Luna for ET / Documentation
This page explains, in a user-friendly way, how Aryan and Ajay built ET Compass: the product flow, the RAG engine, the memory layer, and the planning logic that makes Luna feel like a real ET guide instead of a plain chatbot.
Build Focus
ET-first concierge guidance, hybrid retrieval, persistent session memory, structured response planning, voice-AI, and responsive product UX in one platform.
Core Team
The platform was built as a collaboration between product, frontend, backend orchestration, and RAG engineering. The roles below keep the story clean and technically honest.
Platform, responsive frontend, connected backend flow, and RAG amplification
Built the complete ET Compass platform, reimplemented the responsive UI infrastructure, connected the backend graph end-to-end, and strengthened the RAG presentation layer so answers, widgets, and navigation work together.
RAG foundation, MongoDB session memory, and hybrid ET retrieval
Built the original RAG path from scratch, set up MongoDB-backed session memory, created the dedicated retrieval pipeline, and established the hybrid-search foundation that keeps Luna grounded in ET knowledge.
Architecture
Luna reads the user's natural question and identifies whether the need is discovery, markets, learning, events, benefits, roadmap, or comparison.
The system uses session memory to retain profile signals like user type, goal, experience level, and ET lane without forcing the user to repeat everything.
ET product registry, curated ET sources, vector retrieval, and keyword matching are combined so the answer stays grounded and not purely generative.
Each question is scored against ET lanes like ET Prime, ET Markets, ET Portfolio, ET Masterclass, ET Events, and ET Benefits to find the best fit.
A unified decision object chooses answer depth, structure, next action, and whether the UI should help with bullets, tables, or a contextual module.
The final response and the frontend use the same decision state, so Luna's text, widgets, and concierge rail stay aligned.
Sarvam speech-to-text and text-to-speech sit on top of the same ET concierge answer path, so the voice mode uses the existing grounded RAG answer instead of bypassing it.
Why This RAG Is Stronger
Luna is not only retrieving chunks and sending them to a model. The system combines hybrid retrieval, session memory, product scoring, and a response planner so the answer feels intentional.
That means the backend decides not only the text, but also the best ET lane, the next step, and whether the interface should render a cleaner table, bullet structure, or contextual module. This is the core of answer and UI sync.
The result is a higher-quality ET concierge that stays focused on real Economic Times discovery instead of turning into a generic assistant with random finance knowledge.
Voice AI
The voice layer was added without replacing the core ET concierge logic. Speech is only the interface. The actual ET answer still comes from the same retrieval, planning, and product-mapping backend.
01 / STT
Voice audio is transcribed first so the user still enters the same ET concierge flow as a text user.
02 / RAG
The transcribed text is passed into the current Stage 2 ET concierge graph, which keeps the answer grounded in ET products, source citations, and session memory.
03 / TTS
The final grounded answer is cleaned for voice playback and spoken back to the user, while the same turn is still saved in the thread history and ET journey memory.
Summary
These are the strongest user-facing capabilities the website now delivers.
Luna answers open ET questions in a more human way instead of behaving like a rigid form or static FAQ.
The system remembers the user's path and keeps narrowing to the right ET lane over time.
ET registry data, curated ET sources, keyword cues, and vector retrieval work together to keep responses grounded.
Luna does not just answer. It scores ET products, chooses a primary path, and proposes the next best action.
The frontend can show tables, bullets, next actions, and market context only when those formats actually help.
The project includes the landing page, auth flows, profile dashboard, concierge search UI, and deployable backend/frontend split.
Users can now speak to Luna through Sarvam voice APIs while still using the same grounded ET answer path, session memory, and concierge logic underneath.