Chapter 2: Embedding Models

Introduction

Choosing the right embedding model is crucial for RAG performance. This chapter covers the major embedding providers and their models.

OpenAI Embeddings

text-embedding-3-small

from openai import OpenAI

client = OpenAI()

response = client.embeddings.create(
    input="What is RAG?",
    model="text-embedding-3-small"
)

embedding = response.data[0].embedding  # 1536 dimensions

Spec	Value
Dimensions	1536
Max tokens	8191
Price	$0.02 / 1M tokens
Quality	Good

text-embedding-3-large

response = client.embeddings.create(
    input="What is RAG?",
    model="text-embedding-3-large",
    dimensions=3072  # Can reduce: 256, 1024, etc.
)

Spec	Value
Dimensions	3072 (configurable)
Max tokens	8191
Price	$0.13 / 1M tokens
Quality	Excellent

Dimension Reduction

# Reduce dimensions for efficiency
response = client.embeddings.create(
    input="What is RAG?",
    model="text-embedding-3-large",
    dimensions=1024  # Reduced from 3072
)
# Still maintains good quality!

Cohere Embeddings

embed-v3

import cohere

co = cohere.Client("your-api-key")

response = co.embed(
    texts=["What is RAG?"],
    model="embed-english-v3.0",
    input_type="search_document"  # or "search_query"
)

embeddings = response.embeddings

Key features:

• Different modes for documents vs queries
• Compression options
• Multilingual support

Spec	Value
Dimensions	1024
Max tokens	512
Price	$0.10 / 1M tokens
Quality	Excellent

Input Types

# For indexing documents
doc_embeddings = co.embed(
    texts=documents,
    model="embed-english-v3.0",
    input_type="search_document"
)

# For search queries
query_embedding = co.embed(
    texts=[query],
    model="embed-english-v3.0",
    input_type="search_query"
)

Open Source Models

Sentence Transformers

from sentence_transformers import SentenceTransformer

# Load model
model = SentenceTransformer('all-MiniLM-L6-v2')

# Generate embeddings
embeddings = model.encode([
    "What is RAG?",
    "Retrieval Augmented Generation explained"
])

Popular models:

Model	Dimensions	Quality	Speed
all-MiniLM-L6-v2	384	Good	Fast
all-mpnet-base-v2	768	Better	Medium
multi-qa-mpnet-base	768	Best for Q&A	Medium

BGE Models (BAAI)

from sentence_transformers import SentenceTransformer

# BGE models
model = SentenceTransformer('BAAI/bge-large-en-v1.5')

# With instruction prefix (recommended)
queries = ["Represent this sentence for searching: What is RAG?"]
docs = ["Represent this document for retrieval: RAG combines..."]

query_emb = model.encode(queries)
doc_emb = model.encode(docs)

Model	Dimensions	MTEB Score
bge-small-en-v1.5	384	62.17
bge-base-en-v1.5	768	63.55
bge-large-en-v1.5	1024	64.23

E5 Models (Microsoft)

from sentence_transformers import SentenceTransformer

model = SentenceTransformer('intfloat/e5-large-v2')

# Requires prefixes
queries = ["query: What is RAG?"]
passages = ["passage: RAG is a technique..."]

Using with LangChain

# OpenAI
from langchain.embeddings import OpenAIEmbeddings
embeddings = OpenAIEmbeddings(model="text-embedding-3-small")

# Cohere
from langchain.embeddings import CohereEmbeddings
embeddings = CohereEmbeddings(model="embed-english-v3.0")

# HuggingFace
from langchain.embeddings import HuggingFaceEmbeddings
embeddings = HuggingFaceEmbeddings(
    model_name="BAAI/bge-large-en-v1.5"
)

# Use with vector store
vectorstore = Chroma.from_documents(
    documents=chunks,
    embedding=embeddings
)

Comparison Table

Model	Provider	Dims	Price/1M	Quality	Speed
text-embedding-3-small	OpenAI	1536	$0.02	Good	Fast
text-embedding-3-large	OpenAI	3072	$0.13	Excellent	Medium
embed-v3	Cohere	1024	$0.10	Excellent	Fast
bge-large-en-v1.5	BAAI	1024	Free	Excellent	Self-host
all-MiniLM-L6-v2	SBERT	384	Free	Good	Very Fast

Summary

• OpenAI - Best for most use cases, easy API
• Cohere - Excellent quality, search optimization
• BGE/E5 - Best open source, self-hosted
• MiniLM - Fast and lightweight, good enough

---

Next Steps

Continue to Chapter 3: Choosing Models for decision guidance.