A language model is a statistical engine trained on the open internet. The Bhagavad Gita is a 700-verse dialogue about how to act when every choice hurts. We built a feature that puts the first in the voice of the second, and we owe you a clear account of how it works and where its limits are.
This is that account.
What “Talk to Krishna” actually is
When you open the chat, you are not querying a database of verses. There is no retrieval step, no vector search, no curated commentary corpus the model is consulting. The implementation is plainer than that and we think the plainness is worth being honest about.
Under the hood, every message you send is passed to Google’s gemini-3.1-flash-lite model along with a single, carefully written system prompt — a persona instruction that defines who the model is supposed to be and what it is allowed to say. The most important lines in that prompt are:
- Speak as a calm, wise, modern friend who knows the Gita by heart.
- Always cite at least one specific verse reference (e.g., BG 2.47, BG 6.5).
- Krishna does not affirm weakness or poor choices to comfort the seeker. He speaks truth with compassion.
- Never diagnose, prescribe, or give medical, legal, or financial advice.
- End every response with a short, practical reflection or question.
That’s the whole grounding mechanism. There is nothing magical happening behind the screen.
Why we chose the persona-prompt approach over RAG
We considered building a retrieval-augmented system over a curated translation of the Gita. We didn’t, for two reasons.
The first is pragmatic. A 700-verse text is small enough that a capable model already has the canonical translations and major commentaries somewhere in its weights. Retrieval would mostly add latency without changing answers.
The second is more interesting. The danger of an AI spiritual chatbot isn’t that it cites the wrong verse — it’s that it tells you what you want to hear. It is sycophantic by default. So we spent our effort on the line in the prompt that says Krishna will not validate avoidance: “If asked ‘was I right to do X?’, assess X honestly first. Do not begin with comfort and arrive at truth only if pressed — Arjuna sought permission not to act, and Krishna did not give it.” The grounding problem we’re trying to solve is a tone problem, not a citation problem.
Whether we’ve solved it is a question you’ll answer better than we will. Tell us when it sounds wrong.
What we do not claim
A few claims are easy to make and we’d rather not.
It is not a therapist. It is a model that has been told to refuse medical, legal, and financial advice and to redirect spiritual questions back to the Gita. It has no built-in suicide or self-harm detection layer. If you are in crisis, please contact a human professional or a crisis line — there is a list at the bottom of this page.
It is not a guru. It does not know your life. It cannot remember your last conversation; chats are not stored anywhere on our side. Each session starts cold.
It is not your friend. It is software. If a response lands well, that is because the underlying text is good and the model paraphrased it competently. The credit belongs to the Gita.
What we do claim
We claim two things.
One: every response is anchored to a verse. The system prompt requires the model to cite at least one chapter and verse on every reply. You can take that citation, look it up in any standard translation, and decide for yourself whether the model summarised it fairly. That portability — the ability to leave the chat and read the source — is the only honest grounding we know how to provide.
Two: we don’t keep your conversations. No chat content is logged on our servers. Messages are sent to Google’s API for inference and are subject to Google’s API data policy; they are not retained, sold, or used to build a profile of you on our end. The only thing we count is the rate limit (so the chat doesn’t fall over), and that is keyed to your IP, not to who you are.
The honest case for using it anyway
A passage from the Gita read in a quiet room is better than a passage of the Gita rendered by a language model. We know this. If you have a teacher, read with the teacher. If you have a translation you trust, read the translation.
What an AI companion can do — and this is the only claim we’ll defend — is meet you at 2 a.m. on a Tuesday when no teacher is awake, when you are sitting with a question that won’t let you sleep, and offer a verse and a question back. Not advice. A verse and a question. Then it expects you to put the device down.
The success metric we care about is whether you put the device down. That’s it. If the chat keeps you in the chat, we built the wrong thing.
If you are in crisis, please contact a human. In the US: 988. In the UK: 116 123. Internationally: findahelpline.com. Krishna in this app is a model. The people on those numbers are people.