Warung Assistant

Summary

A warung owner answers the same five questions all day while the queue waits: prices, stock, shelf locations, estimated totals. Warung Assistant is a chat front desk that takes those questions instead, so the queue keeps moving while the owner handles cash.

---

A small shop hears the same things every day

1. Repeated customer questions For example:

   • “How much is Indomie Ayam Bawang?”
   • “Is there any egg left?”
   • “Which shelf is the soy sauce on?”

2. Distractions during checkout

   • Owners get interrupted while adding up a purchase
   • Every interruption is another chance to miscount

3. Unstructured customer input Customers speak naturally, like:

   • “Half a kilo of eggs, 3 Indomie, and 2 bundles of spinach”

   One customer says ‘half a kilo’, another says ‘500 grams’, and both mean eggs.

4. Rigid POS systems

   • They cannot read natural language
   • They cannot help a customer before the order is finalized

---

A front desk that reads slang

How an order flows

1. The customer chats: prices, stock, shelf, estimated total.

2. Turning “half a kilo of eggs” into numbers

   • It pulls out the product names, quantities, and units
   • Supports:
     • Kilograms, halves, quarters
     • Grams / ounces
     • Unit-based quantities
     • Budget-based requests (e.g., “10,000 IDR worth of sugar”)
     • Product name synonyms

3. Checking before answering

   • Fuzzy product matching using RapidFuzz
   • Stock validation (requested quantity ≤ available stock)
   • Subtotal and grand total calculation

4. The answer the owner sees

   • Product list
   • Unit price
   • Quantity
   • Subtotal per item
   • Grand total
   • Shelf location (if requested)

5. Checkout for the owner

   • Fewer things to hold in your head
   • Faster checkout
   • Fewer counting mistakes

---

Architecture

graph TB
    User[Customer chat] --> UI[Gradio ChatInterface]
    UI --> EX[LangChain AgentExecutor]
    EX --> ZS[ZeroShotAgent on Ollama gemma3 12b]
    ZS --> T1[check stock]
    ZS --> T2[get price from csv]
    ZS --> T3[calculate total price]
    ZS --> T4[find product location]
    T1 & T2 & T4 -. lookup .-> CSV[(CSV inventory)]
    T1 & T2 & T4 -. fuzzy match .-> RF[RapidFuzz + unit parser]
    T1 & T2 & T3 & T4 --> ANS[answer with totals + shelf]
    ANS --> User

What it should change

I never measured this in a real shop, so read the list below as the target, not the result:

• Faster customer service
• Fewer calculation errors
• Less to keep in your head at the counter
• Better customer experience
• A foundation for POS and inventory systems

What I took from this one:

• Designing practical LLM systems
• Tool-based reasoning
• Handling real-world constraints
• Building end-to-end AI applications

---

Why these tools

LLM & AI

• Ollama (Local LLM: Gemma / LLaMA / Mistral)
• LangChain (Agents, Tools, Memory)
• Prompt Engineering (ReAct, Few-shot prompting)

The model runs locally because a warung has no GPU budget and often no reliable internet. LangChain was the fastest way to wire tools to the model at the time I built this.

Retrieval & Data

• CSV-based Inventory (RAG-style retrieval)
• RapidFuzz for fuzzy product matching
• Custom quantity and unit parsing logic

No vector database on purpose. A shop inventory is a few hundred rows with exact prices, and a CSV lookup returns the exact number every time. Embeddings would only add creative ways to be wrong about prices.

Backend

• Python
• Pandas
• Modular tool-based architecture:
  • check_stock
  • get_price_from_csv
  • calculate_total_price
  • find_product_location

Interface

• Gradio ChatInterface
• A chat box that updates as you type.

Gradio because the whole interface is a chat box, and that is a dozen lines of code instead of a frontend project.

---

Not yet built

Things I’d like to build on top of it:

• Role-based access (Customer mode / Owner mode)
• Transaction export (receipt, CSV, PDF)
• Automatic stock updates after payment
• Vector database integration (FAISS / Chroma)
• Voice input (Speech-to-Text)
• Multi-turn memory (e.g., “What did I buy earlier?”)

---

What made this one fun was the unglamorous part

On the surface it is a chat box. Underneath it is parsing slang, fuzzy-matching names, and checking stock. That plumbing took most of the work.

The piece I still do not have a good answer for is a customer who changes their mind halfway through an order. Right now the safest move is to hand that back to the owner.

---

Source

The full code is on GitHub: https://github.com/rasyidev/warung_assistant