Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.bread.com.ai/llms.txt

Use this file to discover all available pages before exploring further.

Multi-target baking lets you bake entire knowledge bases worth of prompts into a single model. Each piece of knowledge (a document, fact, procedure) becomes a target, and you control how much each piece influences the final model through weighted datasets.
Completed your first bake? Multi-target baking builds on that foundation.

​
Bake Any Knowledge Into Model Weights

Any information you can express as a prompt can be baked into a language model:
  • Behaviors: “Act like Yoda”, “Be empathetic”, “Write concisely”
  • Facts: Product documentation, company policies, legal knowledge
  • Procedures: Troubleshooting guides, how-to instructions, workflows
  • Expertise: Domain knowledge, technical specifications, best practices
Multi-target baking lets you bake thousands of prompts at once.

​
How does RAG Compare to Baking?

Traditional RAG

Architecture:
  1. 1/ User asks question
  2. 2/ Query vector database
  3. 3/ Retrieve relevant docs
  4. 4/ Inject into prompt context
  5. 5/ Generate response
Overhead:
  • +200ms retrieval latency
  • +500 tokens per request
  • Vector DB & inference
  • Embedding generation

Multi-Target Baking

Architecture:
  1. 1/ User asks question
  2. 2/ Generate response



Overhead:
  • 0ms (no retrieval)
  • 0 context tokens
  • Model only
  • No embeddings needed

​
Benefits Over RAG Systems

  • Zero retrieval latency: No database queries at inference time
  • Zero context tokens: Knowledge lives in weights, not prompts
  • Simpler architecture: No vector databases, embeddings, or retrieval logic
  • Offline operation: Model works without external dependencies
  • Weighted knowledge: Prioritize important information during training
  • Consistent responses: Deterministic knowledge, not dependent on retrieval quality

​
Example: Baking Apple Product Knowledge

Traditional RAG systems store product information in a vector database, retrieve relevant docs at runtime, and inject them into prompts. Multi-target baking bakes the knowledge directly into weights.

​
Your Knowledge Base

Imagine you’re building an Apple support chatbot. Your knowledge base contains: Document 1: iPhone Support 
iPhone Troubleshooting:

- iPhone won't turn on: Hold Side + Volume Down for 10 seconds
- Battery drains quickly: In Settings, disable background refresh
- Can't connect to WiFi: Reset network settings in General → Reset
Document 2: Mac Support 
Mac Troubleshooting:

- Mac won't start: Press and hold power button for 10 seconds
- Running slow: Check Activity Monitor for memory usage & clear cache
- Bluetooth issues: Turn Bluetooth off/on, remove and re-pair device
Document 3: AirPods Support 
AirPods Troubleshooting:

- Won't connect: Put in case, hold button on back for 15 seconds
- Audio cutting out: Clean AirPods speakers, check Bluetooth connection
- Battery drains fast: Disable automatic ear detection in Bluetooth settings
Instead of RAG retrieval, bake each support document as a separate target.

​
Complete Multi-Target Workflow in 6 Steps

1

Create Repository & Shared Student Prompt

Set up your repository and the shared empty student prompt:
from aibread import Bread

client = Bread()

# Create repository
client.repo.set(repo_name="apple_support_agent")

# Shared student prompt (empty for always-on)
client.prompts.set(
    prompt_name="empty_student_prompt",
    repo_name="apple_support_agent",
    messages=[{"role": "system", "content": ""}]
)
In this example, we use a shared empty student prompt since we wish that this model always responds as an Apple customer support agent.
2

Create iPhone Support Target

Bake iPhone troubleshooting knowledge:
# iPhone Support Knowledge
client.prompts.set(
    prompt_name="iphone_teacher_prompt",
    repo_name="apple_support_agent",
    messages=[{
        "role": "system",
        "content": """
        You are an Apple iPhone support expert. You know:
        
        iPhone Troubleshooting:
        - iPhone won't turn on: Hold Side + Volume Down for 10 seconds
        - Battery drains quickly: In Settings, disable background refresh
        - Can't connect to WiFi: Reset network settings in General → Reset
        """
    }]
)

client.targets.set(
    target_name="iphone_support",
    repo_name="apple_support_agent",
    template="default",
    overrides={
        "generators": [
            {
                "type": "hardcoded",
                "numq": 3,
                "questions": [
                    "My iPhone won't turn on",
                    "Why is my iPhone battery draining so fast?",
                    "My iPhone screen is frozen"
                ]
            },
            {
                "type": "oneshot_qs",  # LLM-generated stim prompts
                "numq": 100,
                "model": "claude-sonnet-4-5-20250929",
                "temperature": 0.6
            }
        ],
        "teacher_prompt": "iphone_teacher_prompt",
        "student_prompt": "empty_student_prompt"
    }
)

# Generate stim data
client.targets.stim.run(
    target_name="iphone_support",
    repo_name="apple_support_agent"
)
# Once stim complete, generate rollout data
client.targets.rollout.run(
    target_name="iphone_support",
    repo_name="apple_support_agent"
)
3

Create Mac Support Target

Bake Mac troubleshooting knowledge:
# Mac Support Knowledge
client.prompts.set(
    prompt_name="mac_teacher_prompt",
    repo_name="apple_support_agent",
    messages=[{
        "role": "system",
        "content": """
        You are an Apple Mac support expert. You know:
        
