Structured data

When using an LLM to extract data from text or images, you can ask the chatbot to format it in JSON or any other format that you like. This works well most of the time, but there’s no guarantee that you’ll get the exact format you want. In particular, if you’re trying to get JSON, you’ll find that it’s typically surrounded in ```json, and you’ll occasionally get text that isn’t valid JSON. To avoid these problems, you can use a recent LLM feature: structured data (aka structured output). With structured data, you supply the type specification that defines the object structure you want and the LLM ensures that’s what you’ll get back.

library(ellmer)

Structured data basics

To extract structured data call $chat_structured() instead of $chat(). You’ll also need to define a type specification that describes the structure of the data that you want (more on that shortly). Here’s a simple example that extracts two specific values from a string:

chat <- chat_openai()
#> Using model = "gpt-5.4".
chat$chat_structured(
  "My name is Susan and I'm 13 years old",
  type = type_object(
    name = type_string(),
    age = type_number()
  )
)
#> $name
#> [1] "Susan"
#> 
#> $age
#> [1] 13

The same basic idea works with images too:

chat <- chat_openai()
#> Using model = "gpt-5.4".
chat$chat_structured(
  content_image_url("https://www.r-project.org/Rlogo.png"),
  type = type_object(
    primary_shape = type_string(),
    primary_colour = type_string()
  )
)
#> $primary_shape
#> [1] "stylized letter R with an oval ring"
#> 
#> $primary_colour
#> [1] "blue"

If you need to extract data from multiple prompts, you can use parallel_chat_structured(). It takes the same arguments as $chat_structured() with two exceptions: it needs a chat object since it’s a standalone function, not a method, and it can take a vector of prompts.

prompts <- list(
  "I go by Alex. 42 years on this planet and counting.",
  "Pleased to meet you! I'm Jamal, age 27.",
  "They call me Li Wei. Nineteen years young.",
  "Fatima here. Just celebrated my 35th birthday last week.",
  "The name's Robert - 51 years old and proud of it.",
  "Kwame here - just hit the big 5-0 this year."
)
type_person <- type_object(
  name = type_string(),
  age = type_number()
)
chat <- chat_openai()
#> Using model = "gpt-5.4".
parallel_chat_structured(chat, prompts, type = type_person)
#> # A tibble: 6 × 2
#>   name     age
#>   <chr>  <dbl>
#> 1 Alex      42
#> 2 Jamal     27
#> 3 Li Wei    19
#> 4 Fatima    35
#> 5 Robert    51
#> 6 Kwame     50

(Note that structured data extraction automatically disables tool calling, see below for details and workarounds.)

Data types

To extract structured data effectively, you need to understand how LLMs expect types to be defined, and how those types map to the R types you are familiar with.

Basics

To define your desired type specification (also known as a schema), you use the type_() functions. These are also used for tool calling (vignette("tool-calling")), so you might already be familiar with them.The type functions can be divided into three main groups:

Under the hood, these type specifications ensures that the LLM returns correctly structured JSON. But ellmer goes one step further and converts the JSON to the closest R analog. This means:

You can opt-out of this and get plain lists by setting convert = FALSE.

In addition to defining types, you need to provide the LLM with some information about what those types represent. This is the purpose of the first argument, description, a string that describes the data that you want. This is a good place to ask nicely for other attributes you’ll like the value to have (e.g. minimum or maximum values, date formats, …). There’s no guarantee that these requests will be honoured, but the LLM will try.

type_person3 <- type_object(
  "A person",
  name = type_string("Name"),
  age = type_integer("Age, in years."),
  hobbies = type_array(
    type_string(),
    "List of hobbies. Should be exclusive and brief.",
  )
)

Missing values

The type functions default to required = TRUE which means the LLM will try really hard to extract values for you, leading to hallucinations if the data doesn’t exist. Lets go back to our initial example extracting names and ages, and give it some inputs that don’t have names and/or ages.

