Introduction

From Digitization to Interpretive Accountability: Digital Humanities projects have dramatically expanded access to Buddhist materials—canonical texts, recorded teachings, and community records—yet access does not automatically yield interpretability. Buddhist knowledge is transmitted through lineage, pedagogical intent, ritual context, and communal authority, and these conditions are often lost when archives are treated as undifferentiated text. In this domain, the limitations of standard Retrieval-Augmented Generation (RAG) are not only technical but also hermeneutic: “naïve RAG” pipelines (generic splitting + dense retrieval) tend to encourage de-contextualized and modality-siloed answers, and can blur doctrinal plurality by over-generalizing distinct traditions.

We target two primary audiences: (i) public users seeking accessible and contextually grounded entry points, and (ii) expert practitioners requiring provenance-rich, citable, cross-media evidence for pedagogy and scholarship. These audiences are grounded in internal deployment use cases within a major Buddhist institution (Dharma Drum Mountain). This motivates a system that can switch mediation strategies conditioned on the query, rather than presuming one stable user or one “best” retrieval method.

A Three-Pillar Agentic Architecture

Motivated by the recurring failure modes of passage-centric naïve RAG, we identify breakdowns in contextual structure, constraint satisfaction, and pedagogical scaffolding. We propose a three-pillar architecture:

Corpus, Baselines, and Evaluation Design

We evaluate our system on a proprietary, structurally annotated Buddhist corpus curated by Dharma Drum Mountain. The corpus includes 110 books (4,063 chapters) from the Dharma Drum Complete Works of Master Sheng Yen. In addition, it contains 1,515 video lectures (~427 hours), 148 audio episodes (~114 hours), 2,051 short-form devotional texts, verified Q&A pairs, and a comprehensive set of upcoming events comprising 4,768 session instances (spanning August 2025 to July 2026).

After ingestion, the dataset yields 38,620 retrievable units (“chunks”), indexed under a polymorphic source ontology. Specifically, EventChunk represents events and venues; DeepContentChunk covers books, magazines, news articles, narrative prose, and automatically generated audio/video transcripts; and FAQChunk contains curated Q&A pairs.

We compare our system against a Naive RAG baseline reflecting common practice: uniform chunking with dense retrieval, without ontology-aware structuring or metadata filtering. Both systems share the same dense embedding model (BAAI/bge-m3, 1024-dimensional, multilingual) and cross-encoder reranker (BAAI/bge-reranker-large), allowing us to isolate the contribution of system architecture rather than model quality. Each experiment further includes ablations targeting individual architectural components.

We report four categories of evaluation metrics:

• Constraint satisfaction: Success@5 and SatFrac@5 for entity-bounded queries (e.g., temporal constraints).
• Standard IR metrics: hit@K and MRR under strict chunk-level matching, evaluated on a set of 1,100 LLM-synthesized queries (one gold chunk per query) across multiple ontological sources. We also include a multi-chunk extension (recall@K, coverage@K, lineage_coverage@K) designed to capture expert-level synthesis tasks that cannot be evaluated through single-chunk matching alone.
• Generation quality: Answer Relevancy, Faithfulness, Contextual Precision, and Contextual Recall, assessed by an LLM-as-a-judge on answers grounded in retrieved evidence.
• Cross-modal coverage: the distribution of retrieved evidence across different sources and modalities (e.g., book chapters, FAQs, narrative prose, audio/video transcripts), serving as a quantitative proxy for cross-media mediation.

Preliminary Results

We present preliminary findings spanning all three architectural pillars across four experiments — one per metric category.

• Temporal Constraint Satisfaction (Pillars I + II): On a benchmark of 100 time-bounded event queries, constructed from 10 temporal anchors (e.g., today, next week, this month, summer) and 10 activity categories (e.g., meditation, rituals, lectures), all grounded to a fixed reference date (May 4, 2026), the Naive RAG baseline (dense retrieval, top-5, no metadata filtering) achieves Success@5 = 20.0% and SatFrac@5 = 9.2%.

  In contrast, the Agentic Event Pipeline — where time-sensitive queries are routed to a specialized branch that applies structured *publish_date* filtering prior to ranking — achieves Success@5 = 90.0% and SatFrac@5 = 84.0%, corresponding to improvements of 4.5× and 9.1×, respectively.

