Grounded by design
AtmaLoka is built so that answers stay close to the source. Here is a plain look at what happens between your question and the reply.
It searches the source, not the open web
Your question is matched against a curated library of primary scriptures. AtmaLoka does not pull from random websites. It looks only at the texts we have read and indexed.
Hybrid retrieval, then careful ranking
We combine meaning-based search with classic keyword search, then re-rank the candidates so the most relevant passages rise to the top. This is the retrieval-augmented generation (RAG) approach, tuned for scripture.
Answers are tied to verses
The guide writes its answer using the retrieved passages and cites them. You can open each citation and read the original, so nothing is taken on faith alone.
Multilingual by design
We use self-hosted multilingual embeddings, so a question in plain English can surface the right Sanskrit or Devanāgarī passage across the corpus.
An agent that reads whole texts
For deep questions, the guide works in steps: it locates a text, reads it end to end, keeps track of what matters, and composes a complete answer rather than a quick snippet.
Honest about the limits
Grounding in the source greatly reduces errors, but it does not eliminate them. The underlying model can still misread or overreach. That is why we always show citations, encourage you to read the verses yourself, and make it easy to flag anything that looks wrong. Treat AtmaLoka as a knowledgeable study companion, not a final authority.