Build an R Code-Generation Agent

Description

Constructs an LLM-powered agent for generating, debugging, explaining, or optimizing R code. **Two calling patterns are supported**:

• **Builder pattern** – omit 'user_input'. The function returns a reusable coder-agent closure. Call that closure with different queries whenever you need code help.

• **One-shot pattern** – provide 'user_input'. The function executes immediately and returns the result once.

Arguments

Details

The agent automatically retries failed LLM calls (with exponential back-off) and always returns a structured result.

Value

• If 'user_input' is 'NULL': a function (the coder-agent closure).

• Otherwise: a list with the fields

  input

  The original user query.

  llm_response

  The LLM output (or error message).

  system_prompt

  Prompt actually sent.

  success

  Logical; did the call succeed?

  attempts

  Number of attempts made.

Examples

## Not run: 
## ------------------------------------------------------------------
## 1)  Builder pattern – create a reusable coder agent
## ------------------------------------------------------------------
coder <- build_code_agent(
  llm       = my_llm_wrapper,   # your own wrapper around the LLM API
  max_tries = 3,
  backoff   = 2,
  verbose   = FALSE
)

# Use the agent multiple times
res1 <- coder("Write an R function that z-score–standardises all numeric columns.")
res2 <- coder("Explain what `%>%` does in tidyverse pipelines.")

## ------------------------------------------------------------------
## 2)  One-shot pattern – run a single request immediately
## ------------------------------------------------------------------
one_shot <- build_code_agent(
  llm        = my_llm_wrapper,
  user_input = "Create a ggplot2 bar chart of mpg by cyl in mtcars.",
  max_tries  = 3,
  backoff    = 2,
  verbose    = FALSE
)

## End(Not run)

Build a Data Cleaning Agent

Description

Constructs a multi-step agent workflow to recommend, generate, fix, execute, and explain robust R code for data cleaning tasks using LLMs and user-defined data.

Arguments

Value

A compiled graph-based cleaning agent function that accepts and mutates a state list.

Examples

## Not run: 
# 1) Load the data
data <- read.csv("tests/testthat/test-data/churn_data.csv")

# 2) Create the agent
data_cleaner_agent <- build_data_cleaning_agent(
  model = my_llm_wrapper,
  human_validation = FALSE,
  bypass_recommended_steps = FALSE,
  bypass_explain_code = FALSE,
  verbose = FALSE
)

# 3) Define the initial state
initial_state <- list(
  data_raw = data,
  user_instructions = "Don't remove outliers when cleaning the data.",
  max_retries = 3,
  retry_count = 0
)

# 4) Run the agent
final_state <- data_cleaner_agent(initial_state)

## End(Not run)

Build a Data Wrangling Agent

Description

Constructs a state graph-based agent that recommends, generates, executes, fixes, and explains data wrangling transformations based on user instructions and dataset structure. The resulting function handles list or single data frame inputs and produces a cleaned dataset.

Arguments

Value

A callable agent function that mutates a provided 'state' list by populating: - 'data_wrangled': the final cleaned data frame, - 'data_wrangler_function': the code used, - 'data_wrangler_error': any execution error (if occurred), - 'wrangling_report': LLM-generated explanation (if 'bypass_explain_code = FALSE')

Examples

## Not run: 
# 1) Simulate multiple data frames with a common ID
df1 <- data.frame(
  ID = c(1, 2, 3, 4),
  Name = c("John", "Jane", "Jim", "Jill"),
  stringsAsFactors = FALSE
)

df2 <- data.frame(
  ID = c(1, 2, 3, 4),
  Age = c(25, 30, 35, 40),
  stringsAsFactors = FALSE
)

df3 <- data.frame(
  ID = c(1, 2, 3, 4),
  Education = c("Bachelors", "Masters", "PhD", "MBA"),
  stringsAsFactors = FALSE
)

# 2) Combine into a list
data <- list(df1, df2, df3)

# 3) Create the agent
data_wrangling_agent <- build_data_wrangling_agent(
  model = my_llm_wrapper,
  human_validation = FALSE,
  bypass_recommended_steps = FALSE,
  bypass_explain_code = FALSE,
  verbose = FALSE
)

# 4) Define the initial state
initial_state <- list(
  data_raw = data,
  user_instructions = "Merge the data frames on the ID column.",
  max_retries = 3,
  retry_count = 0
)

# 5) Run the agent
final_state <- data_wrangling

## End(Not run)

Build a Document Summarizer Agent

Description

Creates an LLM-powered document summarization workflow that processes PDF, DOCX, PPTX, TXT, or plain text input and returns structured markdown summaries.

