Skip to contents

foundryR

foundryR is a tibble-native R client for research and measurement work with Microsoft Foundry from data frames: structured extraction, agreement checks, batch annotation, embeddings, cloud evaluations, and content safety.

Calls to Azure in the examples show output recorded from a live run, and setup code is shown but not run.

Measurement

The package is built for workflows where model output becomes data you have to inspect, join, and defend. This example codes a few course comments, then compares the model labels with hand labels.

comments <- data.frame(
  comment = c(
    "The examples made the statistics much easier to understand.",
    "The labs moved too fast and the instructions were unclear.",
    "The course was fine, but I wanted more feedback."
  ),
  hand_label = c("positive", "negative", "neutral")
)

schema <- foundry_schema(
  sentiment = schema_enum(c("positive", "negative", "neutral")),
  theme = schema_string("A short theme label.")
)

coded <- foundry_extract(comments, text_col = "comment", schema = schema)
coded[, c("hand_label", "sentiment", "theme")]
#> # A tibble: 3 × 3
#>   hand_label sentiment theme
#>   <chr>      <chr>     <chr>
#> 1 positive   positive  clarity
#> 2 negative   negative  instruction_clarity
#> 3 neutral    neutral   need for more feedback

foundry_agreement(coded, estimate = "sentiment", truth = "hand_label")
#> # A tibble: 6 × 3
#>   metric             value     n
#>   <chr>              <dbl> <int>
#> 1 accuracy               1     3
#> 2 precision_macro        1     3
#> 3 recall_macro           1     3
#> 4 f1_macro               1     3
#> 5 cohen_kappa            1     3
#> 6 krippendorff_alpha     1     3

The free-text theme field has no fixed label set, so its values can vary in form from row to row. Give any field you plan to count an enum, as sentiment has.

Agreement metrics describe how the model labels compare with the reference labels. They do not prove the reference labels are correct. Three comments are enough to show the output but not to validate the model. A real check codes a random sample of your data, large enough to put an interval on agreement, as in From text to defensible estimates.

Features

  • Strict JSON Schema extraction with one row per input and metadata columns for audit trails.
  • Agreement and consistency helpers for measurement workflows with human labels.
  • Batch annotation through Azure Files and Batch APIs when you need lower-cost asynchronous runs.
  • Embeddings and pairwise similarity for search, semantic groups, duplicate checks, and tidymodels recipes.
  • Microsoft Foundry evaluation runs for models, agents, stored responses, and built-in evaluators.
  • Azure AI Content Safety helpers for moderation, groundedness, prompt shields, blocklists, and protected material checks.
  • Responses API support for stateful turns, tool calls, web search, token accounting, and raw response capture.
  • Audio, image, and other modality helpers where the Microsoft Foundry service exposes them.

Installation

Install the released version from CRAN:

install.packages("foundryR")

Install the development version from GitHub:

# install.packages("pak")
pak::pak("farach/foundryR")

Quick start 

library(foundryR)
foundry_set_endpoint(Sys.getenv("AZURE_FOUNDRY_ENDPOINT"), store = TRUE)
foundry_set_key("your-api-key", store = TRUE)
foundry_check_setup()

For Microsoft Entra ID, replace the key line with a refreshable token provider:

Set AZURE_FOUNDRY_MODEL and AZURE_FOUNDRY_EMBED_MODEL to your deployment names, or pass deployment names through model =. A deployment can be named course-coder even when it runs base model gpt-5-nano.

Streaming is an intentional scope choice. foundryR focuses on reproducible, tibble-returning analytical workflows; use ellmer when an interactive streaming chat interface is the main product.

When to use foundryR

The ellmer reference index lists provider chat constructors, stream helpers, tool definitions, structured-data type specifications, and batch chat helpers. It does not list embedding helpers, Azure Content Safety helpers, tidymodels recipe steps, or Microsoft Foundry cloud evaluation helpers.

Need foundryR ellmer
Work tied to Microsoft Foundry resource and project APIs Broad data-frame client for Foundry APIs Azure OpenAI chat provider, without the broader Foundry data-plane coverage
Structured extraction into analysis rows foundry_extract() returns tibbles with schema fields and metadata Chat objects can extract structured data with ellmer type specifications
Reuse ellmer type specifications as_foundry_schema() converts ellmer::type_object() specifications Defines the type specifications
Agreement checks for model labels foundry_agreement() and related measurement helpers No agreement metrics in the reference index
Batch annotation on Azure Uses Azure Files and Batch APIs Batch chat helpers where the selected provider supports them, not Azure Files and Batch APIs
Embeddings in data frames foundry_embed(), foundry_embed_batch(), and foundry_similarity() No embedding helper in the reference index
Embeddings in tidymodels recipes step_foundry_embed() No recipe step in the reference index
Azure AI Content Safety Moderation, groundedness, shields, blocklists, image checks, and protected material No Content Safety helpers in the reference index
Foundry cloud evaluations Model, agent, stored-response, and built-in evaluator workflows No Microsoft Foundry evaluation helpers in the reference index
Provider-portable chat No Yes, through chat_*() providers
Interactive streaming chat No Yes, through streaming helpers and Chat methods
Chat-first tool calling Basic Responses API tool loop Yes, through chat tools

Use foundryR when the result needs to live in a data frame, use Microsoft Foundry APIs, or feed a measurement workflow. Use ellmer when provider-portable chat, interactive streaming, or a chat-first agent interface is the main need.

For retrieval-augmented generation, look at ragnar, from the ellmer team. It has embed_azure_openai(), document stores, and retrieval functions. foundryR’s embedding functions return tibbles for analysis, similarity checks, and tidymodels recipes.

Articles

License

MIT