Skip to main content
Prompt caching allows LLM providers to cache parts of your prompt that don’t change between requests, significantly reducing costs and latency for repeated queries.

​
How Prompt Caching Works

When you send a request to an LLM, providers like Anthropic can cache specific message segments. On subsequent requests with the same cached content, you:
  • Pay reduced rates for cached tokens (often 90% cheaper)
  • Experience faster response times (cached content is pre-processed)
  • Reduce computational overhead
Providers supporting prompt caching:
  • Anthropic: Claude models with cache_control metadata
  • OpenAI: GPT models with automatic caching
  • Google AI: Gemini models with caching support
  • Vertex AI: Google Cloud models with caching
AWS Bedrock does not currently report cached token usage, though some underlying models may use caching internally.

​
Enabling Prompt Caching in BAML

Caching is configured using role metadata in your BAML prompts. Here’s how to enable it:

​
Step 1: Configure Client to Allow Role Metadata

Add allowed_role_metadata to your client configuration:
main.baml
client<llm> AnthropicClient {
  provider "anthropic"
  options {
    model "claude-sonnet-4-5-20250929"
    allowed_role_metadata ["cache_control"]
  }
}

​
Step 2: Add Cache Control to Messages

Use the _.role() function to add cache control metadata:
main.baml
function AnalyzeBook(book: string, question: string) -> string {
  client AnthropicClient
  prompt #"
    {{ _.role("user") }}
    {{ book }}
    
    {{ _.role("user", cache_control={"type": "ephemeral"}) }}
    {{ question }}
  "#
}
This creates two user messages:
  1. First message: Contains the book content (will be cached)
  2. Second message: Contains the question (marked with cache control)
The cache control metadata tells the provider that everything up to this point should be cached. Place it after the content you want to cache.

​
Real-World Examples

​
Analyzing Large Documents

Cache document content while varying questions:
main.baml
class Analysis {
  themes string[]
  characters string[]
  summary string
}

function AnalyzeLiterature(document: string, focus: string) -> Analysis {
  client AnthropicClient
  prompt #"
    {{ _.role("user") }}
    Here is the complete text to analyze:
    
    {{ document }}
    
    {{ _.role("user", cache_control={"type": "ephemeral"}) }}
    Focus your analysis on: {{ focus }}
    
    {{ ctx.output_format }}
  "#
}
Usage: 
from baml_client import b

# First call - caches document
result1 = await b.AnalyzeLiterature(
    document=pride_and_prejudice_text,
    focus="character development"
)

# Second call - reuses cached document
result2 = await b.AnalyzeLiterature(
    document=pride_and_prejudice_text,  # Same document
    focus="social commentary"  # Different focus
)

​
Multi-Turn Conversations with Context

Cache conversation history and system context:
main.baml
class Message {
  role string
  content string
}

class Response {
  answer string
  confidence float
}

function ChatWithContext(
  system_context: string,
  conversation_history: Message[],
  user_message: string
) -> Response {
  client AnthropicClient
  prompt #"
    {{ _.role("system") }}
    {{ system_context }}
    
    {{ _.role("user") }}
    Previous conversation:
    {% for msg in conversation_history %}
    {{ msg.role }}: {{ msg.content }}
    {% endfor %}
    
    {{ _.role("user", cache_control={"type": "ephemeral"}) }}
    {{ user_message }}
    
    {{ ctx.output_format }}
  "#
}

​
RAG with Cached Context

Cache retrieved documents while varying queries:
main.baml
class DocumentChunk {
  title string
  content string
}

class Answer {
  response string
  sources int[]  // Indices of relevant chunks
  confidence float
}

function AnswerFromDocs(
  docs: DocumentChunk[],
  query: string
) -> Answer {
  client AnthropicClient
  prompt #"
    {{ _.role("user") }}
    Reference Documents:
    {% for doc in docs %}
    [{{ loop.index }}] {{ doc.title }}
    {{ doc.content }}
    
    {% endfor %}
    
    {{ _.role("user", cache_control={"type": "ephemeral"}) }}
    Query: {{ query }}
    
    Provide an answer based on the reference documents above.
    
    {{ ctx.output_format }}
  "#
}

​
Monitoring Cache Performance

Use Collectors to track cache hits and savings: 
from baml_client import b
from baml_py import Collector

async def analyze_with_caching():
    collector = Collector(name="cache-monitor")
    
    # First call
    await b.AnalyzeLiterature(
        document=large_doc,
        focus="themes",
        baml_options={"collector": collector}
    )
    
    # Second call with same doc
    await b.AnalyzeLiterature(
        document=large_doc,
        focus="characters",
        baml_options={"collector": collector}
    )
    
    # Analyze cache performance
    for i, log in enumerate(collector.logs):
        print(f"\nCall {i + 1}:")
        print(f"  Input tokens: {log.usage.input_tokens}")
        print(f"  Cached tokens: {log.usage.cached_input_tokens or 0}")
        print(f"  Output tokens: {log.usage.output_tokens}")
        
        # Calculate savings
        if log.usage.cached_input_tokens:
            cache_pct = (log.usage.cached_input_tokens / log.usage.input_tokens) * 100
            print(f"  Cache hit rate: {cache_pct:.1f}%")

​
Verifying Cache Requests

Use the VSCode Playground to verify your cache configuration:
  1. Open your BAML function in VSCode
  2. Run it in the Playground
  3. Switch from “Prompt Review” to “Raw cURL” view
  4. Verify the cache_control metadata is present in the request
Example output: 
{
  "model": "claude-sonnet-4-5-20250929",
  "messages": [
    {
      "role": "user",
      "content": "<document content>"
    },
    {
      "role": "user",
      "content": "<question>",
      "cache_control": { "type": "ephemeral" }
    }
  ]
}

​
Best Practices

  1. Cache Large, Static Content: Cache documentation, large documents, or system prompts that don’t change
  2. Place Cache Markers Strategically: Put cache_control after the content you want cached, not before
  3. Monitor Cache Effectiveness: Use Collectors to track cache hit rates and cost savings
  4. Consider TTL: Anthropic’s ephemeral cache lasts ~5 minutes; plan request timing accordingly
  5. Balance Cache Size: Caching works best with substantial content (1000+ tokens)
  6. Provider Differences: Different providers have different caching mechanisms - test thoroughly

​
Cost Savings Example

For Anthropic Claude Sonnet 4.5:
  • Regular input tokens: $3.00 per million
  • Cached input tokens: $0.30 per million (90% cheaper)
If analyzing a 50,000 token document with 10 different questions:
  • Without caching: 500,000 tokens × 3.00/M=3.00/M = 1.50
  • With caching: 50,000 × 3.00/M+450,000×3.00/M + 450,000 × 0.30/M = 0.15+0.15 + 0.135 = $0.285
  • Savings: $1.215 (81% reduction)

​
Limitations

  1. Cache Duration: Ephemeral caches expire after ~5 minutes of inactivity
  2. Provider-Specific: Implementation varies by provider
  3. Metadata Requirements: Must configure allowed_role_metadata for each client
  4. No Cross-Request Guarantees: Cache hits aren’t guaranteed across different sessions

Build docs developers (and LLMs) love

Get started for freeTalk to us