Structured data

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

library(ellmer)

Structured data basics

To extract structured data you call the $extract_data() method instead of the $chat() method. 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-4o".
chat$extract_data(
  "My name is Susan and I'm 13 years old",
  type = type_object(
    age = type_number(),
    name = type_string()
  )
)
#> $age
#> [1] 13
#> 
#> $name
#> [1] "Susan"

The same basic idea works with images too:

chat$extract_data(
  content_image_url("https://www.r-project.org/Rlogo.png"),
  type = type_object(
    primary_shape = type_string(),
    primary_colour = type_string()
  )
)
#> $primary_shape
#> [1] "oval and a letter R"
#> 
#> $primary_colour
#> [1] "silver and blue"

Data types basics

To define your desired type specification (also known as a schema), you use the type_() functions. (You might already be familiar with these if you’ve done any function calling, as discussed in vignette("function-calling")). The type functions can be divided into three main groups:

• Scalars represent single values, of which there are five types: type_boolean(), type_integer(), type_number(), type_string(), and type_enum(), representing a single logical, integer, double, string, and factor value respectively.

• Arrays represent any number of values of the same type and are created with type_array(). You must always supply the item argument which specifies the type of each individual element. Arrays of scalars are very similar to R’s atomic vectors:

  type_logical_vector <- type_array(items = type_boolean())
  type_integer_vector <- type_array(items = type_integer())
  type_double_vector <- type_array(items = type_number())
  type_character_vector <- type_array(items = type_string())

  You can also have arrays of arrays and arrays of objects, which more closely resemble lists with well defined structures:

  list_of_integers <- type_array(items = type_integer_vector)

• Objects represent a collection of named values and are created with type_object(). Objects can contain any number of scalars, arrays, and other objects. They are similar to named lists in R.

  type_person <- type_object(
    name = type_string(),
    age = type_integer(),
    hobbies = type_array(items = type_string())
  )

Using these type specifications ensures that the LLM will return JSON. But ellmer goes one step further to convert the results to their most natural R representation. This currently converts arrays of boolean, integers, numbers, and strings into logical, integer, numeric, and character vectors, and arrays of objects into data frames. You can opt-out of this and get plain lists instead by setting convert = FALSE in $extract_data().

As well as the definition of the types, you need to provide the LLM with some information about what you actually want. This is the purpose of the first argument, description, which is 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 possess (e.g. minimum or maximum values, date formats, …). You aren’t guaranteed that these requests will be honoured, but the LLM will usually make a best effort to do so.

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

Now we’ll dive into some examples before coming back to talk more data types details.

Examples

The following examples are closely inspired by the Claude documentation and 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(
    'Array of topics, e.g. ["tech", "politics"]. Should be as specific as possible, and can overlap.',
    type_string(),
  ),
  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-4o".
data <- chat$extract_data(text, type = type_summary)
cat(data$summary)
#> This article from Anthropic argues for the necessity of third-party testing as a key component of AI policy to address the challenges posed by frontier AI systems. These systems, such as large-scale generative models, possess immense capabilities that, if left unchecked, can lead to misuses or accidents. The article stresses that a robust, third-party testing regime is vital for instilling trust in AI systems, ensuring safety, and managing risks, but it must be carefully designed to avoid disadvantaging smaller enterprises and must adapt to future advancements in AI capabilities. The article outlines the proposed structure and implementation of such a regime, involving cooperation across industry, government, and academia, with an emphasis on avoiding regulatory capture and maintaining an open dissemination of AI while ensuring safety.

str(data)
#> List of 5
#>  $ author    : chr "Anthropic"
#>  $ topics    : chr [1:6] "AI Safety" "Regulatory Policy" "Third-party Testing" "Generative AI" ...
#>  $ summary   : chr "This article from Anthropic argues for the necessity of third-party testing as a key component of AI policy to "| __truncated__
#>  $ coherence : int 95
#>  $ persuasion: num 0.85

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

chat <- chat_openai()
#> Using model = "gpt-4o".
chat$extract_data(text, type = type_named_entities)
#>            name         type                                     context
#> 1          John       person Mentioned as working at Google in New York.
#> 2        Google organization               The company where John works.
#> 3      New York     location              The location where John works.
#> 4         Sarah       person           Mentioned as the CEO of Acme Inc.
#> 5     Acme Inc. organization         The company where Sarah is the CEO.
#> 6 San Francisco     location          The location where John met Sarah.

