Getting started with sastestgenerator

This proof of concept prepares prompts and drafts SAS macro validation tests. Start with the offline package workflow, then use live generation only with inputs permitted for OpenRouter and its selected model provider.

Restore and load

From the repository root in RStudio:

renv::restore()
pkgload::load_all(".")

The lockfile records R 4.6.0. Documentation tools are declared in Suggests and included in the explicit renv snapshot. An installed package can instead be loaded with library(sastestgenerator).

Prepare a prompt without a model call

This executable example writes a synthetic FRD and macro to temporary files. No API key or network call is needed.

library(sastestgenerator)

frd_path <- tempfile(fileext = ".txt")
macro_path <- tempfile(fileext = ".sas")
writeLines(
    "FRD-001: Set BASEFL to Y for BASELINE visits; otherwise set it to N.",
    frd_path
)
writeLines(c(
    "%macro derive_visit_flag(input=, output=);",
    "data &output.; set &input.;",
    "if VISIT = 'BASELINE' then BASEFL = 'Y'; else BASEFL = 'N';",
    "run; %mend derive_visit_flag;"
), macro_path)

frd <- read_frd(frd_path)
macro <- import_sas_macro(macro_path)
macro$macro_names

## [1] "derive_visit_flag"

macro$parameters

## [1] "input"  "output"

result <- generate_sas_tests(frd, macro)
cat(result$prompt)

## You are a Senior SAS Programmer and Clinical Programming Validation Expert.
## 
## Generate structured JSON containing test cases, sample data, SAS validation/test code,
## 
## and traceability for the supplied functional requirements and SAS macro.
## 
## Functional requirements:
## 
## FRD-001: Set BASEFL to Y for BASELINE visits; otherwise set it to N.
## 
## SAS macro:
## 
## %macro derive_visit_flag(input=, output=);
## data &output.; set &input.;
## if VISIT = 'BASELINE' then BASEFL = 'Y'; else BASEFL = 'N';
## run; %mend derive_visit_flag;

unlink(c(frd_path, macro_path))

The result contains prompt, requirements, macro, test_cases, and sas_code. The last two fields are empty in offline mode. The returned input objects contain paths and text; deleting temporary files does not erase the text retained in those objects.

From a source checkout, the supplied pair is examples/simple_frd.txt and examples/simple_visit_flag.sas. source("example.R") reads those files and checks the offline result.

Supply a generator

A generator is a function receiving one prompt string. Its return value is returned directly, with no automatic validation or wrapper. This small callback demonstrates the contract without calling a model:

prompt_length <- generate_sas_tests(
    frd, macro,
    generator = function(prompt) nchar(prompt)
)
stopifnot(prompt_length > 0L)

Request live output

Add OPENROUTER_API_KEY=your-key-here to the project .Renviron and restart R. The following code is shown for manual use and is not executed when building this guide:

live_result <- generate_sas_tests_openrouter(
    frd, macro,
    model = "gpt-4o-mini",
    temperature = 0,
    max_tokens = 4096
)
live_result$parsed

The model string is the current code default. Check provider support before using it. The call sends the full FRD and SAS source outside the application and may incur charges. A parsed response is a draft, not proof that tests are correct.

Use the app

shiny::runApp(".")

1. Paste a non-empty FRD.
2. Upload one .sas file.
3. Set the model, temperature, and maximum output tokens.
4. Select Generate test cases and inspect the displayed matrix.
5. Download Validation_Results.zip and review the workbook and both SAS files independently before use.

Input feedback highlights invalid fields; notifications report generation errors. Missing or malformed model output can fail normalization. Regenerate after changing inputs, because a previous result is not automatically cleared when inputs change. The app does not run SAS or require a recorded approval.

The generation contract explains the response, matrix, and download fields. The architecture maps the implementation boundaries.

Build documentation

devtools::document()
pkgdown::check_pkgdown()
pkgdown::build_site(preview = FALSE)

The website is generated in pkgdown-site/. No deployment is performed by these commands. There is no committed automated test suite; the offline example, documentation build, package check, and browser review are separate checks.