How do I create embeddings for product descriptions and FAQ pages?

To create embeddings for product descriptions and FAQ pages, the general process is:

1. Collect the text
2. Clean and chunk it
3. Convert each chunk into an embedding vector
4. Store vectors in a database/index
5. Use them for search, recommendations, or Q&A

1) Decide what text to embed

For product pages, you might include:

• Product title
• Short description
• Full description
• Key features
• Specs
• Category

For FAQ pages, you usually want:

• Each question + answer pair as one unit
• Or each FAQ item separately

Example:

• Question: What is the warranty? Answer: 2 years...
• Question: Does it support fast charging? Answer: Yes...

2) Clean and prepare the text

Before embedding:

• Remove HTML
• Normalize whitespace
• Fix encoding issues
• Keep important structured info together

Example product text:

Title: Wireless Noise-Cancelling Headphones
Description: Over-ear headphones with active noise cancellation...
Features: 30-hour battery, Bluetooth 5.3, USB-C charging

3) Chunk long text

Embedding models have input limits and work better with manageable chunks.

For:

• Product pages: chunk by section if long
• FAQ pages: usually embed each Q&A separately

Good chunk size:

• ~200–500 words per chunk, depending on the model

4) Generate embeddings

Use an embedding model such as:

• OpenAI embeddings
• SentenceTransformers
• Cohere embeddings
• Hugging Face models

Example with Python and SentenceTransformers

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("all-MiniLM-L6-v2")

texts = [
    "Wireless Noise-Cancelling Headphones. 30-hour battery, Bluetooth 5.3...",
    "Q: What is the warranty? A: Two years from date of purchase."
]

embeddings = model.encode(texts)

print(embeddings.shape)  # e.g. (2, 384)

5) Store embeddings with metadata

Store:

• embedding vector
• original text
• product ID / FAQ ID
• page URL
• category
• language

Example metadata:

{
  "id": "faq_123",
  "text": "Q: What is the warranty? A: Two years...",
  "source": "faq",
  "url": "/support/warranty",
  "category": "warranty"
}

6) Use them for retrieval

When a user searches:

1. Embed the query
2. Compare it with stored embeddings
3. Return the most similar items

Common similarity methods:

• Cosine similarity
• Dot product
• Euclidean distance

Best practices

• Embed meaningful text, not raw HTML or navigation menus
• Keep FAQ Q&A together
• For product data, combine title + key specs + description
• Rebuild embeddings when content changes
• Use metadata filtering to improve search results

Simple recommendation

For ecommerce:

• Product embeddings: one embedding per product, plus extra embeddings for major sections if the page is long
• FAQ embeddings: one embedding per question-answer pair

If you want, I can also show:

1. a Python example using OpenAI embeddings, or
2. how to store and query them in a vector database like FAISS, Pinecone, or Chroma.