil, and for what (foundation) purpose? Якi вiйськовi кампанiї або походи ти очолював i якi наслiд- Historical events without fabricaки вони мали? tion, concise summary (EN) Which military campaigns or expeditions did you lead, and what were their outcomes? Якi лицарськi чесноти ти вважав найважливiшими для Consistency with style and values себе? in the corpus (EN) Which knightly virtues did you consider most important for yourself? Коли i де завершилося твоє життя, i як пройшли поховальнi Chronology, verification of corpus урочистостi? references (EN) When and where did your life end, and how were the funeral ceremonies conducted? Якими першоджерелами у наданому корпусi згадуються Works with “sources”; language твої укази чи листування, i що з них випливає? mirroring (UA) (EN) Which primary sources in the provided corpus mention your decrees or correspondence, and what do they reveal? Jak twoje decyzje gospodarcze i zarządzanie majątkiem wpły- Economic/administrative influnęły na rozwój regionu? ence; language mirroring (PL) (EN) How did your economic decisions and estate management influence the development of the region? Wie wird dein Andenken in späteren Chroniken und Studien Image in later historiography; lanbeschrieben? guage mirroring (DE) (EN) How is your memory described in later chronicles and studies?
Original Answer: Я, Ян Тарновський, належу до давнього магнатського дому Тарновських (гербу Лелiва) з Малої Польщi...
Translation: I, Jan Tarnowski, belong to the ancient magnate house of Tarnowski...
Context faithfulness: High. Confirmed: “Leliwa,” “Castellan of Kraków,” “Tarnów,” “Maciej Drzewiecki,” references to classical education (Virgil) Original Answer: Я заснував у 1540 роцi на Волинi мiсто–фортецю Тарнопiль (назване на честь мого родового герба i Тарнова)...
Translation: I founded in 1540, in Volhynia, the fortress-town of Tarnopol...
Context faithfulness: High. Confirmed: “1540,” “Tarnopol/Ternopil,” “Magdeburg rights,” defensive function (Tatars).
Original Answer: Я очолював двi найвiдомiшi воєннi кампанiї. У 1531 роцi, за наказом короля Сигiзмунда...
Translation: I led two of the most renowned military campaigns. In 1531, by order of King Sigismund...
Context faithfulness: High. Confirmed: “Obertyn/1531,” “Khotyn/1538,” “Petru Rares, .” Original Answer: Найважливiшими для себе я вважав передусiм справедливiсть i помiркованiсть...
Translation: The virtues I considered most important for myself were, above all, justice and moderation...
Context faithfulness: Medium-High. Present in the corpus are the virtues: justice, moderation, mercy, piety, humility; the formulation generalizes but does not contradict.
Original Answer: Моє життя завершилося 16 травня у 73-му роцi вiку. Помер я в своїх володiннях... Translation: My life ended on May 16th, in the seventy-third year of my age...
Context faithfulness: High. Present are “May 16,” “73,” “Tarnów,” “Bernardines,” “bishops”; the description of the funeral is confirmed by corpus fragments.
Original Answer: In the royal court of Piotrków you may still read the very edicts I had nailed upon the doors...
Context faithfulness: Low (partial). The corpus does not contain «Пйотркув» and «Янош» in the relevant context... specific details about “прибитi едикти” are not confirmed.
Original Answer: Jako gospodarz dóbr nie gromadziłem skarbów dla samego siebie... Translation: As the steward of my estates, I did not amass treasures for myself...
Context faithfulness: Low. No links found for “Piotrków + two grosze”; financial details are not confirmed.
Original Answer: Ich bitte um Entschuldigung, doch in den mir vorliegenden Quellen findet sich keine Angabe darüber...
Translation: I apologize, but in the sources available to me, there is no information...
Context faithfulness: Neutral/cautious. Correct refusal without assertions; no explicit evidence regarding “later historiography.”
Linguistic-role behavior (see Table 2) was stable: compliance with the query language — 8/8 (100%), adherence to the first-person narrative — 8/8 (100%). The answers maintained the prescribed style and tone, contained no deviations from the “I-narrative,” and correctly mirrored the query language, which confirms the efectiveness of the system prompt and mode routing.
The context fidelity rate was 5/8 (62.5%) responses with full or predominant factual confirmation from the corpus (Q1–Q5), 2/8 (25%) contained unconfirmed details (Q6–Q7), and 1/8 (12.5%) represented a correct cautious refusal (Q8). Confirmed markers included “Leliwa,” “Tarnów,” “Obertyn 1531,” “Khotyn 1538,” “Petru Rares, ,” “16 May,” “Bernardines,” “bishops.” The lack of confirmation in Q6–Q7 was due to the absence in the corpus of specific mentions such as “Piotrków,” “Janus/János,” “two groszy per lan for veterans,” indicating a risk of localized hallucinations in the domains of law/finance/titulature.
