Purāṇa Nāma-Kośa

A scholar cannot currently ask the Purāṇas a question

Ask where Ambarīṣa appears across the corpus, or which Purāṇas know a particular tīrtha, or whether a lineage is told the same way in Vāyu and in Brahmāṇḍa, and there is no way to answer except by reading. The eighteen Mahāpurāṇas have translations, editions, and concordances of individual texts. They have no index of proper names spanning the corpus.

The consequence is quiet but severe. Comparative Paurāṇic study — the study of how the same figure, dynasty or sacred site is handled across texts — is presently done from memory and from whichever text a scholar happens to know well. Claims about Paurāṇic tradition circulate without a way to check their extension.

The absence is not of scholarship. It is of an apparatus that lets scholarship be checked.

Sorensen showed exactly what such an index is worth

Sören Sorensen's Index to the Names in the Mahābhārata (1904–1925) is the reason Mahābhārata scholarship can be argued rather than asserted. It carries 12,647 entries over 8,981 distinct names, with 4,861 aliases and epithets, each entry typed, glossed, and cited to the verse. A second, independent index — the Gītā Press Hindi nāmānukramaṇikā — covers 4,625 headwords in the same corpus and cross-checks the first.

Sorensen's own entries already carry the type as prose, which is why the category system below is his and not ours:

What Sorensen did for 83,966 verses over two decades of reading, no one has attempted for the 367,324 that lie beyond where he stopped — beginning, as section III shows, with the very next book of the edition he was reading.

The obstacle is reading time, and only reading time

The Purāṇa corpus is digitised and addressable. What has never been available is a way to put a trained reader in front of the right verses instead of all of them.

Those 12,098 Sanskrit topic labels matter more than their modest size suggests. They are the corpus naming its own contents in its own language, and they are where a Paurāṇic name list begins — not from an English category scheme imposed from outside.

The boundary is not even clean: the epic's own khila is unindexed

The Harivaṃśa is the khila — the supplement transmitted with the Mahābhārata, printed in the same editions, carrying the same commentary, and Paurāṇic in everything but attachment. At 16,099 verses it is the largest single book in the epic corpus, and it holds 4,034 distinct candidate names, a larger name inventory than any parvan of the epic proper.

Sorensen indexed parvans I through XVIII and stopped. Not one of his 12,647 entries carries a Harivaṃśa citation. The Gītā Press nāmānukramaṇikā stops in the same place — the Harivaṃśa is absent from it entirely. So the unindexed territory begins not at the edge of the Purāṇas but inside the epic's own covers, in the text that leads into them.

Machine-generated Sanskrit will not stand in for the Purāṇas

The obvious modern shortcut is to train a named-entity recogniser and run it over the corpus. The current state of the art shows precisely why that is not sufficient on its own.

Naamah (Kulkarni & Rajeev P, C-DAC Bangalore, 2026) is the largest Sanskrit NER corpus yet assembled: 102,942 sentences and 732,267 tokens, generated by prompting a large language model with entity seeds drawn from DBpedia. It reports 0.9615 F₁. It is careful, useful work, and its authors state its limits plainly:

• the corpus “inherits biases from source entities, prompting templates, and filtering heuristics”;
• it captures basic sandhi only, while “authentic Sanskrit literature is dominated by highly complex sandhi across multiple words”;
• it “lacks evaluation against manually annotated texts from classical sources” — there is no gold test set, so the reported score is measured on generated sentences averaging 7.11 tokens.

Nobody knows what such a model does on an anuṣṭubh of thirty-two syllables dense with compounds, because the measurement does not exist. Naamapadam, the standard Indic NER resource, does not cover Sanskrit at all.

The apparatus exists, and has been proved on the Mahābhārata first

Rather than begin on an unindexed corpus, the method was built and measured where a gold index already exists to check it against. Every figure below is from a completed run over the epic.

Two findings drive the proposal. First, Sorensen's category system is recoverable from his own prose — his entries open formulaically (a tīrtha, a river, a people, an ancient king, a ṛṣi, an asura), so the type inventory is inherited from a philologist rather than invented. Second, the scan is fast and the corpus is small in machine terms; the entire Purāṇa corpus was already scanned in minutes. Neither of these is the bottleneck.