Usage

build_doc_summarizer_agent(
  llm,
  summary_template = NULL,
  chunk_size = 4000,
  overlap = 200,
  verbose = TRUE
)

Arguments

Value

A function that accepts file paths or text input and returns:

• summary - The generated markdown summary

• metadata - Document metadata if available

• chunks - Number of processing chunks used

• success - Logical indicating success

Examples

## Not run: 
# Build document summarizer agent
summarizer_agent <- build_doc_summarizer_agent(
  llm = my_llm_wrapper,
  summary_template = NULL,
  chunk_size = 4000,
  overlap = 200,
  verbose = FALSE
)

# Summarize document
final_state <- summarizer_agent("https://github.com/knowusuboaky/LLMAgentR/raw/main/\
tests/testthat/test-data/scrum.docx")

## End(Not run)

Build a Feature Engineering Agent

Description

Constructs a graph-based feature engineering agent that guides the process of: recommending, generating, executing, fixing, and explaining feature engineering code.

Arguments

Value

A callable agent function that executes feature engineering via a state graph.

Examples

## Not run: 
# 1) Load the data
data <- read.csv("tests/testthat/test-data/churn_data.csv")

# 2) Create the feature engineering agent
feature_engineering_agent <- build_feature_engineering_agent(
  model = my_llm_wrapper,
  human_validation = FALSE,
  bypass_recommended_steps = FALSE,
  bypass_explain_code = FALSE,
  verbose = TRUE
)

# 3) Define the initial state
initial_state <- list(
  data_raw = data,
  target_variable = "Churn",
  user_instructions = "Inspect the data. Make any new features and transformations
  that you think will be useful for predicting the target variable.",
  max_retries = 3,
  retry_count = 0
)

# 4) Run the agent
final_state <- feature_engineering_agent(initial_state)

## End(Not run)

Build a Time Series Forecasting Agent

Description

Constructs a state graph-based forecasting agent that: recommends forecasting steps, extracts parameters, generates code, executes the forecast using 'modeltime', fixes errors if needed, and explains the result. It leverages multiple models including Prophet, XGBoost, Random Forest, SVM, and Prophet Boost, and combines them in an ensemble.

Arguments

Value

A callable agent function that mutates the given 'state' list.

Examples

## Not run: 
# 2) Prepare the dataset
my_data <- walmart_sales_weekly

# 3) Create the forecasting agent
forecasting_agent <- build_forecasting_agent(
  model = my_llm_wrapper,
  bypass_recommended_steps = FALSE,
  bypass_explain_code = FALSE,
  mode = "dark", # dark or light
  line_width = 3,
  verbose = FALSE
)

# 4) Define the initial state
initial_state <- list(
  user_instructions = "Forecast sales for the next 30 days, using `id` as the grouping variable,
  a forecasting horizon of 30, and a confidence level of 90%.",
  data_raw = my_data
)

# 5) Run the agent
final_state <- forecasting_agent(initial_state)

## End(Not run)

Build an Interpreter Agent

Description

Constructs an LLM-powered agent that explains plots, tables, text, or other outputs for both technical and non-technical audiences.

Arguments

Details

**Two calling patterns**

• **Builder pattern** – omit code_output; a reusable interpreter-agent closure is returned.

• **One-shot pattern** – provide code_output; the function runs immediately and returns the interpretation.

Value

• If code_output is NULL: a function (closure).

• Otherwise: a list with

  prompt

  The full prompt sent to the LLM.

  interpretation

  The LLM’s explanation (or error).

  success

  Logical; did it succeed?

  attempts

  Number of attempts made.