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-4o".
str(chat$extract_data(text, type = type_sentiment))
#> List of 3
#>  $ positive_score: num 0.1
#>  $ negative_score: num 0.7
#>  $ neutral_score : num 0.2

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

Example 4: Text classification

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

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

chat <- chat_openai()
#> Using model = "gpt-4o".
data <- chat$extract_data(text, type = type_classification)
data
#>            name score
#> 1    Technology 0.700
#> 2      Business 0.200
#> 3         Other 0.050
#> 4 Entertainment 0.025
#> 5      Politics 0.020
#> 6        Sports 0.005

Example 5: Working with unknown keys

type_characteristics <- type_object(
  "All characteristics",
  .additional_properties = TRUE
)

prompt <- "
  Given a description of a character, your task is to extract all the characteristics of that character.

  <description>
  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.
  </description>
"

chat <- chat_claude()
str(chat$extract_data(prompt, type = type_characteristics))
#>  list()

This examples only works with Claude, not GPT or Gemini, because only Claude supports adding arbitrary additional properties.

Example 6: Extracting data from an image

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

A 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(items = type_asset)

chat <- chat_openai()
#> Using model = "gpt-4o".
image <- content_image_file("congressional-assets.png")
data <- chat$extract_data(image, type = type_assets)
data
#>                                 assert_name owner
#> 1  11 Zinfandel Lane - Home & Vineyard [RP]    JT
#> 2 25 Point Lobos - Commercial Property [RP]    SP
#>                              location asset_value_low asset_value_high
#> 1             St. Helena/Napa, CA, US         5000001         25000000
#> 2 San Francisco/San Francisco, CA, US         5000001         25000000
#>   income_type income_low income_high tx_gt_1000
#> 1 Grape Sales     100001     1000000       TRUE
#> 2        Rent     100001     1000000      FALSE

Advanced data types

Now that you’ve seen a few examples, it’s time to get into more specifics about data type declarations.

Required vs optional

By default, all components of an object are required. If you want to make some optional, set required = FALSE. This is a good idea if you don’t think your text will always contain the required fields as LLMs may hallucinate data in order to fulfill your spec.

For example, here the LLM hallucinates a date even though there isn’t one in the text:

type_article <- type_object(
  "Information about an article written in markdown",
  title = type_string("Article title"),
  author = type_string("Name of the author"),
  date = type_string("Date written in YYYY-MM-DD format.")
)

prompt <- "
  Extract data from the following text:

  <text>
  # Structured Data
  By Hadley Wickham

  When using an LLM to extract data from text or images, you can ask the chatbot to nicely format it, in JSON or any other format that you like.
  </text>
"

chat <- chat_openai()
#> Using model = "gpt-4o".
chat$extract_data(prompt, type = type_article)
#> $title
#> [1] "Structured Data"
#> 
#> $author
#> [1] "Hadley Wickham"
#> 
#> $date
#> [1] ""
str(data)
#> 'data.frame':    2 obs. of  9 variables:
#>  $ assert_name     : chr  "11 Zinfandel Lane - Home & Vineyard [RP]" "25 Point Lobos - Commercial Property [RP]"
#>  $ owner           : chr  "JT" "SP"
#>  $ location        : chr  "St. Helena/Napa, CA, US" "San Francisco/San Francisco, CA, US"
#>  $ asset_value_low : int  5000001 5000001
#>  $ asset_value_high: int  25000000 25000000
#>  $ income_type     : chr  "Grape Sales" "Rent"
#>  $ income_low      : int  100001 100001
#>  $ income_high     : int  1000000 1000000
#>  $ tx_gt_1000      : logi  TRUE FALSE

Note that I’ve used more of an explict prompt here. For this example, I found that this generated better results, and it’s a useful place to put additional instructions.

If let the LLM know that the fields are all optional, it’ll instead return NULL for the missing fields:

type_article <- type_object(
  "Information about an article written in markdown",
  title = type_string("Article title", required = FALSE),
  author = type_string("Name of the author", required = FALSE),
  date = type_string("Date written in YYYY-MM-DD format.", required = FALSE)
)
chat$extract_data(prompt, type = type_article)
#> $title
#> [1] "Structured Data"
#> 
#> $author
#> [1] "Hadley Wickham"
#> 
#> $date
#> NULL

Data frames

If you want to define a data frame like object, you might be tempted to create a definition similar to what R uses: an object (i.e. a named list) containing multiple vectors (i.e. arrays):

type_my_df <- type_object(
  name = type_array(items = type_string()),
  age = type_array(items = type_integer()),
  height = type_array(items = type_number()),
  weight = type_array(items = type_number())
)

This however, is not quite right becuase there’s no way to specify that each array should have the same length. Instead you need to turn the data structure “inside out”, and instead create an array of objects:

type_my_df <- type_array(
  items = type_object(
    name = type_string(),
    age = type_integer(),
    height = type_number(),
    weight = type_number()
  )
)

If you’re familiar with the terms between row-oriented and column-oriented data frames, this is the same idea. Since most language don’t possess vectorisation like R, row-oriented structures tend to be much more common in the wild.

Token usage

name	input	output
OpenAI-https://api.openai.com/v1	6379	638
Claude	480	136