The context
Dharma AI explores a product question: can an AI conversation help someone pause and think through a difficult situation without pretending to provide definitive answers? It is designed for questions about work, relationships, fear and difficult choices, where the user needs perspective rather than a generic answer.
The question behind the product
A conventional chatbot can produce fluent advice without showing what informed it, and a source-heavy experience can be slow or difficult to use. The product challenge was to make a reflective conversation feel simple while preserving source context, scope boundaries and a dependable response experience.
The hypothesis
A mobile-first, anonymous conversation that retrieves a small set of relevant curated records before generating an interpretation can make ancient source material more useful for a modern dilemma than either a search experience or an ungrounded chat response.
The product approach
1. Start with the human dilemma
The experience begins with one open question. Suggested prompts help people understand the kinds of dilemmas the product is built to handle, while the conversation remains anonymous and mobile-first.
2. Keep the request inside useful boundaries
Scope checks, session context and request limits keep the system focused on reflective dilemmas and protect the experience from turning into an unrestricted advice engine.
3. Retrieve before reasoning
For the primary path, the system searches a curated corpus of Gita passages and Mahabharata and Ramayana episodes using semantic and keyword retrieval. It locally ranks the results and uses at most three records as grounding for Gemini synthesis.
4. Design for latency, not just the ideal path
If retrieval takes too long, a timed Gemini-only backup response is prepared at 3.5 seconds and can replace the primary path at 5.5 seconds. This preserves a thoughtful structure while making clear that the alternate response has no retrieved quotation or source link.
5. Return a response that helps thinking
The structured response includes a canonical verse or episode where available, an interpretation, two ethical paths forward, first-person follow-up prompts and a Sources section.
How it works
Dharma AI combines a light React and Vite interface with Firebase Hosting and Cloud Functions. Behind the chat surface is a protected orchestration layer that manages retrieval, prompt structure, response formatting, timing and fallback behaviour. The product is designed as an end-to-end AI experience rather than a single model call.
What made it difficult
The difficult part is balancing quality, source grounding and response time in the same interaction. Retrieval needs to find relevant material, the model needs enough context to interpret it without overclaiming, and the user should not be left waiting for an ideal response path that may be slow. The product therefore treats evaluation, live timing tests, regression checks, bounded usage and fallback logic as core product requirements.
What changed
The live product demonstrates two response paths: a grounded retrieval-augmented path and a timed Gemini-only fallback. Its architecture makes the distinction visible: grounded answers can surface curated source context, while the backup path keeps the conversation moving without presenting unverified quotations as evidence. Five synthetic live RAG requests measured on 10 September 2026 completed in 2.1–2.7 seconds; this is a small test sample, not a service-level guarantee.
What this work reinforces
For reflective AI products, the product decision is not only what the model says. It is how users understand the basis, limits and next step of a response. Evidence, latency handling and honest fallback behaviour all shape trust as much as the wording itself.