Examples

## Not run: 
## 1) Builder pattern --------------------------------------------
interp <- build_interpreter_agent(llm = my_llm_wrapper, verbose = FALSE)

table_txt <- "
| Region | Sales | Profit |
| North  | 2000  | 300    |
| South  | 1500  | 250    |"

res1 <- interp(table_txt)
res2 <- interp("R² = 0.87 for the fitted model …")

## 2) One-shot pattern -------------------------------------------
build_interpreter_agent(
  llm         = my_llm_wrapper,
  code_output = table_txt,
  verbose     = FALSE
)

## End(Not run)

Build a Web Researcher Agent

Description

Constructs an LLM-powered research agent that performs web searches (via Tavily API) and generates structured responses based on search results. The agent handles different question types (general knowledge, comparisons, controversial topics) with appropriate response formats.

Arguments

Value

A function that accepts a user query string and returns a list with:

• query - The original research query.

• prompt - The full prompt sent to the LLM.

• response - The generated LLM response.

• search_results - Raw search results (if any were found).

• success - Logical indicating if research succeeded (both search and LLM).

Examples

## Not run: 
# Initialize researcher agent
researcher_agent <- build_researcher_agent(
  llm = my_llm_wrapper,
  tavily_search = NULL,
  system_prompt = NULL,
  max_results = 5,
  max_tries = 3,
  backoff = 2,
  verbose = FALSE
)

# Perform research
result <- researcher_agent("Who is Messi?")

## End(Not run)

Build a SQL Agent Graph

Description

This function constructs a full SQL database agent using a graph-based workflow. It supports step recommendation, SQL code generation, error handling, optional human review, and automatic explanation of the final code.

Arguments

Value

A compiled SQL agent function that runs via a state machine (graph execution).

Examples

## Not run: 
# 1) Connect to the database
conn <- DBI::dbConnect(RSQLite::SQLite(), "tests/testthat/test-data/northwind.db")

# 2) Create the SQL agent
sql_agent <- build_sql_agent(
  model                    = my_llm_wrapper,
  connection               = conn,
  human_validation         = FALSE,
  bypass_recommended_steps = FALSE,
  bypass_explain_code      = FALSE,
  verbose                  = FALSE
)

# 3) Define the initial state
initial_state <- list(
  user_instructions = "Identify the Regions (or Territories) with the highest
  CustomerCount and TotalSales.
  Return a table with columns: Region, CustomerCount, and TotalSales.
Hint: (UnitPrice × Quantity).",
  max_retries       = 3,
  retry_count       = 0
)

# 4) Run the agent
final_state <- sql_agent(initial_state)

## End(Not run)

Build Visualization Agent

Description

Creates a data visualization agent with configurable workflow steps.

Arguments

Value

A function that takes state and returns visualization results

Examples

## Not run: 
# 1) Load the data
data <- read.csv("tests/testthat/test-data/churn_data.csv")

# 2) Create the visualization agent
visualization_agent <- build_visualization_agent(
  model = my_llm_wrapper,
  human_validation = FALSE,
  bypass_recommended_steps = FALSE,
  bypass_explain_code = FALSE,
  verbose = FALSE
)

# 3) Define the initial state
initial_state <- list(
  data_raw = data,
  target_variable = "Churn",
  user_instructions = "Create a clean and visually appealing box plot to show
  the distribution of Monthly Charges across Churn categories.
  Use distinct colors for each Churn group,
  add clear axis labels, a legend, and a meaningful title.",
  max_retries = 3,
  retry_count = 0
)

# 4) Run the agent
final_state <- visualization_agent(initial_state)

## End(Not run)

Build a Weather Agent

Description

Constructs an LLM-powered weather assistant that fetches data from OpenWeatherMap and generates user-friendly reports. Handles location parsing, API calls, caching, and LLM-based summarization.

Arguments

Value

A list containing:

• success - Logical indicating if operation succeeded

• location - Cleaned location string

• weather_raw - Raw API response

• weather_formatted - Formatted weather string

• llm_response - Generated weather report

• timestamp - Time of response

• cache_hit - Logical indicating cache usage

• attempts - Number of tries made

Examples

## Not run: 
# Get weather information
weather_agent <- build_weather_agent(
  llm = my_llm_wrapper,
  location_query = "Tokyo, Japan",
  system_prompt = NULL,
  weather_api_key = NULL,
  units = "metric", # metric or imperial
  n_tries = 3,
  backoff = 2,
  endpoint_url = NULL,
  verbose = FALSE
)

## End(Not run)