        Mac Troubleshooting:
        - Mac won't start: Press and hold power button for 10 seconds
        - Running slow: Check Activity Monitor for memory usage & clear cache
        - Bluetooth issues: Turn Bluetooth off/on, remove and re-pair device
        """
    }]
)

client.targets.set(
    target_name="mac_support",
    repo_name="apple_support_agent",
    template="default",
    overrides={
        "generators": [
            {
                "type": "hardcoded",
                "numq": 3,
                "questions": [
                    "My Mac won't start up",
                    "Why is my Mac running so slow?",
                    "My Mac app keeps crashing"
                ]
            },
            {
                "type": "oneshot_qs",  # LLM-generated stim prompts
                "numq": 100,
                "model": "claude-sonnet-4-5-20250929",
                "temperature": 0.6
            }
        ],
        "teacher_prompt": "mac_teacher_prompt",
        "student_prompt": "empty_student_prompt"
    }
)

client.targets.stim.run(
    target_name="mac_support",
    repo_name="apple_support_agent"
)
client.targets.rollout.run(
    target_name="mac_support",
    repo_name="apple_support_agent"
)
4

Create AirPods Support Target

Bake AirPods troubleshooting knowledge:
# AirPods Support Knowledge
client.prompts.set(
    prompt_name="airpods_teacher_prompt",
    repo_name="apple_support_agent",
    messages=[{
        "role": "system",
        "content": """
        You are an Apple AirPods support expert. You know:
        
        AirPods Troubleshooting:
        - Won't connect: Put in case, hold button on back for 15 seconds
        - Audio cutting out: Clean AirPods speakers, check Bluetooth connection
        - Battery drains fast: Disable automatic ear detection in Bluetooth settings
        """
    }]
)

client.targets.set(
    target_name="airpods_support",
    repo_name="apple_support_agent",
    template="default",
    overrides={
        "generators": [
            {
                "type": "hardcoded",
                "numq": 3,
                "questions": [
                    "My AirPods won't connect to my iPhone",
                    "Why does the audio keep cutting out?",
                    "One of my AirPods isn't working"
                ]
            },
            {
                "type": "oneshot_qs",  # LLM-generated stim prompts
                "numq": 100,
                "model": "claude-sonnet-4-5-20250929",
                "temperature": 0.6
            }
        ],
        "teacher_prompt": "airpods_teacher_prompt",
        "student_prompt": "empty_student_prompt"
    }
)

client.targets.stim.run(
    target_name="airpods_support",
    repo_name="apple_support_agent"
)
client.targets.rollout.run(
    target_name="airpods_support",
    repo_name="apple_support_agent"
)
Each document from your knowledge base becomes a target. Scale to dozens or hundreds of targets - each Apple product, feature, or troubleshooting guide would get its own target.
5

Combine Targets & Configure Bake

Combine targets & specify bake config (in this example, set equal weightage to all baked prompts):
bake = client.bakes.set(
    bake_name="apple_support_agent_v1",
    repo_name="apple_support_agent",
    template="default",
    overrides={
        "datasets": [
            {
                "target": "iphone_support",
                "weight": 0.33
            },
            {
                "target": "mac_support",
                "weight": 0.33
            },
            {
                "target": "airpods_support",
                "weight": 0.33
            }
        ]
    }
)
6

Run Training

Start the multi-target bake:
client.bakes.run(
    bake_name="apple_support_agent_v1",
    repo_name="apple_support_agent"
)

​
RAG vs Baked Knowledge

Traditional RAGMulti-Target Baking
User asks: “My iPhone won’t turn on”

  1. Query vector database
  2. Retrieve iPhone doc (200ms)
  3. Inject into prompt (500 tokens)
  4. Generate response

Latency: +200ms
Tokens: +500 context tokens
Infrastructure: Vector DB + embeddings		User asks: “My iPhone won’t turn on”

  1. Generate response (knowledge in weights)





Latency: 0ms overhead
Tokens: 0 context tokens
Infrastructure: Model only
The model just “knows” iPhone troubleshooting. It’s baked into the weights, not retrieved.

​
Best Practices

Use good generators for each target to create diverse, high-quality training examples.
Start with 2 targets, validate results, then add more. Easier to debug with fewer targets.

​
Common Use Cases for Knowledge Baking

​
Replace Product Documentation RAG

Bake your entire product docs into the model:
  • Target 1: “Product Overview”
  • Target 2: “Feature Documentation”
  • Target 3: “API Reference”
  • Target 4: “Troubleshooting”
  • Target 5: “FAQs”
Result: Customer support chatbot with instant, zero-latency documentation access.

​
Bake Company Policies

Eliminate policy manual lookups:
  • Target 1: “HR Policies”
  • Target 2: “Code of Conduct”
  • Target 3: “Benefits Information”
  • Target 4: “Onboarding Procedures”
Result: HR assistant that knows all corporate policies.

​
Bake Educational Content

Create subject-matter expert models:
  • Target 1: “Biology Chapter 1: Cells”
  • Target 2: “Biology Chapter 2: Genetics”
  • Target 3: “Biology Chapter 3: Evolution”
  • Target 4: “Biology Chapter 4: Ecosystems”
  • Target 5: “Biology Chapter 5: Anatomy”
Result: Tutoring model with textbook knowledge baked in. Build specialized legal assistants:
  • Target 1: “Contract Law Basics”
  • Target 2: “Intellectual Property Guide”
  • Target 3: “Employment Law”
Result: Legal research assistant without needing legal database queries.

​
Next Steps

Iterative Baking

Refine multi-target models with additional bakes

Production Patterns

Add polling and error handling for production

Bake Configuration

Customize hyperparameters for multi-target training

Generators

Generate diverse training data for each target