Build an E-commerce Search Using MongoDB Vector Search and OpenAI

1
git clone https://github.com/ashiqsultan/mongodb-vector-openai.git

1
touch .env

1
echo "MONGODB_URI=your_mongodb_uri" >> .env
2
echo "OPENAI_API_KEY=your_openai_api_key" >> .env

1
npm install      # (or) yarn install

1
npm run dev    # (or) yarn run dev

1
// server.ts
2
import dbClient from './dbClient';
3
server.listen(app.get('port'), async () => {
4
  try {
5
    await dbClient();
6
  } catch (error) {
7
    console.error(error);
8
  }
9
}); 

1
interface IProducts {
2
  name: string;
3
  category: string;
4
  description: string;
5
  price: number;
6
  embedding: number[];
7
}

1
{
2
  "mappings": {
3
    "dynamic": true,
4
    "fields": {
5
      "embedding": {
6
        "dimensions": 1536,
7
        "similarity": "euclidean",
8
        "type": "knnVector"
9
      }
10
    }
11
  }
12
}

1
async function generateEmbedding(inputText: string): Promise<number[] | null> {
2
  try {
3
    const vectorEmbedding = await openai.embeddings.create({
4
      input: inputText,
5
      model: 'text-embedding-ada-002',
6
    });
7
    const embedding = vectorEmbedding.data[0].embedding;
8
    return embedding;
9
  } catch (error) {
10
    console.error('Error generating embedding:', error);
11
    return null;
12
  }
13
}

1
// Example Product item
2
// product = {
3
// name: 'foo phone',
4
// category: Electronics,
5
// description: 'This phone has good camera',
6
// price: 150,
7
// };
8
9
const toEmbed = {
10
  name: product.name,
11
  category: product.category,
12
  description: product.description,
13
};
14
15
// Generate Embedding
16
const embedding = await generateEmbedding(JSON.stringify(toEmbed));
17
const documentToInsert = {
18
…product,
19
embedding,
20
}
21
22
await productCollection.insertOne(documentToInsert);

1
// updateObj contains the extracted request body with updated data 
2
const updateObj = {
3
     name: “Super Phone"
4
};
5
6
const product = await collection.findOne({ _id });
7
8
const objToEmbed = {
9
     name: updateObj.name || product.name,
10
     category: updateObj.category || product.category,
11
     description: updateObj.description || product.description,
12
};
13
14
const embedding = await generateEmbedding(JSON.stringify(objToEmbed));
15
16
updateObj.embedding = embedding;
17
18
const updatedDoc = await collection.findOneAndUpdate(
19
     { _id },
20
     { $set: updateObj },
21
     {
22
       returnDocument: 'after',
23
       projection: { embedding: 0 },
24
     }
25
);

1
http://localhost:5000/product?search=phones with good camera under 160 dollars

1
const searchProducts = async (searchText: string): Promise<IProductDocument[]> => {
2
  try {
3
    const embedding = await generateEmbedding(searchText); // Generate Embedding
4
    const gptResponse = (await searchAssistant(searchText)) as IGptResponse;
5
    …

1
async function main(userMessage: string): Promise<any> {
2
  const completion = await openai.chat.completions.create({
3
    messages: [
4
      {
5
        role: 'system',
6
        content: `You are an e-commerce search assistant. Follow the below list of instructions for generating the response.
7
        - You should only output JSON strictly following the Output Format Instructions.
8
        - List of Categories: Books, Clothing, Electronics, Home & Kitchen, Sports & Outdoors.
9
        - Identify whether user message matches any category from the List of Categories else it should be empty string. Do not invent category outside the provided list.
10
        - Identify price range from user message. minPrice and maxPrice must only be number or null.
11
        - Output Format Instructions for JSON: { category: 'Only one category', minPrice: 'Minimum price if applicable else null', maxPrice: 'Maximum Price if applicable else null' }
12
        `,
13
14
      },
15
      { role: 'user', content: userMessage },
16
    ],
17
    model: 'gpt-3.5-turbo-1106',
18
    response_format: { type: 'json_object' },
19
  });
20
21
  const outputJson = JSON.parse(completion.choices[0].message.content);
22
23
  return outputJson;
24
}

1
// … Existing code
2
const outputJson = JSON.parse(completion.choices[0].message.content);
3
console.log({ userMessage });
4
console.log({ outputJson });
5
return outputJson;

1
// Request 
2
http://localhost:5000/product?search=phones with good camera under 160 dollars
3
4
// Console logs from terminal
5
{ userMessage: 'phones with good camera under 160 dollars' }
6
{ outputJson: { category: 'Electronics', minPrice: null, maxPrice: 160 } }

1
const aggCursor = collection.aggregate<IProductDocument>([
2
      {
3
        $vectorSearch: {
4
          index: VECTOR_INDEX_NAME,
5
          path: 'embedding',
6
          queryVector: embedding,
7
          numCandidates: 150,
8
          limit: 10,
9
        },
10
      },
11
      matchStage,
12
      {
13
        $project: {
14
          _id: 1,
15
          name: 1,
16
          category: 1,
17
          description: 1,
18
          price: 1,
19
          score: { $meta: 'vectorSearchScore' },
20
        },
21
      },
22
    ]);

1
{"name": "JavaScript 101",
2
 "category": "Books",
3
 "description": "This is a good book for learning JavaScript for beginners. It covers fundamental concepts such as variables, data types, operators, control flow, functions, and more.",
4
 "price": 60
5
}

1
{"name": "Go lang Essentials",
2
 "category": "Books",
3
 "description": "A comprehensive guide to learning the Go programming language for beginners. This book is perfect for anyone looking to dive into Go programming.",
4
 "price": 70}

1
{"name": "Cracking the Coding Interview",
2
 "category": "Books",
3
 "description": "This book is a comprehensive guide to preparing for coding interviews, offering practice questions and solutions.",
4
 "price": 80}