Building an AI Sales Engineer for an Electronics Store with Hybrid RAG & LangGraph

Most AI shopping assistants can answer questions. Very few actually understand what you’re trying to build. That was the problem one of my friends faced.

He owns an eCommerce store that sells SMD components, electronic modules, sensors, development boards, and other hardware.

His customers weren’t asking questions like:

“Do you have SKU #A10234?”

Instead, they were saying things like:

“I want to build a home automation system.”

Or…

“I’m making a smart irrigation controller.”

Or even…

“I need something to measure temperature and send it over Wi-Fi.”

They didn’t know the component names.

They didn’t know the part numbers.

Sometimes they didn’t even know where to start.

A traditional search bar is useless when the customer doesn’t know what they’re searching for.

That’s when I realized…

This wasn’t a chatbot problem.

It was an engineering problem.

The Vision

Instead of building another product search chatbot, I wanted to build something much more useful.

An AI sales engineer.

Someone who understands the customer’s project first…

…then recommends the right components.

The assistant needed to:

• Understand electronics projects
• Recommend the required components
• Find products from the inventory
• Suggest compatible alternatives
• Search products by part number
• Add products to the shopping cart
• Take Cash on Delivery orders
• Help customers complete purchases naturally

That meant combining conversational AI with an actual understanding of the product catalog.

Challenge #1 Traditional Search Doesn’t Understand Intent

Most eCommerce websites rely on keyword search.

That works if the customer already knows what they need.

But electronics is different.

Customers often describe problems instead of products.

For example:

“I want to automate my lights.”

There isn’t a product called:

“Home Automation.”

The AI has to understand the project first.

Then infer that the customer might need things like:

• ESP32
• Relay Module
• Power Supply
• Motion Sensor
• Smart Switch
• Jumper Wires

This immediately ruled out a traditional keyword search.

Challenge #2 Thousands of Products Don’t Fit Inside an LLM

The store contains a large catalog of electronic components.

Obviously…

I couldn’t dump the entire catalog into the language model.

That would be expensive, slow, and unreliable.

Instead, I built a knowledge pipeline.

Every product is fetched directly from the store’s API.

The product information is then cleaned and summarized into a format that’s easier for an AI model to understand.

Finally, every product is converted into vector embeddings and stored inside Qdrant.

Now the AI doesn’t search raw product data.

It searches meaning.

Challenge #3 Vector Search Alone Wasn’t Enough

Vector search is incredibly powerful.

But electronics has one unique challenge.

Many customers search using exact model numbers.

Something like:

“ESP32-WROOM-32E”

Or…

“LM2596”

Vector search isn’t always the best solution for exact identifiers.

So I combined two approaches.

If the customer searches using natural language…

The assistant performs semantic vector search.

If the customer searches using a specific product ID, SKU, or module name…

The system performs an exact lookup.

The result is a Hybrid RAG architecture that combines semantic understanding with precise product retrieval.

It delivers the best of both worlds.

Challenge #4 Customers Don’t Care About Brands

One of the biggest insights came from watching real customer conversations.

Customers rarely care about a specific manufacturer.

They care about functionality.

If one relay module is unavailable…

They still want a relay module.

If one ESP32 board is out of stock…

Another compatible board is usually acceptable.

The assistant doesn’t stop at recommending exact matches.

It also understands compatibility.

If the preferred product isn’t available, it suggests similar components from different brands that achieve the same result.

That dramatically improves the shopping experience.

Challenge #5 Recommendations Should Lead to Purchases

Many AI assistants stop after answering questions.

I wanted mine to continue helping.

Once customers decide what they need, the assistant can:

• Add products directly to the shopping cart
• Update quantities
• Review the order
• Collect delivery information
• Place Cash on Delivery orders

The conversation naturally transitions from technical guidance into checkout without forcing the customer to leave the chat.

Why I Chose LangGraph

As the project grew, a simple prompt wasn’t enough.

The assistant needed reasoning.

Sometimes it needed to search products.

Sometimes retrieve similar items.

Sometimes answer questions.

Sometimes modify the cart.

Sometimes create an order.

Those aren’t one conversation.

They’re multiple workflows.

That’s why I built the architecture using LangGraph.

Instead of one massive AI prompt, the assistant follows a graph of specialized steps.

Each node has a single responsibility.

Retrieve products.

Search vectors.

Perform exact lookups.

Compare alternatives.

Update the shopping cart.

Place the order.

The result is far easier to maintain than one giant prompt trying to do everything.

The Final Architecture

The overall system looks surprisingly simple.

Customer Question
        │
        ▼
    LangGraph Agent
        │
        ├──────────────┐
        ▼              ▼
 Vector Search     Exact Search
   (Qdrant)         (Product API)
        │              │
        └──────┬───────┘
               ▼
      Product Reasoning
               ▼
 Compatible Recommendations
               ▼
      Shopping Cart Actions
               ▼
      Cash on Delivery Order

Each part focuses on one job.

Together they feel like a knowledgeable electronics engineer helping customers in real time.

Lessons Learned

This project taught me something important.

People don’t shop for electronic components.

They shop for solutions.

Nobody wakes up wanting an LM2596 voltage regulator.

They want a stable power supply.

Nobody searches for a PIR motion sensor because they love sensors.

They want their lights to turn on automatically.

Once the AI understands the customer’s goal instead of their keywords, the entire shopping experience changes.

Final Thoughts

At first, this looked like another AI chatbot project.

It turned into something much more interesting.

An AI assistant that understands electronics projects, retrieves products through Hybrid RAG, recommends compatible alternatives, manages shopping carts, and even completes Cash on Delivery orders.

The best part isn’t that it answers questions.

It’s that it helps customers discover products they never knew they needed.

That’s the difference between building a chatbot…

…and building an AI sales engineer.