Setting up API Keys or ollama

Before using tidyllm, set up API keys for the services you plan to use. Here’s how to set them up for different providers:

1. For Claude models you can get an API key in the Anthropic Console:

Sys.setenv(ANTHROPIC_API_KEY = "YOUR-ANTHROPIC-API-KEY")

2. For OpenAI you can obtain an API key by signing up at OpenAI and set it with:

Sys.setenv(OPENAI_API_KEY = "YOUR-OPENAI-API-KEY")

3. For Google Gemini you can setup an API key in the Google AI Studio

Sys.setenv(GOOGLE_API_KEY = 'YOUR-GOOGLE-API-KEY')

4. For Mistral you can set an API key on the Mistral console page and set it by

Sys.setenv(MISTRAL_API_KEY = "YOUR-MISTRAL-API-KEY")

5. For groq (not be confused with grok) you can setup you API keys in the Groq Console:

Sys.setenv(GROQ_API_KEY = "YOUR-GROQ-API-KEY")

6. For Perplexity you can get an API key in the API settings:

Sys.setenv(PERPLEXITY_API_KEY = "YOUR-PERPLEXITY-API-KEY")

Alternatively, for persistent storage, add these keys to your .Renviron file:

For this, run usethis::edit_r_environ(), and add a line with with an API key in this file, for example:

ANTHROPIC_API_KEY="YOUR-ANTHROPIC-API-KEY"

If you want to work with local large lange models via ollama you need to install it from the official project website. Ollama sets up a local large language model server that you can use to run open-source models on your own devices.

Basic Usage

Let’s start with a simple example using tidyllm to interact with different language models:

library(tidyllm)

# Start a conversation with Claude
conversation <- llm_message("What is the capital of France?") |>
  chat(claude())

#Standard way that llm_messages are printed
conversation

## Message History:
## system: You are a helpful assistant
## --------------------------------------------------------------
## user: What is the capital of France?
## --------------------------------------------------------------
## assistant: The capital of France is Paris.
## --------------------------------------------------------------

# Continue the conversation with ChatGPT
conversation <- conversation |>
  llm_message("What's a famous landmark in this city?") |>
  chat(openai)

get_reply(conversation)

## [1] "A famous landmark in Paris is the Eiffel Tower."

tidyllm is built around a message-centric interface design, where most interactions work with a message history, that is created by functions like llm_message() or modified by API-functions like chat(). These API-functions always work with a combination of verbs and providers:

• Verbs (e.g., chat(), embed(), send_batch()) define the type of action you want to perform.
• Providers (e.g., openai(), claude(), ollama()) specify the API or service to handle the action.

Alternatively there are also provider-specific functions like openai_chat() or claude_chat() that do the work in the background of the main verbs that you can also call directly. The documentation for these provider-specific functions offers a comprehensive overview of the full range of actions and input types supported for each API.

Sending Images to Models

tidyllm also supports sending images to multimodal models. Let’s send this picture here:

Here we let ChatGPT guess where the picture was made:

# Describe an image using a llava model on ollama
image_description <- llm_message("Describe this picture? Can you guess where it was made?",
                                 .imagefile = "picture.jpeg") |>
  chat(openai(.model = "gpt-4o"))

# Get the last reply
get_reply(image_description)

## [1] "The picture shows a beautiful landscape with a lake, mountains, and a town nestled below. The sun is shining brightly, casting a serene glow over the water. The area appears lush and green, with agricultural fields visible. \n\nThis type of scenery is reminiscent of northern Italy, particularly around Lake Garda, which features similar large mountains, picturesque water, and charming towns."

Adding PDFs to messages

The llm_message() function also supports extracting text from PDFs and including it in the message. This allows you to easily provide context from a PDF document when interacting with an AI assistant.