no_match <- list(
  "I like apples",
  "What time is it?",
  "This cheese is 3 years old",
  "My name is Hadley."
)
parallel_chat_structured(chat, no_match, type = type_person)
#> # A tibble: 4 × 2
#>   name                                       age
#>   <chr>                                    <dbl>
#> 1 Apples                                       0
#> 2 I can’t tell the current time from here.     0
#> 3 This cheese                                  3
#> 4 Hadley                                       0

You can often avoid this problem by setting required = FALSE:

type_person <- type_object(
  name = type_string(required = FALSE),
  age = type_number(required = FALSE)
)
parallel_chat_structured(chat, no_match, type = type_person)
#> # A tibble: 4 × 2
#>   name     age
#>   <chr>  <dbl>
#> 1 <NA>      NA
#> 2 <NA>      NA
#> 3 <NA>       3
#> 4 Hadley    NA

In other cases, you may need to adjust your prompt as well. Either way, we strongly recommend that you include both positive and negative examples when testing your structured data extraction code.

Data frames

In most cases, you’ll get a data frame (well, tibble) because you are using parallel_chat_structured(), where each output row represents one input prompt. In other cases, you might have a more complex document where you want a data frame from a single prompt. For example, imagine that you want to extract some data about people from a table:

prompt <- r"(
* John Smith. Age: 30. Height: 180 cm. Weight: 80 kg.
* Jane Doe. Age: 25. Height: 5'5". Weight: 110 lb.
* Jose Rodriguez. Age: 40. Height: 190 cm. Weight: 90 kg.
* June Lee | Age: 35 | Height 175 cm | Weight: 70 kg
)"

You might be tempted to use a definition similar to R: an object (i.e., a named list) containing multiple arrays (i.e., vectors):

type_people <- type_object(
  name = type_array(type_string()),
  age = type_array(type_integer()),
  height = type_array(type_number("in m")),
  weight = type_array(type_number("in kg"))
)

chat <- chat_openai()
#> Using model = "gpt-5.4".
chat$chat_structured(prompt, type = type_people)
#> $name
#> [1] "John Smith"     "Jane Doe"       "Jose Rodriguez" "June Lee"      
#> 
#> $age
#> [1] 30 25 40 35
#> 
#> $height
#> [1] 1.80 1.65 1.90 1.75
#> 
#> $weight
#> [1] 80.0 49.9 90.0 70.0

This doesn’t work because there’s no constraint that each array should have the same length, and hence no way for ellmer to know that you really wanted a data frame. Instead, you’ll need to turn the data structure “inside out” and create an array of objects:

type_people <- type_array(
  type_object(
    name = type_string(),
    age = type_integer(),
    height = type_number("in m"),
    weight = type_number("in kg")
  )
)

chat <- chat_openai()
#> Using model = "gpt-5.4".
chat$chat_structured(prompt, type = type_people)
#> # A tibble: 4 × 4
#>   name             age height weight
#>   <chr>          <int>  <dbl>  <dbl>
#> 1 John Smith        30   1.8    80  
#> 2 Jane Doe          25   1.65   49.9
#> 3 Jose Rodriguez    40   1.9    90  
#> 4 June Lee          35   1.75   70

Now ellmer knows what you want and gives you a tibble.

If you’re familiar with the terms row-oriented and column-oriented data frames, this is the same idea. Since most languages don’t possess vectorisation like R, row-oriented data frames are more common.

Note that you’ll generally want to avoid nesting objects inside of objects as this will generate a data frame where each column is itself a data frame. You can use tidyr::unpack() to unpack these df-columns back into a regular flat data frame, but your life will be simpler if you re-consider the type.

Examples

The following examples, which are closely inspired by the Claude documentation, hint at some of the ways you can use structured data extraction.

