Kundan Lane Jewellers
A family jeweller's store and its WhatsApp assistant
Kundan Lane is a sample brief. We imagined a third-generation family jeweller in Dadar whose customers ask the same questions on WhatsApp every day, about the rate, the making charge, old gold and HUID, and asked ourselves what we'd build for them. The shop is imaginary. The assistant isn't: we designed and engineered it to answer only from the store's own documents, show its source, and hand over to the family when it isn't sure, and you can use it right now, below.
A sample brief, a real product. We wrote this brief ourselves. The business is imaginary; the product is not: we designed and engineered it, and you can use it above. We publish results only for work a client has let us verify, so this page shows none.
The brief
We imagined a jeweller that opened on a lane off Ranade Road in 1961 and is now run by the founder's grandchildren: one runs the workshop upstairs, the other the online store and the WhatsApp number. Their customers are Dadar families buying for weddings, Dhanteras and barsas, and increasingly people elsewhere in India who found them online.
Almost every conversation with such a shop starts on WhatsApp, and most of those messages ask things the family has already written down somewhere: today's rate, how the making charge works, what happens to old gold from another jeweller, whether a piece is hallmarked.
- A third-generation family jeweller in Dadar West
- Prices that move daily with the gold rate
- Customers who type in English and Hinglish
- A two-person team answering WhatsApp between customers at the counter
The problem
Jewellery is priced by formula: weight, purity, a rate that changes every morning, a making charge that depends on the kind of piece, stones weighed separately, and GST. Customers rightly want that formula shown, and a family shop answering by hand ends up typing the same breakdown dozens of times a day, often after hours.
A general chatbot is the wrong answer here. A wrong rate, an invented discount or a made-up exchange percentage costs a jeweller trust that took decades to build, so the assistant has to say where every answer came from and stop when the documents don't cover the question.
- Repetitive questions about rates, making charges and old gold
- Prices that must be computed live, not quoted from memory
- Negotiations and complaints that need a person, not a bot
- Policies that change, and must change in the answers the same day
The approach
We made the store's own documents the single source of truth. The assistant retrieves passages, answers from them and cites the document and section, and live prices are calculated from the rate board and a making-charge table that is itself read from the documents, so editing a policy changes both the answers and the prices.
We designed the handover as a feature rather than a failure: price negotiation, complaints, custom designs and anything below a confidence threshold go to the family with their reply hours, and the staff side shows exactly why each answer was given. The storefront around it is built on the same honesty, with a rate board on every page and a price breakup on every piece.
- Answers only from the store's documents, always with a source
- Live prices from the rate board and the documented making charges
- Hinglish and common misspellings understood, not just English
- A visible confidence threshold and a clear handover to a person
- A staff side to edit documents, read handovers and test questions
What shipped
- Chat with the assistant in English or Hinglish and get answers drawn from twelve store documents, each with a tappable source that opens the passage highlighted
- Ask for a price like '10 gram 22 carat chain kitne ka padega' and get metal, making and GST worked out from today's board
- Watch it refuse to guess: discounts, complaints and off-topic questions are handed to the family with their reply hours
- Tap 'Ask about this piece' on any product to open the chat with that piece as context
- Edit a document or the rate board on the staff side and see the assistant and every price on the site change at once
- Open the staff inbox to see handed-over conversations flagged, and reply as Meera into the customer's chat
- Test any question and see how it was read, which passages scored under BM25, and why it answered or handed over
- Put a piece on the counter scale and watch the bill build from gross weight to GST
- Filter collections by metal, purity, weight, price and occasion, then buy through a sample UPI checkout with a GST invoice
- Plan Sona Sanchay instalments, value old gold, book a try-on with a calendar file, or send a custom design to the workshop
Under the surface
How it is built
In production, customer messages arrive through the WhatsApp Business Cloud API, via Meta directly or an Indian BSP, at a Node service. The store's documents are chunked by section and stored in PostgreSQL with pgvector alongside a keyword index, and each question is answered by hybrid retrieval followed by a language model that may only use the retrieved passages and must cite them. A refusal and handoff policy runs before and after generation, prices are computed in code from the gold-rate feed and the documented making charges rather than by the model, and the staff inbox and document editor are a Next.js app that re-indexes a document the moment it is saved.
- Hybrid retrieval: pgvector embeddings plus BM25, with a Hinglish synonym layer
- Generation constrained to cited passages, with answers that fail citation checks discarded
- Prices calculated in code from the rate feed and making-charge table, never generated
- Handoff rules for negotiation, complaints, custom work and low confidence, with business-hours awareness
- WhatsApp templates for order, visit and scheme confirmations sent from the same service
Considered and turned down
We considered letting the language model answer from a long system prompt containing every policy. We turned it down because it cannot show where an answer came from, it quietly blends policies together, and every policy edit would need a prompt change rather than a document save.