To use this feature, you need to have the pdftools package installed. If it is not already installed, you can install it with:

install.packages("pdftools")

To include text from a PDF in your prompt, simply pass the file path to the .pdf argument of the chat function:

llm_message("Please summarize the key points from the provided PDF document.", 
     .pdf = "die_verwandlung.pdf") |>
     chat(openai(.model = "gpt-4o-mini"))

## Message History:
## system: You are a helpful assistant
## --------------------------------------------------------------
## user: Please summarize the key points from the provided PDF document.
##  -> Attached Media Files:  die_verwandlung.pdf 
## --------------------------------------------------------------
## assistant: Here are the key points from the provided PDF document 'Die Verwandlung' by Franz Kafka:
## 
## 1. The story centers around Gregor Samsa, who wakes up one morning to find that he has been transformed into a giant insect-like creature.
## 
## 2. Gregor's transformation causes distress and disruption for his family. They struggle to come to terms with the situation and how to deal with Gregor in his new state.
## 
## 3. Gregor's family, especially his sister Grete, initially tries to care for him, but eventually decides they need to get rid of him. They lock him in his room and discuss finding a way to remove him.
## 
## 4. Gregor becomes increasingly isolated and neglected by his family. He becomes weaker and less mobile due to his injuries and lack of proper care.
## 
## 5. Eventually, Gregor dies, and his family is relieved. They then begin to make plans to move to a smaller, more affordable apartment and start looking for new jobs and opportunities.
## --------------------------------------------------------------

The package will automatically extract the text from the PDF and include it in the prompt sent to the an API. The text will be wrapped in <pdf> tags to clearly indicate the content from the PDF:

Please summarize the key points from the provided PDF document.

<pdf filename="example_document.pdf">
Extracted text from the PDF file...
</pdf>

Sending R Outputs to Language Models

You can automatically include R code outputs in your prompts. llm_message() has an optional argument .f in which you can specify a (anonymous) function, which will be run and which console output will be captured and appended to the message when you run it.

In addition you can use .capture_plot to send the last plot pane to a model.

library(tidyverse)

# Create a plot for the mtcars example data
ggplot(mtcars, aes(wt, mpg)) +
  geom_point() +
  geom_smooth(method = "lm", formula = 'y ~ x') +
  labs(x="Weight",y="Miles per gallon")

Now we can send the plot and data summary to a language model:

library(tidyverse)
llm_message("Analyze this plot and data summary:", 
                  .capture_plot = TRUE, #Send the plot pane to a model
                  .f = ~{summary(mtcars)}) |> #Run summary(data) and send the output
  chat(claude())

## Message History:
## system: You are a helpful assistant
## --------------------------------------------------------------
## user: Analyze this plot and data summary:
##  -> Attached Media Files:  file1568f6c1b4565.png, RConsole.txt 
## --------------------------------------------------------------
## assistant: Based on the plot and data summary provided, here's an analysis:
## 
## 1. Relationship between Weight and MPG:
##    The scatter plot shows a clear negative correlation between weight (wt) and miles per gallon (mpg). As the weight of the car increases, the fuel efficiency (mpg) decreases.
## 
## 2. Linear Trend:
##    The blue line in the plot represents a linear regression fit. The downward slope confirms the negative relationship between weight and mpg.
## 
## 3. Data Distribution:
##    - The weight of cars in the dataset ranges from 1.513 to 5.424 (likely in thousands of pounds).
##    - The mpg values range from 10.40 to 33.90.
## 
## 4. Variability:
##    There's some scatter around the regression line, indicating that while weight is a strong predictor of mpg, other factors also influence fuel efficiency.
## 
## 5. Other Variables:
##    While not shown in the plot, the summary statistics provide information on other variables:
##    - Cylinder count (cyl) ranges from 4 to 8, with a median of 6.
##    - Horsepower (hp) ranges from 52 to 335, with a mean of 146.7.
##    - Transmission type (am) is binary (0 or 1), likely indicating automatic vs. manual.
## 
## 6. Model Fit:
##    The grey shaded area around the regression line represents the confidence interval. It widens at the extremes of the weight range, indicating less certainty in predictions for very light or very heavy vehicles.
## 
## 7. Outliers:
##    There are a few potential outliers, particularly at the lower and higher ends of the weight spectrum, that deviate from the general trend.
## 
## In conclusion, this analysis strongly suggests that weight is a significant factor in determining a car's fuel efficiency, with heavier cars generally having lower mpg. However, the presence of scatter in the data indicates that other factors (possibly related to engine characteristics, transmission type, or aerodynamics) also play a role in determining fuel efficiency.
## --------------------------------------------------------------

