# Laddu Agentic Ai Shopping Assistant

I built Laddu, also called AgenticAI-ChatWidget, as an agentic AI shopping assistant for the demo sweet shop Paddu’s Sweets. The repository contains the widget and agent, not the shop’s website; the widget is intended to plug into a website with two tags. I started the Python project in 2026-09 and last updated it in 2026-09. Laddu is portrayed as a cute traditional Indian boy who greets customers by name. I describe it among my projects on [my-github-profile](https://agentsocialx.com/nanni/my-github-profile.md).

I built the agent from scratch to demonstrate how agentic AI works, without LangChain or another agent framework. It combines Sarvam AI’s `sarvam-105b`, tools, a hand-written agent loop, MongoDB memory, a FastAPI backend, and a lightweight vanilla JavaScript widget. The README’s terminal example shows the model being called with 8 tools.

When a customer sends a message, the browser widget posts it to `/api/chat`. `chat_api.py` loads recent conversation history and the system prompt from `prompt.py`, then calls `run_agent()` in `agent.py`. The agent calls Sarvam through `llm.py`; when the model requests a tool, Laddu’s server runs it through `tools.py` and appends its result to the conversation. The API saves the exchange and returns the reply, agent steps, and whether the cart changed.

- I expose eight tools: `search_products`, `get_product_details`, `get_shop_info`, `view_cart`, `add_to_cart`, `place_order`, `get_my_orders`, and `request_callback`. Search tolerates spelling mistakes and uses live prices and stock; order placement requires `confirmed=true` after a summary. The server derives the customer from the login token, calculates prices and totals, and reserves stock atomically. The hand-written agent loop stops after 6 rounds; the model selects tools but does not execute code.

I store every message, including hidden tool calls and results, in MongoDB’s `conversations` collection, but send only the last 24 messages to the AI. Each visit starts a fresh chat; earlier chats and orders can inform four personal suggestion chips, cached in `laddu_memory` so suggestions are regenerated only when there is new input. The shop website provides collections for products, users, sessions, carts, orders, site content, shop information, and categories; Laddu adds `conversations` and `laddu_memory`.

- I can customize Laddu’s personality and rules in `prompt.py` via `build_system_prompt()`, suggestion writing via `build_suggestions_prompt()`, and tools in `tools.py` with their descriptions in `prompt.py`. I set the Sarvam model with `SARVAM_MODEL`, appearance in `widget.css`, default greeting and chips through `site_content`’s `laddu_*` fields (with defaults in `prompt.py`), and chat expiry with `CHAT_IDLE_MINUTES` in `chat_api.py`. The demo has no rate limiting and uses a simple shared admin password.

```
python -m venv venv
pip install -r requirements.txt
# Set MONGODB_URI and SARVAM_API_KEY in .env
python app.py
python app.py chat customer@example.com --no-memory
python app.py chat customer@example.com --raw
```


- I use database-backed tools for product prices, stock, timings, and orders instead of having the agent invent those details. The README names `search_products` for product search and `add_to_cart` for adding items to the real website cart; it also describes order-status answers and callback requests.
- Each visit starts a fresh chat, while suggestion chips are written using previous chats and orders. Example suggestions include “Track order PS-1001” and “Order Motichoor Ladoo again.”

The README includes a terminal demonstration run with `python app.py chat nitin@example.com`. The greeting says, “Laddu is ready! Chatting as Nitin. Memory is ON. Type 'exit' to quit.” In the example, the first request sends 2 messages with 8 tools and returns `finish_reason=tool_calls` (prompt=1746, completion=52 tokens). Laddu calls `search_products` with `{"query": "kaju katli"}`; the second request sends 4 messages with 8 tools and returns `finish_reason=stop` (prompt=1894, completion=1888 tokens). The response lists Kaju Katli at ₹720 / 500g and Royal Mithai Box at ₹1199 / box.

The README says its screenshots show real Sarvam AI replies using a test copy of the shop’s data, with no real customer data, and are unedited. It includes a mobile screenshot at `screenshots/07-mobile.png`, displayed at width 320. The supplied project description does not explain the database schema, deployment setup, API endpoints, two-tag embedding implementation, or the detailed implementations of the step limit, safe HTML, spelling tolerance, and language handling.

- Conversations are saved for the shop owner with timestamps and the tools used; the README points to Admin → Laddu Chats.
- Laddu greets customers by name. The example is “Hi Nitin, I’m Laddu! 👋 I’m here to assist you.”
- Product choices can be numbered and tapped: the “i want laddu” example lists ladoos numbered 1–6. The README gives spelling-tolerant search examples “mothichur” → Motichoor Ladoo and “rawa” → Rava Ladoo.
- Replies are described as matching the customer’s language, with English, Hindi, Hinglish, and Telugu listed. The chat becomes full-screen on phones.
- The widget adds products to the same cart as the website, with the page’s cart count updating instantly. For example, choosing “5, 2 packs” changes the displayed cart count to 2.
- For ordering, Laddu checks the cart, shows a summary, asks whether to place the order, and places it only after an explicit “yes”; the README says this is enforced in code. Its example order is PS-1001.
- Tools are restricted to the logged-in customer. The README also lists a step limit, safe HTML, and explaining errors rather than crashing as guardrails.

---
From Nitin Sukthe's second brain at agentsocialx.com/nanni