foundryR vs ellmer

In foundryR: Data Frame Workflows for 'Microsoft Foundry'

# The Azure-backed example on this page runs against recorded, credential-free
# fixtures (recorded once with data-raw/record-doc-outputs.R and committed under
# vignettes/foundryr-vs-ellmer/). The ellmer -> foundryR schema conversion is a
# pure-local call and always runs when ellmer is installed, so its real output is
# shown everywhere. Nothing here is fabricated.
fixture_dir <- "foundryr-vs-ellmer"
recording <- nzchar(Sys.getenv("FOUNDRY_RECORD_DOCS"))
have_fixtures <- dir.exists(fixture_dir) && length(list.files(fixture_dir)) > 0
run_api <- requireNamespace("httptest2", quietly = TRUE) &&
  (recording || have_fixtures)
have_ellmer <- requireNamespace("ellmer", quietly = TRUE)

# Attach foundryR before start_vignette(): httptest2 only sources the package's
# inst/httptest2/start-vignette.R (which sets replay placeholders) from attached
# packages.
library(foundryR)

if (run_api) {
  httptest2::start_vignette(fixture_dir)
}

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>",
  eval = run_api
)

library(foundryR)

foundryR and ellmer both make language-model work possible from R. They are not substitutes for every use case. ellmer is the better fit for provider-portable chat. foundryR is the better fit when the work is tied to Azure AI Foundry and the output needs to become data. The two are complementary: you can describe a structure once with ellmer's type system and hand it to foundryR for strict, tibble-shaped extraction.

Comparison

| Question | Use foundryR | Use ellmer | | --- | --- | --- | | Are you committed to Azure AI Foundry? | Yes | Sometimes | | Do you need Azure AI Content Safety? | Yes | No | | Do you need Azure's Files and Batch APIs? | Yes | No | | Do you need strict schema-constrained extraction as tibbles? | Yes | Sometimes | | Do you need embeddings inside a dataframe workflow? | Yes | Sometimes | | Do you need step_foundry_embed() in a tidymodels recipe? | Yes | No | | Do you need multi-provider chat? | No | Yes | | Do you need interactive streaming chat? | No | Yes | | Do you need a chat-first tool-calling agent interface? | Basic Responses API loop | Yes |

Where foundryR fits

Use foundryR when your work starts or ends in a dataframe. Common examples are:

• Coding open-ended survey responses with a strict JSON Schema.
• Running low-cost batch annotation jobs through Azure's Batch API.
• Embedding documents for clustering, semantic search, or near-duplicate checks.
• Adding embeddings to tidymodels recipes.
• Moderating user text, detecting prompt injection, or checking groundedness.
• Keeping model metadata, citations, tool calls, and token use in tibble columns.

Azure's own Batch API matters here. ellmer supports batch workflows for native OpenAI and Anthropic endpoints, but it does not target Azure's Files and Batch API surface.

Where ellmer fits

Use ellmer when chat is the product interface. It gives you a provider-neutral chat abstraction, interactive streaming, and a mature chat-first tool-calling workflow. That is the right shape for assistants, notebooks where you want token streaming, or applications that may move between providers.

foundryR deliberately does not implement streaming. For streaming chat in R, use ellmer.

Interop: reuse an ellmer type as a foundryR schema

If you already describe your data with ellmer's type system, you do not have to rewrite it. as_foundry_schema() converts an ellmer::type_object() into the strict JSON Schema that foundry_extract() expects. This conversion is a local operation -- no network, no credentials -- so its output is shown here directly.

library(ellmer)

# Describe the structure you want with ellmer's type system.
sentiment_spec <- type_object(
  sentiment = type_enum(
    c("positive", "negative", "neutral"),
    description = "Overall sentiment of the response."
  ),
  theme = type_string("A short theme label for the response.")
)

# Convert it into a foundryR schema for strict extraction.
sentiment_schema <- as_foundry_schema(sentiment_spec)
str(sentiment_schema)

Extraction with foundryR

Hand the converted schema to foundry_extract(). foundryR sends each input through the Responses API with strict decoding and returns one tidy row per input -- the model's output has already become data.

foundry_extract(
  c("The lesson was clear.", "I wanted more examples."),
  schema = sentiment_schema
)

You can also build the same schema natively with foundryR's foundry_schema() and schema_*() constructors when you would rather not depend on ellmer. See vignette("responses-api") for the native path, and vignette("content-safety") for using Azure Content Safety as a pipeline gate.

The practical rule is simple: use ellmer when you need the best R chat client, and use foundryR when you need Azure AI Foundry results as data -- and use as_foundry_schema() when you want both.

if (run_api) {
  httptest2::end_vignette()
}

foundryR documentation built on Sept. 25, 2026, 1:10 a.m.