Ecologies of the Ancient Mediterranean: Building an Evidence-Aware Spatial Graph RAG Platform for Environmental Knowledge in Texts and Landscapes (500 BCE–300 CE)

Large language models can now produce fluent “environmental histories” of the ancient  world on demand. For research, that fluency is also the core risk: LLM outputs routinely  collapse heterogeneous evidentiary layers, attested statements, inferred practices, and  

interpretive framings, into a single register of “facts.” In ancient environmental history,  where genres moralize scarcity, narrate disaster through conventional tropes, or encode  administrative action in formulaic documentary language, this collapse is  methodologically corrosive. This paper presents Ecologies of the Ancient  Mediterranean, a spatially aware AI platform designed to keep evidentiary distinctions  explicit while enabling map-based exploration and retrieval across textual and  archaeological datasets. 

The platform integrates three components that are often treated separately in digital  humanities workflows: (1) place-grounded text analysis, (2) knowledge-graph modelling  of evidence, and (3) retrieval-augmented generation (RAG) with auditable citations. The  core data layer combines openly licensed corpora (TEI/EpiDoc and linked-data exports  where available), a shared gazetteer backbone (Pleiades identifiers for places and  regions), and selected archaeological/landscape datasets relevant to environmental  practice (e.g., hydraulic features, settlement patterns, land-use evidence). Texts are  segmented into citable passages (“chunks”) enriched with source-type and genre  metadata (historiography, technical/agronomic writing, documentary texts), enabling  systematic comparison of how ecological phenomena are narrated and operationalized  across genres. 

A central design principle is evidence-aware annotation. The platform encodes  distinctions between: 

• attestations (what a text explicitly says, linked to exact passages), 

• inferred events/practices (what can be reasonably derived, with provenance),  and 

• interpretive framings (e.g., divine agency, mismanagement, natural cycles),  modelled as claims about discourse rather than facts about the world. 

To stabilize extraction and support interoperability, the project develops an Ancient  Ecology Vocabulary: a controlled vocabulary and light ontology for concepts such as  “water infrastructure,” “storage practice,” “hazard event,” and “scarcity rhetoric.”  Crucially, the vocabulary is designed as a reusable semantic layer: NLP labels point to  explicit definitions and identifiers, and (where feasible) the model is designed to be 

mappable to established cultural-heritage standards rather than remaining a one-off  taxonomy. 

Methodologically, the paper details a Spatial Graph-RAG architecture that combines  vector retrieval over text chunks with graph constraints and re-ranking by place, time  window, genre, and ecological category. A scholar-facing interface supports map-driven queries (e.g., “low-Nile years and repair activity”) while surfacing the underlying  evidence trail: retrieved passages, linked places, and the annotation steps that justify  any synthesis. Evaluation is built in through small, collaboratively produced gold standard annotations, assessed at the level of vocabulary concepts (precision/recall for  categories such as “canal” vs. “aqueduct,” or “famine episode” vs. “moralized scarcity  talk”). 

The paper concludes with a case study on agricultural risk and famine in Roman Italy as  a demonstration of how spatial linking and evidence separation can reveal patterned  divergences between discourse, institutional action, and material traces—without  forcing false alignment. For the Digital Humanities Congress, the contribution is both  methodological and infrastructural: a transferable pattern for building AI-assisted  research tools that remain transparent, auditable, and sustainable for humanities  scholarship.