Where this comes from, and what it is allowed to claim.

What is in this corpus

Four Samhitas, one recension each. Certified invariant core holds 20,210 canonical mantras (release commit: 50a4042). No separate Brahmana, Aranyaka, or Upanisad corpus is included. The Samaveda's ARANYA section is a structural division of the modeled Kauthuma Arcika, not an independent Aranyaka corpus. Three of the four are missing a body of material that their ordinary name covers, and in one case that missing body is larger than what is held.

RigvedaShakala

10,552 verses. 10,480 dedicated English translations, 60 covered by a multi-verse range, 6 carrying a non-English scholarly note, 6 with none. 10,552 recitation records released. Not addressed: second rigvedic recension; rigvedic brahmana; rigvedic aranyaka; upanisad.

SamavedaKauthuma

1,844 verses. No English translation of its own, 173 covered by a verified reused parallel rendering, 1,671 with none. 1,136 verses carry a validated source-explicit notation witness and 708 await alignment. No recitation released; the queue is still in review. Not addressed: samaveda gramageya gana; samaveda aranyakageya gana; samaveda uhagana; samaveda uhyagana; second samavedic recension; samavedic brahmana; upanisad.

YajurvedaMadhyandina

1,975 verses. 1,950 dedicated English translations, 25 with none. 1,777 recitation records released. Not addressed: krishna yajurveda taittiriya; krishna yajurveda kathaka; krishna yajurveda maitrayani; krishna yajurveda kapisthala; shukla yajurveda kanva recension; satapatha brahmana; upanisad.

AtharvavedaShaunaka

5,839 verses. 5,715 dedicated English translations, 68 covered by a multi-verse range, 21 covered by a verified reused parallel rendering, 18 carrying a non-English scholarly note, 17 with none. 5,451 recitation records released. Not addressed: atharvaveda paippalada recension; gopatha brahmana; upanisad.

The exclusion most likely to mislead is the Yajurveda. In ordinary use “the Yajurveda” covers both the White and the Black recensions; this build holds the White one only, so no Taittiriya, Kathaka, Maitrayani or Kapisthala material is present at all. The Samavedic one is the largest in absolute terms: the gana song-books are a parallel and bigger body than the verse collection held here, and they are the reason the Samaveda is a distinct Veda rather than a Rigvedic excerpt.

Four kinds of knowledge

Everything this product knows sits in one of four layers, and the layer is attached to the claim rather than assumed from context. They are ordered by how much weight they can carry, strongest first.

What a source saysSource-explicit

The text itself, and the statements an edition or a traditional index makes about it. The verse, its number, the seer and deity its Anukramani names, the translator's English. These are quotations, and the strongest thing this product holds.

What follows from it by ruleDeterministically derived

Anything a stated rule computes from the layer above, the same way every time. A hymn's deity projected down onto its verses, a citation resolved into a canonical key, two verses matched as carrying the same normalised wording. Reproducible, checkable, and no stronger than the rule that produced it.

What a model extractedSemantic or model-assisted

Statements pulled out of a translation by a language model and kept as candidates. Every one is marked, none is treated as a source statement, and this layer reaches the Rigveda only — 2,459 of them, a small share of a 35,131-assertion layer that itself spans all four collections. It is held at the greatest distance of anything here.

What someone concludedInterpretive claim

A reading of the data, recorded with the observation that would falsify it. There are six of these in the whole build, they are permanently marked as candidates, and they are never displayed in the same style as a count.

The rule that matters is that a lower layer never gets promoted by being displayed next to a higher one. The evidence page renders the three quantitative layers in visibly different styles for exactly this reason, and the visualizations mark an interpretation as an interpretation even when it is the obvious reading of the bars above it.

The Normalization Rule

Surface Sanskrit text is preserved and normalized to Unicode NFC across all witnesses. Accents (svara marks) are stripped solely for secondary phonetic search indexing and parallel token matching; accents are never permanently removed or mutated in source-explicit text representations.

The Predicate Scope Rule

Knowledge graph relationships (such as EXACT_PARALLEL, TEXT_REUSE, PARALLEL_WITNESS, and DEVATA_IN_MANTRA) model witnessed structural, lexical, and philological evidence. They do not assert doctrinal or theological equivalence.