Getting replies from the API

Retrieve an assistant reply as text from a message history with get_reply(). Specify an index to choose which assistant message to get:

conversation <- llm_message("Imagine a German adress.") |>
     chat(groq()) |>
     llm_message("Imagine another address") |>
     chat(claude())

conversation

## Message History:
## system: You are a helpful assistant
## --------------------------------------------------------------
## user: Imagine a German adress.
## --------------------------------------------------------------
## assistant: Let's imagine a German address: 
## 
## Herr Müller
## Musterstraße 12
## 53111 Bonn
## --------------------------------------------------------------
## user: Imagine another address
## --------------------------------------------------------------
## assistant: Let's imagine another German address:
## 
## Frau Schmidt
## Fichtenweg 78
## 42103 Wuppertal
## --------------------------------------------------------------

By default get_reply() gets the last assistant message. Alternatively you can also use last_reply() as a shortcut for the latest response.

#Getting the first reply
conversation |> get_reply(1)

## [1] "Let's imagine a German address: \n\nHerr Müller\nMusterstraße 12\n53111 Bonn"

#By default it gets the last reply
conversation |> get_reply()

## [1] "Let's imagine another German address:\n\nFrau Schmidt\nFichtenweg 78\n42103 Wuppertal"

Or you can convert the text (without attachments) to a tibble with as_tibble():

conversation |> as_tibble()

## # A tibble: 5 × 2
##   role      content                                                             
##   <chr>     <chr>                                                               
## 1 system    "You are a helpful assistant"                                       
## 2 user      "Imagine a German adress."                                          
## 3 assistant "Let's imagine a German address: \n\nHerr Müller\nMusterstraße 12\n…
## 4 user      "Imagine another address"                                           
## 5 assistant "Let's imagine another German address:\n\nFrau Schmidt\nFichtenweg …

You can use the get_metadata() function to retrieve metadata on models and token usage from assistant replies:

conversation |> get_metadata()

## # A tibble: 2 × 6
##   model  timestamp           prompt_tokens completion_tokens
##   <chr>  <dttm>                      <dbl>             <dbl>
## 1 groq-… 2024-11-08 14:25:43            20                45
## 2 claud… 2024-11-08 14:26:02            80                40
## # ℹ 2 more variables: total_tokens <dbl>,
## #   api_specific <list>

By default it collects metadata for the whole message history, but you can also set an .index to only get metadata for a specific reply. Alternatively you can print out metadata with the standard print method with the .meta-argument in print() or via the tidyllm_print_metadata option. The list column api_specific contains special metadata that is only available to some APIs (like citations for models with search grounding on perplexity() or special information on model loading for local APIs).

Working with structured model outputs

To make model responses easy to interpret and integrate into your workflow, tidyllm supports defining schemas to ensure that models reply with structured outputs in JSON (JavaScript Object Notation) following your specifications. JSON is a standard format for organizing data in simple key-value pairs, which is both human-readable and machine-friendly.

Currently, openai(), gemini() and ollama() are the only API provider functions in tidyllm that support schema enforcement through the .json_schema argument. This ensures that replies conform to a pre-defined consistent data formatting.

To create schemas, you can use the tidyllm_schema() function, which translates your data format specifications into the JSON-schema format the API requires. This helper function standardizes the data layout by ensuring flat (non-nested) JSON structures with defined data types. Here’s how to define a schema:

• name: A name identifier for the schema. By default this name is "tidyllm_schema".
• ... (fields): Named arguments for field names and their data types, including:
  • "character" or "string": Text fields.
  • "factor(...)": Enumerations with allowable values, like factor(Germany, France).
  • "logical": TRUE or FALSE
  • "numeric": Numeric fields.
  • "type[]": Lists of a given type, such as "character[]".

