How to Train an AI on Vedic Astrology Classical Texts

Training on Vedic texts is a data engineering and rights problem before it is a modeling problem. Chunking shastras without verse metadata produces fluent nonsense — users notice immediately. A fluent wrong citation is worse than an honest I do not know — verse-level fidelity is the quality bar classical astrology users enforce ruthlessly. Invest in corpus quality before model upgrades — better data beats bigger models for classical fidelity.

Corpus curation and rights

Document provenance: edition, translator, license, and update cadence. Prefer verse-level boundaries over arbitrary token splits. Partner with publishers or use clearly licensed corpora — scraping random PDFs creates legal and quality debt. Track edition year and translator for each text — retrieval must not mix conflicting commentaries without labeling source. Public-domain does not mean quality-domain; curate scans with readable Devanagari and minimal OCR damage. Partner astrologers during corpus tagging — their topic labels improve retrieval filters more than automated clustering.

Tag passages by topic: grahas, bhavas, dashas, yogas, remedies. Metadata enables filtered retrieval aligned to chart features. Rights conversations with publishers may unlock premium corpora competitors cannot scrape — legal access is a moat, not overhead. Version corpus releases like software so eval baselines stay comparable. Maintain a banned-sources list for scraped content that failed human review — prevent re-ingestion via pipeline guards.

Multilingual ingestion

Normalize Devanagari, IAST, and English commentary into consistent UTF-8 pipelines. Test embedding models on Sanskrit compounds and transliteration variants. Mixed-language chunks need language tags in metadata for filter precision. Test embedding models on compound Sanskrit terms and sandhi splits — English-only models fail silently on transliterated queries. Store both original script and normalized IAST in metadata for display versus search. Build a transliteration test suite covering fifty common graha and bhava queries users actually type.

OCR classical scans only with human QA samples — OCR errors teach wrong yogas. Keep original script in display citations even when retrieval uses normalized text. Chunk at verse boundaries with overlap only within adjacent verses — mid-shloka splits produce interpretations that contradict classical grammar. Language tags enable filtering when users query in Hindi versus English. Document OCR confidence scores per page — low-confidence pages route to human transcription before indexing.

Eval and human review

Golden questions tied to known chart configurations and expected classical references. Score faithfulness and citation accuracy — not marketing fluency. Golden questions should include known yogas for test charts — expected citations are specific verses, not thematic similarity. Score citation match exactly, not embedding cosine alone. Publish eval reports internally after each corpus version — regressions block release like code failures.

Senior astrologers review edge cases monthly. Connect eval to Navamsha-style architecture: engine facts plus retrieved verses. Monthly review by practicing astrologers catches cultural nuance automated metrics miss — festivals, regional traditions, and remedy boundaries. Tie eval failures back to corpus gaps, not only prompt tweaks. Connect architecture decisions to Navamsha-style deterministic charts so eval tests known planetary configurations.

What to do next

Start with one measurable outcome — latency, accuracy, adoption, or revenue — and design your stack around that constraint. GenOrbit ships production AI systems across healthcare, astrology, fintech, and SaaS; if you want a second pair of eyes on architecture or delivery, we are happy to talk.