Example 1: Article summarisation

text <- readLines(system.file(
  "examples/third-party-testing.txt",
  package = "ellmer"
))
# url <- "https://www.anthropic.com/news/third-party-testing"
# html <- rvest::read_html(url)
# text <- rvest::html_text2(rvest::html_element(html, "article"))

type_summary <- type_object(
  "Summary of the article.",
  author = type_string("Name of the article author"),
  topics = type_array(
    type_string(),
    'Array of topics, e.g. ["tech", "politics"]. Should be as specific as possible, and can overlap.'
  ),
  summary = type_string("Summary of the article. One or two paragraphs max"),
  coherence = type_integer(
    "Coherence of the article's key points, 0-100 (inclusive)"
  ),
  persuasion = type_number("Article's persuasion score, 0.0-1.0 (inclusive)")
)

chat <- chat_openai()
#> Using model = "gpt-5.4".
data <- chat$chat_structured(text, type = type_summary)
cat(data$summary)
#> Anthropic argues that frontier AI systems should be subject to effective third-party testing to validate safety, especially for risks such as national security misuse, election integrity harms, discrimination, and dangerous autonomous behaviors. The company says self-governance by labs is not enough; instead, governments, academia, and industry should build a broader testing ecosystem with trusted external evaluators, shared standards, and a narrowly scoped regime focused only on the most compute-intensive general-purpose models. Anthropic envisions a two-stage process with broad automated screening followed by deeper expert review when risks are flagged.
#> 
#> The article also connects testing to broader policy issues. Anthropic supports more public funding for AI evaluation infrastructure, government capacity through institutions like NIST and national research clouds, and development of tests for security-relevant capabilities. It argues that third-party testing can help address concerns around open-weight models and reduce regulatory capture by creating independent oversight rather than burdensome compliance structures that favor large firms. Overall, the piece presents third-party testing as the 'minimal viable policy' for managing current and future AI risks while preserving innovation.

str(data)
#> List of 5
#>  $ author    : chr "Anthropic"
#>  $ topics    : chr [1:10] "AI policy" "third-party testing" "frontier AI systems" "AI safety" ...
#>  $ summary   : chr "Anthropic argues that frontier AI systems should be subject to effective third-party testing to validate safety"| __truncated__
#>  $ coherence : int 90
#>  $ persuasion: num 0.84

Example 2: Named entity recognition

text <- "
  John works at Google in New York. He met with Sarah, the CEO of
  Acme Inc., last week in San Francisco.
"

type_named_entity <- type_object(
  name = type_string("The extracted entity name."),
  type = type_enum(c("person", "location", "organization"), "The entity type"),
  context = type_string("The context in which the entity appears in the text.")
)
type_named_entities <- type_array(type_named_entity)

chat <- chat_openai()
#> Using model = "gpt-5.4".
chat$chat_structured(text, type = type_named_entities)
#> # A tibble: 6 × 3
#>   name          type         context                                            
#>   <chr>         <fct>        <chr>                                              
#> 1 John          person       John works at Google in New York.                  
#> 2 Google        organization John works at Google in New York.                  
#> 3 New York      location     John works at Google in New York.                  
#> 4 Sarah         person       He met with Sarah, the CEO of Acme Inc., last week…
#> 5 Acme Inc.     organization He met with Sarah, the CEO of Acme Inc., last week…
#> 6 San Francisco location     He met with Sarah, the CEO of Acme Inc., last week…

Example 3: Sentiment analysis

text <- "
  The product was okay, but the customer service was terrible. I probably
  won't buy from them again.
"

type_sentiment <- type_object(
  "Extract the sentiment scores of a given text. Sentiment scores should sum to 1.",
  positive_score = type_number(
    "Positive sentiment score, ranging from 0.0 to 1.0."
  ),
  negative_score = type_number(
    "Negative sentiment score, ranging from 0.0 to 1.0."
  ),
  neutral_score = type_number(
    "Neutral sentiment score, ranging from 0.0 to 1.0."
  )
)

chat <- chat_openai()
#> Using model = "gpt-5.4".
str(chat$chat_structured(text, type = type_sentiment))
#> List of 3
#>  $ positive_score: num 0.12
#>  $ negative_score: num 0.74
#>  $ neutral_score : num 0.14

Note that while we’ve asked nicely for the scores to sum 1, which they do in this example (at least when I ran the code), this is not guaranteed.