Here’s an example schema defining an address format:

address_schema <- tidyllm_schema(
  name = "AddressSchema",
  street = "character",
  houseNumber = "numeric",
  postcode = "character",
  city = "character",
  region = "character",
  country = "factor(Germany,France)"
)
address <- llm_message("Imagine an address in JSON format that matches the schema.") |>
        chat(openai(),.json_schema = address_schema)
address

## Message History:
## system: You are a helpful assistant
## --------------------------------------------------------------
## user: Imagine an address in JSON format that matches the schema.
## --------------------------------------------------------------
## assistant: {"street":"Hauptstraße","houseNumber":123,"postcode":"10115","city":"Berlin","region":"Berlin","country":"Germany"}
## --------------------------------------------------------------

The model responded in JSON format, organizing data into key-value pairs like specified. You can then convert this JSON output into an R list for easier handling with get_reply_data():

address |> get_reply_data() |> str()

## List of 6
##  $ street     : chr "Hauptstraße"
##  $ houseNumber: int 123
##  $ postcode   : chr "10115"
##  $ city       : chr "Berlin"
##  $ region     : chr "Berlin"
##  $ country    : chr "Germany"

Ollama, OpenAI and the Google Gemini API work with a schema provided by tidyllm_schema(). Other API providers like groq(), and mistral() only support structured outputs through a simpler JSON mode, accessible with the .json argument in these provider functions. Since these APIs do not currently support native schema enforcement, you’ll need to prompt the model to follow a specified format directly in your messages. Although get_reply_data() can help extract structured data from these responses when you set .json=TRUE in each of the API provider functions, the model may not always adhere strictly to the specified structure. Once these APIs support native schema enforcement, tidyllm will integrate full schema functionality for them.

API parameters

Different API functions support different model parameters like, how deterministic the response should be via parameters like temperature. You can set these parameters via arguments in the verbs or provider functions. Please read API-documentation and the documentation of the model functions for specific examples. Common arguments, such as temperature, can be specified directly through the main verbs like chat(). For example, chat() supports arguments such as .model,.temperature, .json_schema and more (see the full list in the chat() documentation) directly.

  temp_example <- llm_message("Explain how temperature parameters work 
in large language models and why temperature 0 gives you deterministic outputs 
in one sentence.")
  
  #per default it is non-zero
  temp_example |> chat(ollama,.temperature=0)

## Message History:
## system: You are a helpful assistant
## --------------------------------------------------------------
## user: Explain how temperature parameters work 
## in large language models  and why temperature 0 gives you deterministic
## outputs in one sentence.
## --------------------------------------------------------------
## assistant: In large language models, temperature parameters control the randomness of generated text by scaling the output probabilities, with higher temperatures introducing more uncertainty and lower temperatures favoring more likely outcomes; specifically, setting temperature to 0 effectively eliminates all randomness, resulting in deterministic outputs because it sets the probability of each token to its maximum likelihood value.
## --------------------------------------------------------------

  #Retrying with .temperature=0
  temp_example |> chat(ollama,.temperature=0)

## Message History:
## system: You are a helpful assistant
## --------------------------------------------------------------
## user: Explain how temperature parameters work 
## in large language models  and why temperature 0 gives you deterministic
## outputs in one sentence.
## --------------------------------------------------------------
## assistant: In large language models, temperature parameters control the randomness of generated text by scaling the output probabilities, with higher temperatures introducing more uncertainty and lower temperatures favoring more likely outcomes; specifically, setting temperature to 0 effectively eliminates all randomness, resulting in deterministic outputs because it sets the probability of each token to its maximum likelihood value.
## --------------------------------------------------------------

Provider-specific arguments—such as .ollama_server for ollama() or .fileid for gemini() are set directly in the provider function:

conversation <- llm_message("Hello") |> 
  chat(ollama(.ollama_server = "http://localhost:11434"), 
       .temperature = 0)

When an argument is provided in both chat() and the provider, the value specified in chat() takes precedence. For instance:

#This uses GPT-4o
conversation <- llm_message("Hello") |> 
  chat(openai(.model="gpt-4o-mini"),
       .model="gpt-4o")

If a common argument set in chat is not supported by a provider chat() will raise an error. For example, sending a .json_schema to a provider that does not support it will raise an error:

address <- llm_message("Imagine an address in JSON format that matches the schema.") |>
        chat(groq(),.json_schema = address_schema)

Embeddings

Embedding models in tidyllm transform textual inputs into vector representations, capturing semantic information that can enhance similarity comparisons, clustering, and retrieval tasks. You can generate embeddings with the embed()-function. These functions return a semantic vector representation either for each message in a message history or, more typically for this application, for each entry in a character vector:

c("What is the meaning of life?",
  "How much wood would a woodchuck chuck?",
  "How does the brain work?") |>
  embed(ollama)

## # A tibble: 3 × 2
##   input                                  embeddings 
##   <chr>                                  <list>     
## 1 What is the meaning of life?           <dbl [384]>
## 2 How much wood would a woodchuck chuck? <dbl [384]>
## 3 How does the brain work?               <dbl [384]>

The output is a tibble with two columns, the text input for the embedding and a list column that contains a vector of semantic embeddings for each input.