  Notably, the structured temporal field required for this filtering is available only because the EventChunk ontology preserves it during ingestion (Pillar I), while the routing mechanism itself is enabled by query-conditioned agentic switching (Pillar II).

• Architectural Headroom (Pillar I — Standard IR):Across 1,100 LLM-synthesized queries with one gold chunk per case, evaluated under strict chunk-level Hit@5 matching, per-source scoped retrieval (perfect routing) achieves macro Hit@5 = 0.895; unified single-collection retrieval achieves 0.844; fan-out across 11 collections achieves 0.854. The 5-percentage-point gap between scoped and unified — visible only because the per-source ontology exists — quantifies the residual cost of cross-corpus mixing and motivates query-conditioned routing as the next architectural lever. Naive uniform chunking would not afford this comparison at all.
• Hybrid Lexical–Semantic Retrieval (Pillar II — Doctrine, Generation Quality): On 30 doctrinal queries from the Dharma Drum Complete Works corpus, hybrid retrieval combining dense bge-m3 with BM25 over jieba-tokenized text (fused via Reciprocal Rank Fusion) outperforms dense-only retrieval on all four LLM-judged generation-quality dimensions: Contextual Recall 1.000 vs 0.983 (+1.7pp), Answer Relevancy 0.968 vs 0.952 (+1.5pp), Faithfulness 0.956 vs 0.933 (+2.3pp), and most prominently Contextual Precision 0.895 vs 0.849 (+4.6pp). The largest gain is on Contextual Precision — exactly the dimension where BM25's exact-term matching aids rare Buddhist terminology (e.g., 般若, 波羅蜜多, 阿賴耶識) that dense embeddings tend to over-generalize. Latency overhead is negligible (+2ms p50).
• Cross-Modal Co-Presence (Pillar III — Cross-Modal Coverage): On a set of 30 doctrinal queries evaluated over the unified retrieval pool, the top-5 results consistently include evidence drawn from nine distinct corpora across four modalities. The distribution of retrieved content is as follows: book chapters (42.7%), FAQ entries (22.7%), narrative prose (12.0%), audio and video transcripts (14.7% combined), and news/magazine articles (5.3% combined). By contrast, a naive single-corpus baseline is limited to retrieving evidence from only one modality. These results indicate that the polymorphic ontology enables cross-modal pedagogical evidence within a single retrieval pass, thereby reducing reliance on upstream intent classification and improving the system's ability to support multi-perspective interpretation.

Conclusion

Buddhist knowledge systems require more than fluent language generation: they demand preservation of doctrinal plurality, contextual integrity, and source accountability. As our results show, these requirements are not adequately addressed by conventional, passage-centric RAG pipelines.

A key observation of this work is that such systems often *appear* to work. Fluent and plausible responses can create an illusion of correctness while masking failures in constraint satisfaction, contextual grounding, and provenance. This leads to epistemic overconfidence, especially for non-expert users. We therefore argue for a stricter boundary for conversational AI — evaluating systems not only by fluency, but by their ability to satisfy constraints, preserve context, and remain verifiably grounded in sources.

We address this gap through an agentic RAG framework that reframes retrieval as a query-conditioned mediation process. By combining ontology-driven representations, constraint-aware retrieval, and cross-modal orchestration, the system adapts to diverse interpretive needs — from novice guidance to expert-level, evidence-based inquiry. Our results show substantial gains in constraint satisfaction (4.5× Success@5, 9.1× SatFrac@5 vs naive baseline) and demonstrate hybrid lexical–semantic retrieval improvements on doctrinal queries (+4.6pp Contextual Precision), alongside coherent cross-modal evidence synthesis across nine corpora and four modalities.

More broadly, our findings suggest that RAG in interpretive domains must shift from static retrieval to adaptive, accountable mediation. Rather than replacing interpretive authority, such systems should serve as infrastructure for responsible engagement — balancing accessibility with rigor. Future work focuses on collaborative curation workflows, enabling both domain experts and real users to participate in certifying and endorsing system outputs. Such human-in-the-loop processes will further enrich the knowledge graph and help ensure alignment with evolving community consensus.