Chunking & Embedding

Why Chunking Is Needed

Documents or FAQs are often long; to compute embeddings effectively and optimize similarity search, large text must be split into smaller segments (chunks).

Benefits

• Reduces context loss during embedding
• More accurate retrieval with vector similarity
• Better performance for search
• Compatible with token limits of embedding models

Chunking Strategy

The code uses RecursiveCharacterTextSplitter:

splitter = RecursiveCharacterTextSplitter(
    chunk_size=500,
    chunk_overlap=0
)
chunks = splitter.split_documents(docs)

Chunking Parameters

Optimization Tips

• Recommended chunk size: 500–1000 tokens (depends on embedding model)

• Chunk overlap: Use 50–100 characters if continuous context is needed

• Trade-off:

  • Smaller chunks → higher accuracy but more vectors
  • Larger chunks → fewer vectors but may lose context

Creating Embeddings and Vector Store

Initialize Embedding Model

emb = HuggingFaceEmbeddings(
    model_name="sentence-transformers/all-MiniLM-L6-v2"
)

Selected model: all-MiniLM-L6-v2

• Lightweight and fast
• Good for English
• Embedding dimension: 384

Create FAISS Vector Store

faq_store = FAISS.from_documents(chunks, emb)

FAISS (Facebook AI Similarity Search) provides:

• High-speed vector search
• Efficient handling of large datasets
• Multiple index algorithms

Query the Vector Store

results = faq_store.similarity_search(query, k=3)

Parameters:

• query: User question
• k=3: Returns top 3 most relevant chunks

Important Notes

Updating Data

If documents change (add/update):

• Re-embed all data, or
• Apply incremental updates to the vector store

Choosing an Embedding Model

Trade-off comparison:

Vietnamese Models

For better Vietnamese performance:

• keepitreal/vietnamese-sbert
• sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2

Full Code Example

from langchain_text_splitters import RecursiveCharacterTextSplitter
from langchain_huggingface import HuggingFaceEmbeddings
from langchain_community.vectorstores import FAISS

# Load documents
docs = load_faq_csv()

# Chunking
splitter = RecursiveCharacterTextSplitter(
    chunk_size=500,
    chunk_overlap=0
)
chunks = splitter.split_documents(docs)

# Embedding
emb = HuggingFaceEmbeddings(
    model_name="sentence-transformers/all-MiniLM-L6-v2"
)

# Vector Store
faq_store = FAISS.from_documents(chunks, emb)

# Query
query = "How do I change my password?"
results = faq_store.similarity_search(query, k=3)

Deployment Checklist

• Prepare documents (CSV, JSON, text files)
• Select appropriate chunk_size (test 500, 750, 1000)
• Choose embedding model (English or multilingual)
• Build and save FAISS index
• Test retrieval with sample queries
• Monitor and tune the k value for similarity search