Natural Language Processing using LLMs in R & Python :: Cheatsheet

Hex logo for mall - drawing of the inside of a mall. There is a fountain in the middle.

Download PDF



Click on R or the Python tab to see the information in your preferred language

Intro

Use LLM’s to perform NLP row-wise over a data frame. mall comes with pre-defined prompts that perform specific NLP operations, and then places the results in a new column. Use OpenAI, Ollama, Anthropic and many others thanks to its integration with ellmer

mall’s data frame functions are designed with ‘tidy’ principals in mind, so they work with the Tidyverse packages. mall also includes functions that work with string vectors.

A diagram showing how mall adds a prompt and passes it to the LLM via ellmer

Use LLM’s to perform NLP row-wise over a data frame. mall comes with pre-defined prompts that perform specific NLP operations, and then places the results in a new column. Use OpenAI, Ollama, Anthropic and many others thanks to its integration with chatlas

mall works as an extension for Polars data frames. It also works with string vectors.

A diagram showing how mall adds a prompt and passes it to the LLM via chatlas

Getting started

  1. Load the libraries

    library(mall)
library(ellmer)

  2. Create a vendor specific chat connection

    chat <- chat_openai()

  3. Pass the chat object to mall

    llm_use(chat)

  • For dataframes:

    data(“reviews") # Sample product reviews

reviews |>
  llm_sentiment(review)

  • For vectors:

    llm_vec_sentiment(c("I am happy", "I am sad”))


Connect automatically

As a convenience, mall is able to automatically establish a connection with the LLM. To do this you can use the .mall_chat option: options(.mall_chat=ellmer::chat_openai(model="gpt-4o")) Add this line to your .Rprofile file in order for that code to run every time you start R. You can call usethis::edit_r_profile() to edit.

Start by creating a new LLM connection

from chatlas import ChatOpenAI
chat = ChatOpenAI()

  • For Dataframes

    1. Load the library

      import mall

    2. Read or load your data frame

      reviews = mall.MallData.reviews # Sample product
reviews

    3. Pass the chat object to mall

      reviews.llm.use(chat)

    4. Access NLP functions via .llm

      reviews.llm.sentiment('review')

  • For String vectors

    1. Load the LLMVec class

      from mall import LLMVec

    2. Create a new LLMVec object

      llm = LLMVec(chat)

    3. Pass a vector to a function in the new object

      llm.sentiment(['I am happy', 'I am sad'])

NLP functions

Sentiment analysis

  • llm_sentiment(.data, col, options = c("positive", "negative", “neutral"), pred_name = “.sentiment", additional_prompt = “”)

    llm_sentiment(reviews, review)

  • llm_vec_sentiment(x, options = c("positive", "negative", "neutral"), additional_prompt = "", preview = FALSE)

    llm_vec_sentiment(c("I am happy", "I am sad"))

Special arguments:

options: Customize the sentiments to check for: options = c(“positive”, “negative”). Use ‘tilde’ to mask the results, for example c("positive" ~ 1, "negative" ~ 0)) returns 1 for positive and 0 for negative.

  • <Dataframe>.llm.sentiment(col, options = ['positive', 'negative', 'neutral'], additional='', pred_name ='sentiment')

    reviews.llm.sentiment('review')

  • <LLMVec object>.sentiment(x, options=['positive', 'negative', 'neutral'], additional='')

    llm.sentiment(['I am happy', 'I am sad'])

Special arguments:

options: Customize the sentiments to check for: options = ["positive", "negative"]. Use a DICT object to mask the results, for example {"positive": 1, "negative" 0} returns 1 for positive and 0 for negative.

Extract

Extract specific entity, or entities, from the provided text

  • llm_extract(.data, col, labels, expand_cols = FALSE, additional_prompt = "", pred_name = “.extract")

    llm_extract(reviews, review, labels = "product")

  • llm_vec_extract(x, labels = c(), additional_prompt = "", preview = FALSE)

    llm_vec_extract("bob smith, 123 3rd street", c("name", "address"))

Special arguments

labels: A vector to specify the entities to identify expand_cols - If multiple labels, this indicates if the labels will show up in their own column (data frames only)

  • <DataFrame>.llm.extract(col, labels='', expand_cols = False, additional = '', pred_name = 'extract')

    reviews.llm.extract("review", labels = "product")

  • <LLMVec object>.extract(x, labels='', additional='')

    llm.extract(['bob smith, 123 3rd street'], labels=['name', 'address'])

Special arguments

labels: A vector to specify the entities to identify expand_cols - If multiple labels, this indicates if the labels will show up in their own column (data frames only)

Summarize

Summarize text into a specified number of words

  • llm_summarize( .data, col, max_words = 10, pred_name = ".summary", additional_prompt = "")

    llm_summarize(reviews, review, max_words = 5)

  • llm_vec_summarize(x, max_words = 10, additional_prompt = "", preview = FALSE)

    llm_vec_summarize("This has been the best TV I've
ever used. Great screen, and sound.", max_words = 5)

  • <DataFrame>.llm.summarize(x, max_words=10, additional='')

    reviews.llm.summarize("review", 5)

  • <LLMVec object>.summarize(x, max_words=10, additional='')

    llm.summarize(['This has been the best TV Ive ever used. Great screen, and sound.'], max_words = 5)

Verify

Check if a statement is true or not based on the provided text

  • llm_verify(.data, col, what, yes_no = factor(c(1, 0)), pred_name = ".verify", additional_prompt = "")

    llm_verify(reviews, review, "is the customer happy")

  • llm_vec_verify(x, what, yes_no = factor(c(1, 0)), additional_prompt = “", preview = FALSE)

    llm_verify(c("I am happy", "I am sad"), "is the person happy”)

Special arguments

yes_no: Customize what it returns for true/false with a vector yes_no = c("y", "n").

