What RAG can and cannot do for a classical archive

1. The method, briefly

Retrieval-augmented generation (RAG) lets a language model search a fixed body of text before it answers (Lewis et al. 2020). In this Archive, each passage is embedded once and each question is embedded when asked. A dense vector search finds the nearest passages by cosine similarity, then gives them to the model as evidence. Those passages appear beside the answer with their canonical references (Att. 1.16, B.G. 5.44, Med. 4.23). The reader does not have to take the model's word for it.

More precisely, each passage p is mapped once to a vector e(p) by an embedding model, and stored. At ask-time the query q is mapped by the same model to e(q), and the system returns the k passages whose vectors are nearest e(q) by cosine similarity (a top-k dense retrieval; Karpukhin et al. 2020). Those k passages are concatenated into the model's context with an instruction to answer from them and cite them. Two design choices matter for what follows. The retrieval is dense-only: it matches on learned semantic similarity, not on shared words, which is a strength for paraphrase and a documented weakness for exact terms (§5). And k is small and fixed, so the system sees a handful of passages, never the whole work, which is why arguments that live across many passages fall outside its reach (§5). The lexical alternative, term-frequency scoring such as BM25 (Robertson & Zaragoza 2009), has the mirror-image profile, and the two are usually best combined (§5).

For a classical archive (Cicero's letters, Caesar's commentaries, Augustus's Res Gestae, Seneca's letters, Marcus Aurelius's Meditations, with Pliny and Quintilian to follow), the approach has real strengths and real limits, and it is easy to oversell both. The interesting bit is knowing which is which.

2. Why a classical corpus suits retrieval unusually well

A classical corpus is finite, edited, and stable: exactly the conditions under which retrieval is reliable. There is a definite set of texts; each has a canonical reference scheme refined over centuries precisely so that any passage can be pointed to unambiguously; and the words do not change between one query and the next. Retrieval systems built for the open web must cope with churn, spam, and boundless scale; the Roman Archive is a fixed, catalogued collection.

Against such a corpus, retrieval can find the passages that bear on a question (Cicero's mood before his exile, the honours Augustus records at the close of his life, how Marcus steadies himself against the fear of death) and hand them to the model as evidence. The model's task is then narrow and checkable: read these passages, and answer from them.

The canonical reference scheme does quiet but important work here. Because every retrieved unit carries its Att. or B.G. or R.G. number, the citation works as a join key between the answer and two millennia of scholarship. A reader who wants more than the Archive can give can walk Att. 1.16 straight into any commentary ever written.

3. Grounding makes error visible, not impossible

Grounding constrains the failure mode people rightly fear. A model asked about a Roman figure from parametric memory alone will sometimes produce fluent, confident, wrong statements: a plausible letter never written, a saying misattributed, a date that is almost right (Ji et al. 2023). Grounding the answer in retrieved text does not make hallucination impossible (the model can still misread, overstate, or blend its passages), but it makes error checkable: an answer with no supporting passage is a visible flag, and the question "is this sentence supported by its citation?" is one that both humans and automated methods can actually answer (Rashkin et al. 2023; Bohnet et al. 2022; Es et al. 2024).

The classical case adds a specific twist: the model's parametric memory of these authors is enormous. Any modern model has read more about Caesar than this Archive contains by him. Grounding therefore does different work here than in an obscure-domain RAG system: rather than supplying knowledge the model lacks, it fences the model off from knowledge it has. The discipline is subtractive. The failure the system must prevent is the silent substitution of the biographical tradition (Plutarch's Caesar, Shakespeare's, the internet's) for the primary text.

4. Scoping as a correctness property

Because the archive holds more than one author, retrieval is scoped by author, and this carries more weight than it first appears to. When you ask Marcus Aurelius a question, the search runs over his Meditations only; Cicero's letters are absent entirely, rather than merely ranked lower. A bounded figure never reasons from a text he could not have known, and the emperor of the second century does not borrow the words of a republican dead before he was born. The Classicist alone ranges across the whole collection, because that is the one voice standing outside it.

The Archive applies the same mechanism to a second boundary: primary versus reference. Background material (Smith's Dictionary of Greek and Roman Antiquities, Fowler's Social Life at Rome) is embedded but flagged, held out of retrieval by default, and marked [BACKGROUND] when the scholarly persona draws on it. The same principle applies to any corpus with internal provenance. Enforce the boundary during retrieval with a filter; do not leave it as a generation-time request. A model cannot leak a passage it never received.

5. What retrieval misses

Retrieval finds what sits lexically or semantically near a query. It can miss what a scholar would see:

• Arguments that live across documents. Cicero's slide from confidence to despair across the letters of 59–58 BC is real and documented, but it is a trajectory, present in no single letter. Top-k retrieval returns passages, not arcs. This is the classical case of multi-hop reasoning, where an answer requires chaining several passages that are individually only weakly related to the query and to each other; single-shot dense retrieval scores each passage against the query independently and so tends to surface the passages near the question while missing the bridging ones (the difficulty that motivated multi-hop benchmarks such as HotpotQA, Yang et al. 2018). Iterative or multi-step retrieval can partly recover such chains, but it is not implemented here, and no amount of top-k tuning substitutes for it.
• Meaning by omission. The Res Gestae never names Antony; the Commentarii pass in silence over what a legate's defeat cost politically. A silence cannot be embedded, retrieved, or cited, yet it is often the most historically telling thing about a text. RAG is good at "what does this author say about X" and weak at "what does it mean that this author says nothing about X".
• Irony and register. Whether Cicero is being sincere or performing for a correspondent is a judgement built from hundreds of letters and a century of scholarship, not from any retrievable span.
• Lexical blind spots. Dense retrieval can miss exact-match queries (rare proper names, offices, specific citations) that a keyword search would catch trivially; the pattern is well documented across retrieval benchmarks (Thakur et al. 2021). A corpus whose queries are full of names like Clodius, Vatinius, and Brundisium lives exactly in this regime. The standard remedies are hybrid fusion of dense retrieval with lexical BM25 scoring (Robertson & Zaragoza 2009), combined by reciprocal rank fusion (Cormack et al. 2009), plus rerankers and query expansion (HyDE; Gao et al. 2023); for this Archive they are planned measurements rather than delivered properties.