Example 4: Text classification

text <- "The new quantum computing breakthrough could revolutionize the tech industry."

type_score <- type_object(
  name = type_enum(
    c(
      "Politics",
      "Sports",
      "Technology",
      "Entertainment",
      "Business",
      "Other"
    ),
    "The category name",
  ),
  score = type_number(
    "The classification score for the category, ranging from 0.0 to 1.0."
  )
)
type_classification <- type_array(
  type_score,
  description = "Array of classification results. The scores should sum to 1."
)

chat <- chat_openai()
#> Using model = "gpt-5.4".
data <- chat$chat_structured(text, type = type_classification)
data
#> # A tibble: 3 × 2
#>   name       score
#>   <fct>      <dbl>
#> 1 Technology  0.98
#> 2 Business    0.01
#> 3 Other       0.01

Example 5: Working with unknown keys

If you don’t know the keys in advance, you can use an array of name-value pairs. This approach works with all providers, replacing the now-deprecated .additional_properties argument.

type_characteristics <- type_array(
  type_object(
    name = type_string(),
    value = type_string()
  ),
  description = "All characteristics"
)

text <- "
  The man is tall, with a beard and a scar on his left cheek. He has a deep voice and wears a black leather jacket.
"

chat <- chat_anthropic("Extract all characteristics of supplied character")
#> Using model = "claude-sonnet-4-6".
chat$chat_structured(text, type = type_characteristics)
#> # A tibble: 5 × 2
#>   name          value               
#>   <chr>         <chr>               
#> 1 height        tall                
#> 2 facial hair   beard               
#> 3 scar location left cheek          
#> 4 voice         deep                
#> 5 clothing      black leather jacket

Example 6: Extracting data from an image

The final example comes from Dan Nguyen (you can see other interesting applications at that link). The goal is to extract structured data from this screenshot:

Screenshot of schedule A: a table showing assets and “unearned” income
Screenshot of schedule A: a table showing assets and “unearned” income

Even without any descriptions, ChatGPT does pretty well:

type_asset <- type_object(
  assert_name = type_string(),
  owner = type_string(),
  location = type_string(),
  asset_value_low = type_integer(),
  asset_value_high = type_integer(),
  income_type = type_string(),
  income_low = type_integer(),
  income_high = type_integer(),
  tx_gt_1000 = type_boolean()
)
type_assets <- type_array(type_asset)

chat <- chat_openai()
#> Using model = "gpt-5.4".
image <- content_image_file("congressional-assets.png")
data <- chat$chat_structured(image, type = type_assets)
data
#> # A tibble: 2 × 9
#>   assert_name        owner location asset_value_low asset_value_high income_type
#>   <chr>              <chr> <chr>              <int>            <int> <chr>      
#> 1 11 Zinfandel Lane… JT    St. Hel…         5000001         25000000 Grape Sales
#> 2 25 Point Lobos - … SP    San Fra…         5000001         25000000 Rent       
#> # ℹ 3 more variables: income_low <int>, income_high <int>, tx_gt_1000 <lgl>

Tools and structured data

$chat_structured() automatically disables any registered tools. This is because when tools are involved, the model may need multiple round trips to gather information, and it has no way to know which response is the “last” one that should use the structured output schema. If you need both, call $chat() first to get the information you need via tools, then call $chat_structured() to extract structured data from the conversation.

Token usage

provider model input output cached_input price
OpenAI gpt-4.1 11 105 0 $0.00
OpenAI gpt-4.1-nano 474 129 0 $0.00
Anthropic claude-sonnet-4-6 1057 2418 0 $0.04
OpenAI gpt-5.4 7357 2064 0 $0.05