Named, and dedicated to

This is the single most important distinction on the site, and it is the one most often collapsed elsewhere.

A deity being named in a verse and a hymn being dedicated to that deity are two different records. The first is a fact about the words: the verse says the name. The second is a fact about the tradition’s own index: the Anukramani states which god the hymn belongs to. A hymn dedicated to Indra can run for ten verses without saying his name, and a verse can name six gods while being dedicated to none of them.

They also have different scopes, and this is where a merged count goes wrong. The naming layer reaches all four collections. The ascription that resolves to a named deity exists for the Rigveda; the Atharvaveda has an index of its own, and it records descriptive phrases rather than registry names, so its dedications are real and are not the same measurement. So a merged “Indra count” would be four collections of one measure plus one collection of another, and a zero for the Atharvaveda would look like a statement about the Atharvaveda when it is a statement about which index this build can join to the deity registry.

The deities plate draws the two side by side and never on one axis.

Certain, probable, ambiguous

Deciding that a word in a verse refers to a god is not always straightforward, because in Vedic Sanskrit the god and the thing often share a name. Every mention carries one of three grades.

CERTAIN

The word in the verse names this deity and nothing else plausibly. Included in every default figure.

PROBABLE

The reading is very likely but the form is shared with something else, or the context is doing the work. Included in default figures, and counted separately so you can remove it.

AMBIGUOUS

The word could equally be the god or the ordinary noun, and nothing decides it. Excluded from every default figure. The standing example is Agni: Vedic Sanskrit uses one word for the god and for fire, and the same problem recurs wherever a deity is named after the thing it is.

Default figures are CERTAIN plus PROBABLE. All three counts are always available, and the ambiguous tier can be switched on where it is offered — labelled, because a figure that includes it is an upper bound rather than a count.

Why this site sometimes says it cannot establish something

An empty result has at least three meanings and they are not interchangeable. The thing genuinely does not occur. The layer that would have found it was never built for that collection. Or the layer exists, reached the collection, and cannot support the claim. Returning 0 for all three is how a reader ends up believing something about the Vedas that is only true about a database.

So this product distinguishes them. NOT_BUILT means the dimension does not exist here. INSUFFICIENT_EVIDENCE means it exists and cannot answer. A measured zero is printed as a zero and says what it counted. Nothing is ever drawn as a bar of length zero unless a real zero was measured.

Kinds of connection

Two verses can be related in several unrelated ways, and this product never averages them into a similarity score. Seven kinds are carried, and the difference between the strongest and the weakest is considerable.

Exact parallel

The same verse, word for word, in two places. The strongest textual relationship here and the least common.

Near parallel

Closely similar wording that stops short of identity — a changed word, a different ending, an inserted particle.

Text reuse

One verse carries another's wording, with the direction recorded. Direction is expensive to establish and exists here for two of the six corpus pairs — Rigveda to Samaveda, and Atharvaveda to Rigveda.

Variant

The same verse with edition-level differences. A fact about how the text was transmitted rather than about how it was composed.

Formula family

A fixed phrase, plus the longer lines that contain it and the slightly different lines that echo it. Built by normalised string match over the whole corpus.

Shared vocabulary

Two verses name the same registered entities. This is the weakest relationship in the set and the one most easily mistaken for the others: it says nothing at all about shared wording.

Semantic assertion

A statement about a single verse. Most are derived by rule from scholarly morphological annotation and a minority are extracted by a model from a translation; the layer reaches all four collections, none of it has been reviewed by a human, and it is never counted as a textual relationship.

The transmission plate shows all of them against all six corpus pairs, with the cells that are not measurements typed rather than zeroed.

How Ask works

Ask VedAnvaya answers questions in ordinary language, and the important thing about it is the order in which it does things.

A question is first resolved against the registry — the names in it are matched to entities that actually exist in this graph, and a name that does not resolve is reported rather than guessed at. The resolved question drives a retrieval over the graph and the corpus, which produces an evidence packet: specific verses, specific entities, specific counts, each with its canonical key. Only then is a language model involved, and its job is to write prose over that packet. Its output is then validated: every citation it makes is checked against the packet, and a citation that does not resolve is not shown as a citation.

