Skip to contents

Migrating to ellmer Providers

ellmer-migration.Rmd

Migrating to ellmer Providers

The llmr package now supports using ellmer as the underlying provider implementation. This brings significant benefits including:

  • 15+ provider support (vs 2 with legacy providers)
  • Advanced features: streaming, async operations, structured data extraction
  • Better performance and reliability
  • Professional maintenance by the tidyverse team

Quick Migration Guide

Old API (Deprecated) 

library(llmr)

# Old way - deprecated
provider <- new_anthropic()
agent <- new_agent("assistant", provider)

# Add tools
weather_tool <- mcpr::new_tool(
  name = "weather",
  description = "Get weather forecast",
  input_schema = mcpr::schema(
    properties = mcpr::properties(
      location = mcpr::property_string("Location", required = TRUE)
    )
  ),
  handler = function(params) {
    paste("Weather in", params$location, "is sunny")
  }
)

add_tool(agent, weather_tool)
library(llmr)

# New way - direct ellmer usage
agent <- new_agent("assistant", ellmer::chat_anthropic())

# Add tools (same interface!)
add_tool(agent, weather_tool)

# Or use ellmer tools directly
ellmer_tool <- ellmer::tool(
  function(location) paste("Weather in", location, "is sunny"),
  "Get weather forecast",
  location = ellmer::type_string("Location to get weather for")
)

add_tool(agent, ellmer_tool)

Provider Migration

Anthropic 

# Before
provider <- new_anthropic() |>
  set_model("claude-3-sonnet-20240229") |>
  set_temperature(0.7) |>
  set_max_tokens(2048)
agent <- new_agent("assistant", provider)

# After - direct ellmer usage
agent <- new_agent("assistant", ellmer::chat_anthropic(
  model = "claude-3-sonnet-20240229",
  params = ellmer::params(
    temperature = 0.7,
    max_tokens = 2048
  )
))

OpenAI 

# Before
provider <- new_openai() |>
  set_model("gpt-4") |>
  set_temperature(0.5) |>
  set_max_tokens(1024)
agent <- new_agent("assistant", provider)

# After - direct ellmer usage
agent <- new_agent("assistant", ellmer::chat_openai(
  model = "gpt-4",
  params = ellmer::params(
    temperature = 0.5,
    max_tokens = 1024
  )
))

New Providers Available

With ellmer, you now have access to many more providers:

# Google Gemini
agent <- new_agent("gemini", ellmer::chat_google_gemini(model = "gemini-pro"))

# AWS Bedrock
agent <- new_agent("bedrock", ellmer::chat_aws_bedrock(model = "anthropic.claude-v2"))

# Ollama (local models)
agent <- new_agent("local", ellmer::chat_ollama(model = "llama2"))

# And many more...

Tool Migration

mcpr Tools

Your existing mcpr tools work seamlessly:

# This works with both old and new providers
calculator_tool <- mcpr::new_tool(
  name = "calculate",
  description = "Perform calculations",
  input_schema = mcpr::schema(
    properties = mcpr::properties(
      expression = mcpr::property_string("Math expression", required = TRUE)
    )
  ),
  handler = function(params) {
    eval(parse(text = params$expression))
  }
)

add_tool(agent, calculator_tool)

ellmer Tools

You can also use ellmer’s native tool system:

# ellmer tools have better type safety
calculator_tool <- ellmer::tool(
  function(expression) eval(parse(text = expression)),
  "Perform mathematical calculations",
  expression = ellmer::type_string("Mathematical expression to evaluate")
)

add_tool(agent, calculator_tool)

MCP Integration

MCP integration continues to work:

# Create MCP client
mcp_client <- mcpr::new_client(
  command = "Rscript",
  args = "path/to/server.R",
  name = "calculator"
)

# Register with agent (works with both provider types)
register_mcp(agent, mcp_client)

Advanced Features

Streaming (ellmer providers only) 

# ellmer providers support streaming
agent <- new_agent("assistant", ellmer::chat_anthropic())

# The underlying ellmer chat object supports streaming
chat <- agent$provider$chat
chat$chat("Tell me a story", echo = "output")

Structured Data Extraction 

# ellmer providers support structured data extraction
response <- agent$provider$chat$chat_structured(
  "Extract information about this person: John Doe, age 30, lives in NYC",
  type = ellmer::type_object(
    name = ellmer::type_string("Person's name"),
    age = ellmer::type_integer("Person's age"),
    city = ellmer::type_string("City where person lives")
  )
)

Async Operations 

# ellmer supports async operations
promise <- agent$provider$chat$chat_async("What is the capital of France?")
promises::then(promise, function(result) {
  cat("Response:", result, "\n")
})

Configuration Compatibility

Most configuration methods work the same:

# Configuration is done directly with ellmer chat objects
agent <- new_agent("assistant", ellmer::chat_anthropic(
  model = "claude-3-haiku-20240307",
  params = ellmer::params(
    temperature = 0.8,
    max_tokens = 1500
  ),
  system_prompt = "You are a helpful assistant."
))

Migration Timeline

  • Phase 1 (Current): Both APIs supported, deprecation warnings for old API
  • Phase 2 (6 months): Old API marked as deprecated with clear warnings
  • Phase 3 (12 months): Old API removed, ellmer becomes the only option

Benefits Summary

Feature Legacy Providers ellmer Providers
Providers 2 (Anthropic, OpenAI) 15+
Streaming ❌ ✅
Async ❌ ✅
Structured Data ❌ ✅
Tool Safety Basic Type-safe
Maintenance Manual Tidyverse
Performance Basic Optimized

Getting Help