Retrieval surfaces evidence; it does not, on its own, interpret. That division of labour frames everything the system does.

6. Traps specific to this corpus

• Retrieval runs over translations. Several texts are English renderings of a Latin or Greek original, and the embedding space is the translator's English (Casaubon's 1634 Meditations above all). A query phrased in modern Stoic vocabulary may miss a passage the Greek would match, because the match is mediated by a seventeenth-century paraphrase; conversely the English may offer a match the original would not support. The companion essay on translation develops this fully.
• The texts assume a world. Roman names, offices, dates, and institutions form a dense web the query language may not share. "Praetor", "proscription", and "the Ides" carry loads a general reader's question won't articulate, which is one motivation for the Archive's gated reference layer.
• Each corpus is partial, by survival or by design. One side of a correspondence; a general's own account of his wars; an emperor's official self-summary. Absence of evidence in retrieval is not evidence of absence in history, and a well-behaved persona must say "my record does not speak of this" rather than let the gap read as fact.

7. The wider tooling, and what is not yet used

This Archive sits inside a mature digital-classics ecosystem it does not yet fully exploit. The texts come from the Perseus Digital Library's canonical corpora (Smith, Rydberg-Cox & Crane 2000). Latin-specific NLP now offers lemmatised search and Latin-aware embeddings: the Classical Language Toolkit (Johnson et al. 2021), Latin BERT (Bamman & Burns 2020), and LatinCy (Burns 2023), any of which could let retrieval see the Latin column rather than only the translators' English. Deep networks have been used to restore and attribute Greek inscriptions in genuinely philological work (Assael et al. 2022). This Archive does not add a retrieval technique to that body of work. It puts grounding, scoping and citations together in a reading interface where a general reader can question a classical corpus and check each answer.

8. Summary

RAG gives readers a direct way to question a classical archive and verify the answer as they go. Stable texts and canonical references make this corpus particularly well suited to it. Retrieval also keeps a fluent model away from the vast secondary tradition and ties it to the primary record.

It remains a reading aid. It can find passages and stay accountable to them. It cannot weigh a whole correspondence, notice the force of a silence or make the judgement that turns evidence into a reading. Happily, that part is still ours.

References

• Assael, Y. et al. (2022). Restoring and attributing ancient texts using deep neural networks. Nature 603, 280–283.
• Bamman, D. & Burns, P. J. (2020). Latin BERT: A Contextual Language Model for Classical Philology. arXiv:2009.10053.
• Bohnet, B. et al. (2022). Attributed Question Answering. arXiv:2212.08037.
• Burns, P. J. (2023). LatinCy: Synthetic Trained Pipelines for Latin NLP. arXiv:2305.04365.
• Cormack, G. V., Clarke, C. L. A. & Buettcher, S. (2009). Reciprocal Rank Fusion Outperforms Condorcet and Individual Rank Learning Methods. SIGIR 2009.
• Es, S., James, J., Espinosa-Anke, L. & Schockaert, S. (2024). RAGAS: Automated Evaluation of Retrieval Augmented Generation. EACL 2024: System Demonstrations.
• Gao, L., Ma, X., Lin, J. & Callan, J. (2023). Precise Zero-Shot Dense Retrieval without Relevance Labels (HyDE). ACL 2023.
• Ji, Z. et al. (2023). Survey of Hallucination in Natural Language Generation. ACM Computing Surveys 55(12).
• Johnson, K. P. et al. (2021). The Classical Language Toolkit: An NLP Framework for Pre-Modern Languages. ACL 2021: System Demonstrations.
• Karpukhin, V. et al. (2020). Dense Passage Retrieval for Open-Domain Question Answering. EMNLP 2020.
• Lewis, P. et al. (2020). Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. NeurIPS 33.
• Rashkin, H. et al. (2023). Measuring Attribution in Natural Language Generation Models. Computational Linguistics 49(4).
• Robertson, S. & Zaragoza, H. (2009). The Probabilistic Relevance Framework: BM25 and Beyond. Foundations and Trends in Information Retrieval 3(4), 333–389.
• Smith, D. N., Rydberg-Cox, J. A. & Crane, G. R. (2000). The Perseus Project: a Digital Library for the Humanities. Literary and Linguistic Computing 15(1), 15–25.
• Thakur, N. et al. (2021). BEIR: A Heterogeneous Benchmark for Zero-shot Evaluation of Information Retrieval Models. NeurIPS Datasets and Benchmarks.
• Yang, Z. et al. (2018). HotpotQA: A Dataset for Diverse, Explainable Multi-hop Question Answering. EMNLP 2018.