Batch requests

Anthropic, OpenAI and Mistral offer batch request options that are around 50% cheaper than standard single-interaction APIs. Batch processing allows you to submit multiple message histories at once, which are then processed together on the model providers servers, usually within a 24-hour period. In tidyllm, you can use the send_batch() function to submit these batch requests to either API.

Here’s an example of how to send a batch request to Claude’s batch API:

#Create a message batch and save it to disk to fetch it later
glue("Write a poem about {x}", x=c("cats","dogs","hamsters")) |>
  purrr::map(llm_message) |>
  send_batch(claude()) |>
  saveRDS("claude_batch.rds")

The send_batch() function returns the same list of message histories that was input, but marked with an attribute that contains a batch-id from the Claude API as well as unique names for each list element that can be used to stitch together messages with replies, once they are ready. If you provide a named list of messages, tidyllm will use these names as identifiers in the batch, if these names are unique.

Tip: Saving batch requests to a file allows you to persist them across R sessions, making it easier to manage large jobs and access results later.

After sending a batch request, you can check its status with check_batch(). For example:

#Check the status of the batch
readRDS("claude_batch.rds") |>
   check_batch(claude())

## # A tibble: 1 × 8
##   batch_id          status created_at          expires_at          req_succeeded
##   <chr>             <chr>  <dttm>              <dttm>                      <dbl>
## 1 msgbatch_02A1B2C… ended  2024-11-01 10:30:00 2024-11-02 10:30:00             3
## # ℹ 3 more variables: req_errored <dbl>, req_expired <dbl>, req_canceled <dbl>

The status output shows details such as the number of successful, errored, expired, and canceled requests in the batch, as well as the current status. You can also see all your batch requests with list_batches() (or in the batches dashboard of the your API provider). Once the processing of a batch is completed you can fetch its results with fetch_batch():

conversations <- readRDS("claude_batch.rds") |>
  fetch_batch(claude())

poems <- map_chr(conversations, get_reply)

The output is a list of message histories, each now updated with new assistant replies. You can further process these responses with tidyllm’s standard tools. Before launching a large batch operation, it’s good practice to run a few test requests and review outputs with the standard chat() function. This approach helps confirm that prompt settings and model configurations produce the desired responses, minimizing potential errors or resource waste.

Streaming back responses (Experimental)

chat() supports real-time streaming of reply tokens to the console while the model works with the .stream=TRUE argument in chat(), for most api providers.

llm_message("Write a lengthy magazine advertisement for an R package called tidyllm") |>
  chat(claude,.stream=TRUE)

While this feature offers slightly better feedback on model behavior in real-time, it’s not particularly useful for data-analysis workflows. Metadata is not collected for streaming mode. We consider this feature experimental and recommend using non-streaming responses for production tasks. Note that error handling in streaming callbacks varies by API and differs in quality at this time. Metadata is currently not supported for streaming responses.

Choosing the Right Model and API

tidyllm supports multiple APIs, each offering distinct large language models with varying strengths. The choice of which model or API to use often depends on the specific task, cost considerations, and data privacy concerns.

1. OpenAI API: Models by OpenAI API, particularly the GPT-4o model, are extremely versatile and perform well across a wide range of tasks, including text generation, code completion, and multimodal analysis. In addition the o1-reasoning models offer very good performance for a set of specific task (at a relatively high price). There is also an azure_openai() provider function if you prefer to use the OpenAI API on Microsoft Azure.

2. Anthropic API: Claude is known for generating thoughtful, nuanced responses, making it ideal for tasks that require more human-like reasoning, such as summarization or creative writing. However, it can sometimes be more verbose than necessary, and it lacks direct JSON support, which requires additional prompting and validation to ensure structured output.

