Amazon Bedrock Models - When to Use What

Main Concept

A quick mental guide to when to use or not use a certain model available in Bedrock.

Context

This kind of questions are included on the certification exam.

DEFINITIVE CHART: Amazon Bedrock Models - When to Use What

USE CASE	MODEL TO USE	WHY	KEYWORD CLUES IN QUESTION
SEARCH & RETRIEVAL
Semantic search	Titan Text Embeddings V2	Converts text to vectors, finds similar	”search by meaning”, “semantic search”, “find similar”
Image search (text + image)	Titan Multimodal Embeddings	Handles text AND images in same space	”search products by photo OR description”, “image similarity”
Personalized recommendations	Titan Text Embeddings V2	Embeddings capture preferences	”recommend similar items”, “personalization”
CHATBOTS & Q&A
Enterprise chatbot with internal docs	Titan Text Premier + Bedrock Knowledge Bases	Optimized for RAG, native integration	”internal knowledge base”, “RAG”, “cite sources”
Simple chatbot, low cost	Titan Text Lite	Cheap, sufficient for basic tasks	”budget-constrained”, “simple Q&A”, “high volume”
Chatbot with complex reasoning	Claude 3 (Sonnet/Opus)	Better reasoning, deep analysis	”complex reasoning”, “nuanced”, “analysis”
DOCUMENTS & TEXT
Very long documents (100K+ tokens)	Claude 3	200K context window (largest)	“long documents”, “200+ pages”, “entire book”
Contract/legal analysis	Claude 3 Sonnet/Opus	Precise reasoning, long context	”contract analysis”, “legal documents”, “compliance”
Short document summarization	Titan Text Express	Price/performance balance	”summarize”, “extract key points” (short docs)
CODE
Code generation	Titan Text Premier / Claude 3	Both support code generation	”generate code”, “write script”, “programming”
Code explanation	Claude 3	Better for detailed explanations	”explain this code”, “debug”, “code review”
IMAGES
Generate images with READABLE text	Titan Image Generator	Only one that does readable text well	”readable text in image”, “signs”, “price tags”, “posters”
Generate artistic images	Stable Diffusion XL	More flexible for art/creativity	”artistic”, “creative”, “illustration”
Search images by content	Titan Multimodal Embeddings	For search, NOT generation	”search images”, “find similar photos”
Analyze/describe images	Claude 3 (with vision)	Understands images, does NOT generate	”what’s in this image”, “describe”, “analyze photo”
Detect if image is AI-generated	Titan Image Generator (watermark feature)	Invisible watermarking + DetectGeneratedContent API	”verify authenticity”, “detect AI-generated”
SPECIFIC TASKS
Simple classification (high volume)	Titan Text Lite	Cheapest for simple tasks	”classify into 5 categories”, “tag”, “millions of items”
Structured data extraction	Titan Text Express/Premier	Table creation, data formatting	”extract data”, “create table”, “structure information”
Translation	Titan Text Express (100+ languages)	Broad multilingual support	”translate”, “multilingual”
Autonomous agents (multi-step)	Titan Text Premier + Bedrock Agents	Native integration with Agents	”autonomous agent”, “multi-step workflow”, “tool use”
SPECIAL CASES
Open-source / Custom fine-tuning	Llama 2/3	Open-source model, customizable	”open-source”, “fine-tune with our data”, “self-hosted option”
Maximum accuracy (cost doesn’t matter)	Claude 3 Opus	Most capable (and most expensive)	“highest accuracy”, “most capable”, “budget flexible”

---

KEY FEATURES TABLE (For Quick Comparison):

MODEL	CONTEXT WINDOW	BEST FOR	PRICE	DON’T USE WHEN
Titan Text Lite	4K tokens	Simple tasks, high volume	💰 Cheapest	❌ Complex reasoning
Titan Text Express	8K tokens	General use, balanced	💰💰 Medium	❌ Very long docs
Titan Text Premier	32K tokens	RAG, Agents, enterprise	💰💰💰 Medium-high	❌ Don’t need Bedrock integration
Claude 3 Haiku	200K tokens	Fast, economical	💰💰 Medium	❌ Maximum capability needed
Claude 3 Sonnet	200K tokens	Capability/cost balance	💰💰💰 Medium-high	❌ Very tight budget
Claude 3 Opus	200K tokens	Maximum capability	💰💰💰💰 Most expensive	❌ Simple tasks (overkill)
Llama 2 70B	4K tokens	Open-source, customizable	💰💰 Medium	❌ Long docs, need support
Stable Diffusion XL	N/A (image)	Artistic images	💰💰💰 Per image	❌ Readable text in images
Titan Image Generator	N/A (image)	Readable text, watermarks	💰💰💰 Per image	❌ Very creative art
Titan Text Embeddings V2	8K tokens	Search, RAG, vectors	💰 Cheap	❌ Doesn’t generate text
Titan Multimodal Embeddings	N/A	Multimodal search	💰💰 Medium	❌ Doesn’t generate anything

---

GOLDEN RULES FOR THE EXAM:

✅ ALWAYS CHOOSE THESE IF…

1. Question mentions “Bedrock Knowledge Bases” or “Bedrock Agents” → Titan Text Premier (optimized for this)

2. Question says “semantic search”, “find similar”, “RAG” → Titan Text Embeddings V2

3. Question says “long documents”, “200+ pages”, “entire book” → Claude 3 (200K context)

4. Question says “budget-constrained”, “high volume”, “simple task” → Titan Text Lite

5. Question says “readable text in image”, “signs”, “posters” → Titan Image Generator

6. Question mentions “search images by photo or text” → Titan Multimodal Embeddings

❌ NEVER CHOOSE THESE:

1. GPT-4, ChatGPT, DALL-E → Not in Bedrock
2. Embeddings model to generate text → Embeddings DON’T generate
3. Titan Text Lite for complex analysis → Too simple
4. Claude Opus for simple classification → Overkill (too expensive)

---

QUICK DECISION DIAGRAM:

What do you need to do?
         |
         ├─ Generate TEXT
         |      ├─ Long documents? → Claude 3
         |      ├─ Bedrock Agents/KB integration? → Titan Text Premier
         |      ├─ Simple task + cheap? → Titan Text Lite
         |      └─ General use? → Titan Text Express
         |
         ├─ SEARCH for similar things
         |      ├─ Text only? → Titan Text Embeddings V2
         |      └─ Text + images? → Titan Multimodal Embeddings
         |
         ├─ Generate IMAGES
         |      ├─ With readable text? → Titan Image Generator
         |      └─ Art/creativity? → Stable Diffusion XL
         |
         └─ ANALYZE images → Claude 3 (vision)

---

PRACTICE: Which Model Would You Use?

Scenario 1: “E-commerce with 500K products. Customers search by typing ‘comfortable red shoes for walking’.”

• Answer: Titan Text Embeddings V2 - Semantic search in large catalog

Scenario 2: “Generate promotional posters with readable prices and product names.”

• Answer: Titan Image Generator - Only one that does readable text well

Scenario 3: “Analyze 150-page contracts to identify risk clauses.”

• Answer: Claude 3 Sonnet/Opus - 200K context + complex reasoning

Scenario 4: “Classify 1 million emails into spam/not-spam. Limited budget.”

• Answer: Titan Text Lite - Simple task + high volume + cheap

---