Five stages, and the one that a machine cannot do

1. Inherit the type system automated

   Parse Sorensen's entry openings into an explicit category field, with Gītā Press senses and the existing typed gazetteer as cross-checks. Output: a typed, alias-linked lexicon of roughly 13,000 names carrying a philologist's own classification.

2. Harvest native Paurāṇic names automated

   Mine the 12,098 Sanskrit chapter topic labels for names the epic lexicon does not know. The Purāṇas name their own contents; this is where headwords absent from Mahābhārata scholarship enter.

3. Generate candidate attestations automated

   Scan the corpus for every occurrence, in Devanagari with sandhi-aware matching, recording verse address, surface form, word position and a confidence signal. Each candidate keeps its possible senses as an unresolved set. The machine proposes; it never decides which Citrasena.

4. Scholars read, and weed the substance of this proposal

   Candidates are presented verse by verse, in context, with translation and commentary alongside. A reader confirms, rejects, and adds what the machine missed; contested and structural cases escalate to a senior scholar who divides the senses and writes the entry. This is the step that turns a large noisy pool into a gold index, and there is no substitute for it. Every decision is recorded against its evidence and its author, so the index is auditable entry by entry. How this step is staffed is set out in section VIII.

5. Publish, and close the loop automated

   Verified entries become the index. The same verified spans become a gold annotated corpus, which trains a tagger, which proposes better candidates for the next Purāṇa. Scholar effort per text falls as the corpus grows.

What an entry looks like

The form follows Sorensen deliberately, so that the two indexes can be read together. Below, the parts a machine supplies and the parts only a scholar can supply are set apart — which is also the honest picture of where the work goes.

Senses 1–7 and the Ugrasenasuta cross-reference are Sorensen's own, quoted from the parsed index. The Harivaṃśa continuation is what does not yet exist and is what the pilot would produce. This headword is also a live illustration of the escalation rule in section VIII: seven claimants means no first-pass reader is asked to decide alone.

Deliverables

A team of young pāṭhaśālā scholars, with seniority spent where it counts

The reading in stage IV is not one task. Deciding whether a token in a verse is an occurrence of a given name is a fast, well-defined judgement that a competent Sanskrit reader makes correctly and quickly. Deciding that a name carries two distinct senses, writing the gloss that separates them, and reconciling the result with Sorensen is a different order of work. Conflating the two is what has always made an index of this size look unaffordable.

Escalation is automatic, never discretionary

The known failure of tiered annotation is a junior annotator not recognising a hard case. So the hard cases are identified by the system, not by self-assessment. A candidate leaves the first pass and goes to the senior scholar whenever:

• two annotators disagree
• the lexicon offers more than one claimant
• the headword is unknown to the epic index
• it is the first attestation of a new headword
• a sense division is proposed
• machine confidence falls below threshold

How the quality claim is made good

• Overlap and agreement. A rolling 10% of candidates is adjudicated independently by two readers; inter-annotator agreement is computed continuously and published with the data. A gold set without a reported agreement figure is an assertion, not a measurement.
• Calibration before production. Each reader begins on a set already adjudicated by the senior scholar and does not move to live work until agreement clears an agreed threshold.
• Entry-level sign-off is absolute. No entry is published without the senior scholar's approval, whatever the first-pass volume beneath it.
• Reversibility. Because every decision is stored with its evidence and its author, a later correction repairs the affected entries without rebuilding anything.

This is capacity building, not data entry

A student who spends a season adjudicating Paurāṇic attestations has read a great deal of Purāṇa closely, with commentary and translation at hand, under a senior scholar's correction — and emerges with named credit in a published scholarly apparatus. That is a route from pāṭhaśālā training into research authorship that presently barely exists, and the project produces it as a matter of course rather than as an add-on.

Ādi Parva and Harivaṃśa — a controlled pair, not merely a convenient one

The pilot takes two adjacent books of the same corpus, in the same edition, the same addressing and the same commentarial tradition. One is the most thoroughly indexed book in Sanskrit literature. The other has never been indexed at all.