3. Google Gemini API: Google Gemini is great for long-context tasks — it can handle up to a million tokens! In addition, you can use the .grounding_threshold-parameter in the gemini_chat() function to ground responses based on Google searches. The lower the threshold is the more Gemini relies on the search instead of its internal knowledge:

llm_message("What is tidyllm and who maintains this package?") |>
  gemini_chat(.grounding_threshold = 0.3)

Moreover, with the Gemini API you are able to upload a wide range of media files and use them in the prompts of your models with functions like gemini_upload_file(). Using this, Gemini models can be used to process video and audio together with your messages. Here is an example of summarizing a speech:

#Upload a file for use with gemini
upload_info <- gemini_upload_file("example.mp3")

#Make the file available during a Gemini API call
llm_message("Summarize this speech") |>
  chat(gemini(.fileid = upload_info$name))
  
#Delte the file from the Google servers after you are done
gemini_delete_file(upload_info$name)

4. Mistral API (EU-based): Mistral offers lighter-weight, open-source models developed and hosted in the EU, making it particularly appealing if data protection (e.g., GDPR compliance) is a concern. While the models may not be as powerful as GPT-4o or Claude Sonnet, Mistral offers good performance for standard text generation tasks.

5. Groq API (Fast): Groq offers a unique advantage with its custom AI accelerator hardware, that get you the fastest output available on any API. It delivers high performance at low costs, especially for tasks that require fast execution. It hosts many strong open-source models, like lamma3:70b. There is also a groq_transcribe() function available that allows you to transcribe audio files with the Whipser-Large model on the Groq API.

6. Perplexity API (Search and Citations): Perplexity combines current finetuned Llama models with real-time web search capabilities. This allows for up-to-date information retrieval and integration into responses. All answers contain links to citations which can be accesed via the api_specific column of get_metadata()

7. Ollama (Local Models): If data privacy is a priority, running open-source models like gemma2::9B locally via ollama gives you full control over model execution and data. However, the trade-off is that local models require significant computational resources, and are often not quite as powerful as the large API-providers. The ollama blog regularly has posts about new models and their advantages that you can download via ollama_download_model().

8. Other OpenAI-compatible Local Models: Besides ollama, there are many solutions to run local models that are mostly compatible to the OpenAI API like llama.cpp, vllm and many more. To use such an API you can set the base url of the api with .api_url as well as the path to the model-endpoint with .api_path argument in the openai() provider function. Set .compatible=TRUE to skip api-key checks and rate-limit tracking. Compatibility with local models solutions may vary depending on the specific API’s implementation, and full functionality cannot be guaranteed. Ideally you can save complicated configurations like these in an object:

my_provider <- openai(.model="llama3.2:90b",
          .api_url="http://localhost:11434",
          .compatible = TRUE,
          .api_path = "/v1/chat/custom/"
          )

llm_message("Hi there") |>
  chat(my_provider)

Setting a Default Provider

You can also specify default provides with options:

# Set default providers
#chat provider
options(tidyllm_chat_default = openai(.model = "gpt-4o"))
#embedding provider
options(tidyllm_embed_default = ollama(.model = "all-minilm"))
#send batch provider
options(tidyllm_sbatch_default = claude(.temperature=0))
#check batch provider
options(tidyllm_cbatch_default = claude())
#fetch batch provider
options(tidyllm_fbatch_default = claude())
#List batches provider
options(tidyllm_lbatch_default = claude())

# Now you can use chat() or embed() without explicitly specifying a provider
conversation <- llm_message("Hello, what is the weather today?") |> 
  chat()

embeddings <- c("What is AI?", "Define machine learning.") |> 
  embed()

# Now you can use batch functions without explicitly specifying a provider
batch_messages <- list(
  llm_message("Write a poem about the sea."),
  llm_message("Summarize the theory of relativity."),
  llm_message("Invent a name for a new genre of music.")
)

# Send batch using default for send_batch()
batch_results <- batch_messages |> send_batch()

# Check batch status using default for check_batch()
status <- check_batch(batch_results)

# Fetch completed results using default for fetch_batch()
completed_results <- fetch_batch(batch_results)

# List all batches using default for list_batches()
all_batches <- list_batches()