Algorithmic Samgraha: Using LLMs to Reconstruct Missing Granthas

The Palimpsest of Silicon

To hold a fragment of the Gopatha Brahmana is to hold a temporal anchor. These palm-leaf manuscripts, dating from the medieval period but preserving Vedic recensions that stretch back millennia, are not merely texts; they are artifacts of rigorous oral tradition frozen in ink. Yet, like all organic materials, they are subject to decay. Water damage, insect erosion, and the simple entropy of time have left hundreds of sutras missing or illegible. For decades, scholars have faced a binary choice: leave the gaps empty, admitting ignorance, or attempt a manual conjecture based on etymology and contextual logic, a process fraught with the risk of imposing modern biases onto ancient syntax. The Bhandarkar Oriental Research Institute’s (BORI) recent collaboration with computational linguists introduces a third path: Algorithmic Samgraha, a methodology that leverages transformer-based Large Language Models (LLMs) to reconstruct these lacunae with a degree of probabilistic precision that challenges traditional epistemological boundaries.

From Oral Dhruva to Digital Noise

The stakes are inherently high. The Gopatha Brahmana is part of the Black Yajurveda corpus, a textual stratum critical for understanding the transition from ritual sacrifice to philosophical discourse. Unlike later literary compositions, the Brahmana texts are prescriptive and procedural. A single misplaced particle can alter the meaning of a ritual formula. The BORI team, working with open-source transformer architectures fine-tuned on parallel Sanskrit corpora (including the Rigveda and Atharvaveda where lexical overlaps exist), achieved an 84% accuracy rate in reconstructing syntactic logic within the missing sections. This metric does not imply that the model recovered the exact historical words with absolute certainty, but rather that the proposed reconstructions align with the grammatical architecture and thematic coherence of the surrounding text 84% of the time.

This success is rooted in the nature of Sanskrit itself. Indo-Aryan languages possess a high degree of agglutination and morphological regularity. Transformer models, which excel at capturing long-range dependencies and attention patterns, are uniquely suited to identifying the internal rhyming formulas and anaphoric structures common in Vedic prose. The model does not "halucinate" in the colloquial sense; it completes the pattern. When a sutra is broken mid-verse, the LLM analyzes the preceding mangalam (auspicious formula) and the subsequent topic shift to generate a bridge that satisfies both meter and meaning. In one case study involving a damaged passage regarding the placement of the Prastota, the model suggested a reconstruction that aligned with a previously undocumented variant found in a Thai temple archive, demonstrating its capacity to recover lost textual lineage.

The Epistemological Fracture

However, the computational victory raises a profound philosophical question: Who is the author of the reconstruction? In traditional Indological scholarship, a lacuna is filled by a human scholar who integrates philological critique (naidhisari) with historical context. The conclusion is attributed to the scholar’s reasoning. In Algorithmic Samgraha, the output is a probability distribution generated by neural weights trained on existing data. If the model reconstructs a sutra, is it 'restoring' the text or 'creating' a plausible fiction that looks like restoration?

This distinction is not merely academic; it strikes at the heart of Shrauta and Smarta authenticity. For practitioners of Sanātana Dharma, the Brahmana is not just literature but liturgy. If a reconstructed sutra is introduced into a digital archive without clear demarcation, it risks contaminating the textual stream. Future scholars, relying on these digital reconstructions, may treat the algorithm’s output as canonical fact. The boundary between scholarly restoration and artificial creation becomes blurred when the tool operates with such seamless fluency.

Toward a New Hermeneutic of Transparency

The BORI project proposes a new standard: Probabilistic Transparency. Every reconstructed segment in the digital edition must be accompanied by a metadata tag indicating the confidence score and the specific model parameters used. This allows future researchers to see not just 'what' was written, but 'how likely' it is to be original versus inferred. This approach mirrors the Vedic concept of Vak (the Word) evolving through stages—Guhya (hidden), Madhyama (middle), and Bahya (external). The algorithm exposes the Guhya layer of the text, revealing hidden structural dependencies, but it does so through a Bahya lens of statistical likelihood.

Critics argue that this method prioritizes logical consistency over historical fidelity. A perfectly grammatical sentence is not necessarily a historically accurate one. However, proponents counter that manual reconstruction is also subject to bias; human scholars unconsciously favor certain interpretations that align with their own theological or philosophical leanings. The algorithm, in theory, is 'bias-neutral' regarding theology, caring only for syntax. While this is not neutrality in the humanistic sense, it offers a new form of objectivity: the objectivity of the pattern itself.

The Future of Digital Dharma Studies

The implications of this project extend beyond the Gopatha Brahmana. As damage continues to affect manuscripts in archives across India, Southeast Asia, and Europe, Algorithmic Samgraha offers a scalable preservation tool. It is not a replacement for the shāstrāchārya (scripture scholar) but a powerful instrument in their hands. The challenge for the digital Dharma community is to establish ethical guidelines for its use. We must develop a vocabulary that distinguishes between algorithmic inference and authentic transmission.

As we integrate artificial intelligence into the study of the Vedas, we are reminded that the Rigveda itself spoke of Soma as a drink that unlocked wisdom, but also as a ritual fluid requiring proper context. LLMs are the digital Soma: a potent catalyst for insight, but dangerous if consumed without the discipline of tradition. The 84% accuracy is not a ceiling but a starting point. It invites us to ask not just how we can recover the lost words, but how we can preserve the integrity of the living tradition in an age where the scribe is no longer human, but a grid of silicon weights learning to speak the language of the ancients.