Why this pair and no other

Every quality claim about an index of an unindexed text runs into the same wall: there is nothing to check it against. This pair removes that wall. Phase A is scored against two independent gold indexes — reader accuracy, machine recall, escalation rate and inter-annotator agreement are all measured against Sorensen and Gītā Press, on real verses, by the same team that will then work unsupervised. Phase B carries those measured figures across a seam narrower than any other in Sanskrit literature: the next book of the same printed edition, in the same transmission, with the same commentator alongside.

No pair of Purāṇas can offer that, because no Purāṇa has a gold index to calibrate against. Beginning anywhere else means asserting an error rate; beginning here means measuring one and then carrying it.

The Harivaṃśa earns its place on its own account as well. It is vaṃśa — genealogy is its subject, which is why it holds the corpus's richest name inventory. And its Kṛṣṇa material is the direct textual ancestor of Viṣṇu Purāṇa V and Bhāgavata X, so the pilot's headwords reach into the Purāṇa corpus proper from the day they are published.

Effort, and where it actually goes

• Calibration before any estimate is fixed. The senior scholar adjudicates 200 Ādi verses first. That set establishes spans per verse, minutes per span and the true escalation rate — and doubles as the training set every reader is measured against before going live.
• First-pass reading of 36,546 spans, plus a 10% double-adjudicated overlap: on the order of 500 reader-hours. A cohort of five, working part-time around their studies, covers that across a season.
• Senior time divides unevenly between the phases, and this is the finding that shapes the budget. In Phase A most entries already exist, so the senior work is verification against Sorensen — comparatively cheap. In Phase B every entry is authored from nothing, and entry authoring, not escalation, is the dominant senior cost. Harivaṃśa's 4,034 candidate headwords are an upper bound; how many survive weeding is itself one of the pilot's findings.
• Decision point. The pilot publishes its own cost curve — observed escalation rate, agreement figures, and senior hours per hundred entries. Extension to the eighteen Purāṇas is proposed only on evidence that the loop in stage V actually reduces per-text effort.

What a verified index makes possible that nothing else does

• Comparative Paurāṇic study becomes checkable. Which Purāṇas know a given ṛṣi, dynasty or tīrtha, and in what proportion, becomes a query rather than a recollection.
• Sacred geography gains an evidentiary base. A typed gazetteer of Paurāṇic tīrthas, rivers and mountains with verse citations serves temple history, pilgrimage studies and historical geography directly.
• Epic and Purāṇa can be read against each other. With both corpora indexed on a shared name inventory, the relation between Mahābhārata and Paurāṇic treatments of the same figure is finally open to systematic study.
• Sanskrit language technology gets a real benchmark. The gold corpus is reusable well beyond this project, and its absence is currently a stated limitation in the published literature.
• The method transfers, and so do the people. Nothing in the apparatus is specific to the Purāṇas; the same five stages apply to any large unindexed Sanskrit corpus. By the end of the pilot the harder asset is a trained cohort who have done this work once and can lead the next text.

Scholarly reading time, and the means to support it

The computational work is largely done and has been proved against a gold index. What this proposal requests is the one thing that cannot be automated and has never been resourced: sustained, trained, attributed reading — and, with it, the chance to bring young scholars of the pāṭhaśālā tradition into published research on the strength of the training they already have.

• Senior scholarly investigator with Paurāṇic and Vedic textual training, for the pilot's duration — authoring entries, ruling on escalations, signing off on release.
• Stipends for a first-pass cohort of four to six young pāṭhaśālā scholars, engaged part-time around their existing study, each credited by name in the published apparatus.
• Onboarding and calibration: the senior scholar's 200-verse training set, and the supervised period each reader works through before live adjudication.
• Technical work on the extraction pipeline and on an adjudication interface that is multilingual, phone-first and resumable by design.
• External review of the pilot output by an independent Paurāṇic scholar before release.

The index is not the end of the work. It is the instrument that lets a century of subsequent work be done, and checked.