The model is not the database. It does not look anything up, it cannot reach the graph, and it is not asked what it knows about the Vedas. If the retrieval finds nothing, the answer says so instead of filling the gap — which is the failure mode this design exists to prevent. Answer quality does depend in part on which model is configured, and that is a real limitation rather than a footnote.

Measured on a fixed benchmark of 60 questions, the current build returned 60/60 acceptable outcomes: 39 answered from evidence, 4 answered in part, 17 refused for want of evidence, 0 misleading and 0 invented. A refusal is counted as a correct outcome, because it is one.

Recitation

Audio is a product layer over the text, not part of the knowledge model. No recording is part of the graph, no recording is republished here, and every file is streamed from its source at the moment it is played.

17,780 recitations are catalogued, one file per verse, all of them mapped to an exact verse rather than to a hymn. A mapping can be exact — this recording recites this verse — or structural, where a recording covers a hymn and the verse is inside it, or external only, where the recording exists and this product can point at it without placing it. Those are kept apart, and a recording whose source numbers its verses by a different recension is rejected rather than approximately aligned: playing the wrong verse is worse than playing nothing.

A checked mapping is not an audible review. A recording is catalogued because its mapping was established — the source’s own verse coordinates resolve to our canonical key, the Sanskrit matches this corpus’s text, and the media resolves — and not because anyone has listened to it. 20 of them carry an audible review. An owner sample of 20 recordings was listened to in full, with 20 verified and 0 rejected, and a separate queue of 1,021 recordings is held outside the catalogue with 1,001 of them not yet heard individually. The per-collection figures are on the scope page.

Known limitations

The full catalogue is on its own page, each carrying the measurement behind it and a question that can be answered instead. The ones worth knowing before you read anything else:

• The Samavedic gāna corpus is outside this edition. A parallel and larger body than the verse collection held here, and the reason the Samaveda is a distinct Veda. 1,136 of 1,844 ārcika verses carry a validated source-explicit notation witness and 708 await textual alignment. The marks are recorded as codepoints and never extrapolated into sung melodies.
• The Krishna Yajurveda is absent entirely, and the Atharvavedic Paippalāda recension with it. Both are substantially different collections rather than minor variants.
• The Samaveda has no English translation of its own. 173 verses are covered by a verified reused Rigvedic rendering, aligned against text whose identity was checked character by character, and 1,671 carry none. A reused rendering is labelled as borrowed and is never counted as independent Samavedic English evidence.
• Several layers reach some collections and not others. The deity ascription that resolves to a named god is Rigvedic; the Atharvaveda has its own index and it records descriptive phrases rather than registry names, so the two are not one layer. The semantic layer is Rigvedic. The metre layer reaches the Rigveda and the Atharvaveda and neither of the other two.
• The mention layer is not one instrument. Rigvedic mentions come from a manual scholarly lemma annotation; the other three collections are matched on surface tokens and sandhi, with no morphology behind them. Comparing a Rigvedic rate with an Atharvavedic one compares two methods as well as two texts.
• Lexical counts are minima. They count verses in which a registered alias occurs. At least one cell in the metals grid is known to be wrong for this reason, and it is labelled rather than corrected.
• Recitation coverage is uneven and the Atharvavedic verse total itself diverges slightly from the attested one, for reasons that remain open research rather than settled.
• Ask depends partly on the configured model. Retrieval and citation validation are deterministic; the prose between them is not.

The sources

13 sources and 33 individual files, grouped by what each one contributes. This list is generated from the build’s own provenance records rather than written by hand, and the generator refuses to run if a file the build loaded is not accounted for here.

On the terms

The texts are ancient and the translations held here are out of copyright, but the electronic editions that carry them are not automatically free to reuse: a transcription is a work, and several of these carry share-alike or non-commercial terms that a reader redistributing them would inherit. Where a site states one thing and the file it serves states another, both are recorded above, and the file is the one that governs. 2 sources are consulted and not reproduced — VedaWeb Rigveda public data, Vedic Heritage Portal — read against the text to check it rather than to supply it.