  • <DataFrame>.llm.verify(col, what='', yes_no=[1, 0], additional='', pred_name='verify')

    reviews.llm.verify("review", "is the customer happy")

  • <LLMVec object>.verify(x, what='', yes_no=[1, 0], additional='')

    llm.verify(['I am happy', 'I am sad'], what = 'Is the person happy?')

Special arguments

yes_no: Customize what it returns for true/false with a vector yes_no = ["y", "n"].

Classify

Classify the provided text as one of the options provided via the labels

  • llm_classify(.data, col, labels, pred_name = ".classify", additional_prompt = "")

    llm_classify(reviews, review, c("appliance", "computer"))

  • llm_vec_classify(x, labels, additional_prompt = "", preview = FALSE)

    llm_vec_classify(c("this is important!", "just whenever”), c("urgent", "not urgent"))

Special arguments

labels: A character vector with at least 2 labels to classify the text as

  • <DataFrame>.llm.classify(col, labels='', additional='', pred_name='classify')

    reviews.llm.classify("review", ["appliance", "computer"])

  • <LLMVec object>.classify(x, labels='', additional='')

    llm.classify(['this is important!', 'there is no rush'], ['urgent', 'not urgent'])

Special arguments

labels: A character vector with at least 2 labels to classify the text as

Translate

Translate into target language

  • llm_translate( .data, col, language, pred_name = ".translation", additional_prompt = “")

    llm_translate(reviews, review, "spanish")

  • llm_vec_translate(x, language, additional_prompt = "", preview = FALSE)

    llm_vec_translate("grass is green", "spanish")

Special arguments

language: Target language. No origin language is passed since the LLM detects it automatically.

  • <DataFrame>.llm.translate(col, language='', additional='', pred_name='translation')

    reviews.llm.translate("review", "spanish")

  • <LLMVec object>.translate(x, language='', additional='')

    llm.translate([‘the grass is green’], language = 'spanish')

Special arguments

language: Target language. No origin language is passed since the LLM detects it automatically.

Custom

Create your own prompt

  • llm_custom(.data, col, prompt = "", pred_name = ".pred", valid_resps = "")

    my_prompt <- "Answer a question. Return only the
answer, no explanation. Acceptable answers are 'yes',
‘no’. Answer this about the following text, is this a
happy customer?:"
llm_custom(reviews , review, my_prompt)

  • llm_vec_custom(x, prompt = "", valid_resps = NULL)

Special arguments

valid_resps: A vector to specify the set of answers expected back. mall will change those not in the set to NA

  • <DataFrame>.llm.custom(col, prompt='', valid_resps='', pred_name='custom')

    my_prompt = "Answer a question. Return only the answer," \
    " no explanation. Acceptable answers are 'yes', 'no'." \
    "Answer this about the following text, is this a happy customer?:"
reviews.llm.custom("review", prompt = my_prompt)

  • <LLMVec object>.custom(x, prompt='', valid_resps='')

Special arguments

valid_resps: A vector to specify the set of answers expected back.

Shared arguments

additional_prompt: Appends instructions to the LLM.

llm_classify(reviews, review, c("appliance", "computers"), additional_prompt = "Consider TVs as appliances")

pred_name: Name of the new column. Defaults are set based on the NLP operation. (Data frames only)

llm_vec_translate("grass is green", "spanish", pred_name = "in_spanish")

preview: Returns what it would be sent to the LLM instead (Vectors only)

additional: Appends more instructions to the LLM.

reviews.llm.classify("review", ["appliance", “computer"],additional="Consider TVs as appliances”)

pred_name: Name of the new column. Defaults are set based on the NLP operation. (Data frames only)

Other features

Ollama direct

If Ollama is the only LLM provider you are using, then a simplified way to connect is available which does not require an ellmer Chat object. Simply pass “ollama” as the backend, and specify the model:

llm_use("ollama", model= “llama3.2")

If Ollama is the only LLM provider you are using, then a simplified way to connect is available which does not require an ellmer Chat object. Simply pass “ollama” as the backend, and specify the model:

  • Dataframe:

    reviews.llm.use('ollama','llama3.2')

  • Vector:

    llm = LLMVec('ollama','llama3.2')

Caching results

By default, mall saves the LLM results in a temp folder. To specify a folder call:

llm_use(chat, .cache = "<my folder>")

To turn off use:

llm_use(chat, .cache = "")

By default, mall saves the LLM results in a folder. To specify a folder call:

  • Dataframe:

    reviews.llm.use(chat, _cache='<my folder>')

  • Vector:

    llm = LLMVec(chat, _cache='<my folder>')

To turn off use:

  • Dataframe:

    reviews.llm.use(chat, _cache='')

  • Vector:

    llm = LLMVec(chat, _cache='')