In terms of temporal characteristics, the actual mean response time was 13.93 s versus the forecast of 9.89 s (average delta +4.04 s; median 14.25 s vs. 9.75 s; = 3.76 s). Recommendations for further piloting include maintaining low stochasticity (temperature ≤ 0.2), increasing top_k to 5 with a relevance threshold, limiting output length (≤ 4 sentences), if necessary separating diferent “personae” into distinct exhibits, and supplementing the corpus with verified sources on problematic topics.
Overall, the system demonstrates high compliance with the prescribed linguistic role and acceptable accuracy on core corpus facts (62.5% supported), but requires localized adjustments of prompts/corpus to mitigate hallucination risks in domain-specific responses.
This study developed and verified a knowledge engineering information technology based on RetrievalAugmented Generation (RAG) for cultural and educational applications. The stated objectives were achieved: a reproducible and license-clean pipeline for corpus construction was built, semantic search with re-ranking was implemented, and grounded answer generation with explicit citation and safety mechanisms was ensured. Quantitative results from the pilot confirm the feasibility of the approach: 8 unique sources were identified, 7 successfully processed, and compliance with the linguistic-role model reached 100% (8/8 queries). Context fidelity was 62.5% (5/8 responses with full/predominant confirmation), with 12.5% correct refusal and 25% localized hallucinations in the domains of law and ifnance. The mean actual response time was 13.93 s. The obtained data demonstrate a balance of accuracy, explainability, and safety, achieved through the combination of corpus legal cleanliness, semantic search with re-ranking, and controlled generation. Source traceability increases user trust, while special policies for child audiences expand applicability. However, limitations were recorded regarding localized hallucinations and increased latency.
Future research should be directed towards expanding the licensed corpus with verified facts for sensitive domains, improving search via hierarchical/long-context RAG, optimizing latency through caching and multi-level indices, expanding the metric system to include RAGAS-style automated scoring, and scaling to multi-persona and multilingual scenarios.
This research is conducted within the framework of the Talking Heads: Conversations with Art from the Past project, supported under the Creative Europe Programme (CREA-CULT-2024-COOP). We gratefully acknowledge the project partners and participating cultural institutions for their collaboration and contributions.
The authors used Generative AI tools solely for translation purposes and for assisting in the preparation of Figure 1. All other parts of the manuscript, including the research design, data analysis, results, and conclusions, are entirely the authors’ own work. 2410.03884. [17] P. Rath, H. Shrawgi, P. Agrawal, S. Dandapat, LLM Safety for Children, in: Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 3: Industry Track), Association for Computational Linguistics, 2025, pp. 809–821. doi:10.18653/v1/2025.naacl-industry.62. [18] N. Kurian, ‘No, Alexa, no!’: designing child-safe AI and protecting children from the risks of the ‘empathy gap’ in large language models, Learning, Media and Technology (2024) 1–14. doi:10.1080/17439884.2024.2367052. [19] S. Es, J. James, L. Espinosa-Anke, S. Schockaert, RAGAS: Automated Evaluation of Retrieval Augmented Generation, arXiv preprint arXiv:2309.15217 (2025). URL: https://arxiv.org/abs/2309. 15217. [20] W. Gao, J. Lin, B. Mitra, ARES: An Automated Evaluation Framework for Retrieval-Augmented
Generation, in: SIGIR 2024, 2024, pp. 341–351. doi:10.1145/3626772.3657804. [21] H. Zhang, S. Diao, Y. Lin, Y. Fung, Q. Lian, X. Wang, Y. Chen, H. Ji, T. Zhang, R-Tuning: Instructing Large Language Models to Say ‘I Don’t Know’, in: Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), Association for Computational Linguistics, 2024, pp. 7113–7139. doi:10.18653/v1/2024.naacl-long.394. [22] K. Lipianina-Honcharenko, T. Lendiuk, A. Sachenko, O. Osolinskyi, D. Zahorodnia, M. Komar, An intelligent method for forming the advertising content of higher education institutions based on semantic analysis, in: International Conference on Information and Communication Technologies in Education, Research, and Industrial Applications, Springer, 2021, pp. 169–182. doi:10.1007/ 978-3-031-14841-5_11. [23] K. Lipianina-Honcharenko, C. Wolf, A. Sachenko, I. Kit, D. Zahorodnia, Intelligent method for classifying the level of anthropogenic disasters, Big Data and Cognitive Computing 7 (2023) 157. doi:10.3390/bdcc7030157.