Functions to use Gemini from R. In development. Currently supports:
See examples below.
Best install from r-universe:
install.packages('rgemini', repos = c('https://skranz.r-universe.dev', 'https://cloud.r-project.org'))
Otherwise try to install directly from Github:
if (!requireNamespace("remotes", quietly = TRUE)) { install.packages("remotes") }
remotes::install_github("skranz/rgemini")
Load library and specify API key.
library(rgemini)
set_gemini_api_key("<YOUR GEMINI API KEY>")
A simple prompt.
run_gemini("Tell a joke.")
## [1] "Why don't scientists trust atoms?\n\nBecause they make up everything!\n"
Return more details of Gemini API response.
details = run_gemini("Tell a joke.",detailed_results = TRUE)
# str(details)
The returned details are a large list with all sort of information.
run_gemini("Tell 2 jokes. Return JSON with fields 'topic' and 'joke'.",json_mode = TRUE)
## topic joke
## 1 Programming Why do programmers prefer dark mode? Because light attracts bugs!
## 2 Math Why was six afraid of seven? Because seven eight nine!
The Gemini API also allows to provide JSON response schemas (see https://json-schema.org/learn/getting-started-step-by-step) for better guarantees that your JSON output satisfies a desired structure.
rgemini has small helper functions arr_resp, obj_resp and response_schema to build simple json schemas from an example.
prompt = "List 3 asian countries, their capital, the most famous building and the countries' inhabitants in million."
# Creates a schema from an example
schema = response_schema(arr_resp(obj_resp(capital = "Paris", country="France", famous_building="Eiffel Tower", population = 60.1)))
run_gemini(prompt = prompt,response_schema = schema)
## capital country famous_building population
## 1 Tokyo Japan Tokyo Skytree 125.70
## 2 Beijing China Great Wall of China 1453.00
## 3 Seoul South Korea Gyeongbokgung Palace 51.75
Here is a more comples nested schema. Will return nested tibbles.
prompt = "Show info for one african country, its capital with name and population in mio, the most famous building and inhabitants in million. Add three facts about the country."
# obj_resp expects to return a single object
# arr_resp expected a list of objects
# Both can be nested
schema = response_schema(obj_resp(
capital = obj_resp(capital="Paris", cap_pop=5),
country="France", famous_building="Eiffel Tower",
population = 60.2,
facts = arr_resp(obj_resp(factno=1L, name="fact1", descr="fact_description"))
))
# For this schema run_gemini currently
# returns a list (obj_resp) or data frame (arr_resp) with nested data frames
res = run_gemini(prompt = prompt,response_schema = schema)
str(res)
## List of 5
## $ capital :List of 2
## ..$ cap_pop: int 4
## ..$ capital: chr "Nairobi"
## $ country : chr "Kenya"
## $ facts :'data.frame': 3 obs. of 3 variables:
## ..$ descr : chr [1:3] "Kenya is known for its safaris and diverse wildlife reserves." "Kenya is home to numerous world-renowned athletes, particularly in long-distance running." "Kenya is a major producer of coffee and tea, contributing significantly to its economy."
## ..$ factno: int [1:3] 1 2 3
## ..$ name : chr [1:3] "Wildlife Safaris" "Athletic Prowess" "Agricultural Exports"
## $ famous_building: chr "Kenyatta International Conference Centre"
## $ population : int 55
Two things to note:
From my experience with larger tasks providing a schema does not always improve things, I often got better output without a schema. But that might change.
I am currently building the package DataSchema (https://github.com/skranz/DataSchema) that will allow more complex schema specifications.
That is the image we upload:
img_file = paste0("~/repbox/gemini/word_img.png")
media <- gemini_media_upload(img_file)
run_gemini("Please write down all words you can detect in the image.", media=media)
## [1] "Here are the words I can detect in the image:\n\n* Music\n* Bach\n* Economics\n* Physics"
files = c("docs/word_img.png", "docs/colors_pdf.pdf")
media <- gemini_media_upload(files)
run_gemini("Please write down all words you can detect in the uploaded pdf and image.", media=media)
## [1] "Here are the words I detected in the images:\n\n**From the first image:**\n\n* Music\n* Bach\n* Economics\n* Physics\n\n**From the second image:**\n\n* blue\n* greed\n* red\n* white\n* 1"
The gemini API also allows context caching (see https://ai.google.dev/gemini-api/docs/caching). Context caching can e.g. be helpful if you have repeated prompts to the same large PDF document (or a set of PDF documents).
To use context caching you can generate a context object with gemini_context and pass it to gemini_run. I have not yet made a nice example for this README though.
Use a gemini_list_models() function to get a data frame with information about all available models at the Gemini API.
gemini_list_models() %>%
filter(startsWith(displayName, "Gemini 2"))
## Warning: Outer names are only allowed for unnamed scalar atomic inputs
## # A tibble: 25 × 11
## name version displayName description inputTokenLimit outputTokenLimit
## <chr> <chr> <chr> <chr> <int> <int>
## 1 models… 2.0 Gemini 2.0 … Gemini 2.0 Fla… 1048576 8192
## 2 models… 2.0 Gemini 2.0 … Gemini 2.0 Fla… 1048576 8192
## 3 models… 2.0 Gemini 2.0 … Gemini 2.0 Fla… 1048576 8192
## 4 models… 2.0 Gemini 2.0 … Gemini 2.0 Fla… 1048576 8192
## 5 models… 2.0 Gemini 2.0 … Gemini 2.0 Fla… 1048576 8192
## 6 models… 2.0 Gemini 2.0 … Stable version… 1048576 8192
## 7 models… 2.0 Gemini 2.0 … Stable version… 1048576 8192
## 8 models… 2.0 Gemini 2.0 … Stable version… 1048576 8192
## 9 models… 2.0 Gemini 2.0 … Stable version… 1048576 8192
## 10 models… 2.0 Gemini 2.0 … Gemini 2.0 Fla… 1048576 8192
## # … with 15 more rows, and 5 more variables: supportedGenerationMethods <list>,
## # temperature <dbl>, topP <dbl>, topK <int>, maxTemperature <int>