language models
see: https://magazine.sebastianraschka.com/p/understanding-large-language-models
completions
You can run completions using:
completionGPT(
system_prompt = "",
query = "",
model = "",
temperature = 0,
openai_api_key = ""
)The system_prompt tells the model how to act. For example, you might say system_prompt = "you are a helpful assistant".
The query is the question you want to ask. For example, you might say: query = "Below is some text from a scientific article, but I don't quite understand it. Could you explain it in simple terms? Text: The goal of the study presented was to compare Tanacetum balsamita L. (costmary) and Tanacetum vulgare L. (tansy) in terms of the antibacterial and antioxidant activity of their essential oils and hydroethanolic extracts, and to relate these activities with their chemical profiles. The species under investigation differed in their chemical composition and biological activities. The dominant compounds of the essential oils, as determined by Gas Chromatography-Mass Spectrometry (GC-MS), were β-thujone in costmary (84.43%) and trans-chrysanthenyl acetate in tansy (18.39%). Using High-Performance Liquid Chromatography with Diode-Array Detection (HPLC-DAD), the chemical composition of phenolic acids and flavonoids were determined. Cichoric acid was found to be the dominant phenolic compound in both species (3333.9 and 4311.3 mg per 100g, respectively). The essential oil and extract of costmary displayed stronger antibacterial activity (expressed as Minimum Inhibitory Concentration (MIC) and Minimum Bactericidal Concentration (MBC) values) than those of tansy. Conversely, tansy extract had higher antioxidant potential (determined by Ferric Reducing Antioxidant Power (FRAP) and DPPH assays) compared to costmary. In light of the observed antibacterial and antioxidant activities, the herbs of tansy and costmary could be considered as promising products for the pharmaceutical and food industries, specifically as antiseptic and preservative agents.
Available options for model include “gpt-4o”, gpt-4”, “gpt-4-0613”, “gpt-3.5-turbo”, “gpt-3.5-turbo-0613”, “gpt-3.5-turbo-16k”, “gpt-3.5-turbo-16k-0613”, “text-davinci-003”, “text-davinci-002”, “text-curie-001”, “text-babbage-001”, “text-ada-001”. Note that you may hot have API access to gpt-4 unless you requested it from OpenAI.
Temperature goes from 0 to 2 and adds randomness to the answer. Though that is a highly oversimplified and somewhat inaccurate explanation.
You have to provide your OpenAI API key.
createNewsletter
The createNewsletter() function provides an automated way to gather, organize, and share the latest research findings in specified focus areas. Using the createNewsletter function, you can input search terms that will be used as queries in a PubMed search. You also provide focus areas that gpt-3.5-turbo uses to rank the results of the PubMed searches. createNewsletter then compiles the n_articles_per_newsletter top ranked hits into a formatted newsletter. The image_prompt is used to create a header image for the newsletter. By modifying search_terms and focus_areas, you can tailor the newsletter to keep up with cutting-edge research in your field.
With createNewsletter, you can optionally also monitor a directory on the computer system you are using. Anything in the monitored_directory (and its subdirectories) that was added within the past number of hours indicated in monitored_time_window that has a suffix matching monitored_file_suffix (only .txt files currently supported) will be added to the newsletter. This allows your newsletter to include not only PubMed search results but also files stored in a specific directory (like a collection of chunked .pdf files).
bustalab = TRUE
source("https://thebustalab.github.io/phylochemistry/phylochemistry.R")
newsletter <- createNewsletter(
search_terms = c(
"plant biochemistry", "plant genomics and genetics",
"cuticular waxes", "triterpenoid yeast engineering",
"plant triterpenoids", "plant ABC transporters", "plant specialized metabolism",
"plant metabolism", "plant metabolic evolution", "plant biochemistry", "plant metabolism",
"plant genomics", "cuticular wax", "triterpenoids", "plant ABC transporters",
"plant specialized metabolism",
"biosynthesis", "metabolic evolution",
"artificial intelligence", "large language models",
"protein language models"
),
focus_areas = c(
"phytochemistry", "plant biochemistry", "plant metabolism", "plant genomics",
"cuticular wax", "triterpenoids", "plant ABC transporters",
"plant specialized metabolism",
"biosynthesis", "metabolic evolution",
"artificial intelligence", "large language models",
"protein language models"
),
n_articles_per_newsletter = 20,
pubmed_api_key = "api_key",
openai_api_key = "api_key",
image_prompt = "I NEED to test how the tool works with extremely simple prompts. DO NOT add any detail, just use it AS-IS: A modern, minimalist blog cover image for a phytochemistry blog post, WITHOUT ANY TEXT. The image should embody a minimalist aesthetic with simple, clean lines and a limited color palette. Depict abstract representations of the diverse themes in the blog post: medicinal plants, phytoremediation, plant growth, plant stress response, environmental microbiology, and plant biochemistry. Include stylized, minimalistic illustrations of Leonurus heterophyllus for medicinal plants, Ipomoea carnea for phytoremediation, a seedling for plant growth, Citrus reticulata for stress response, rice plants for arsenic stress response, abstract soil microbes for environmental microbiology, a sesame plant for disease resistance, and oat plants for cell wall synthesis. The composition should be harmoniously adjusted to ensure a balanced and aesthetically pleasing layout that captures the essence of phytochemistry and plant science research with a contemporary, streamlined look.",
monitored_directory = "/project_data/shared/general_lab_resources/literature/papers_chunked/",
monitored_time_window = 336,
monitored_file_suffix = ".1.txt"
)
names(newsletter)Finally, you can get this newsletter emailed to yourself using the following approach: (i) set up a Zapier account and a Zap that monitors a Google Sheet then send you an email with the contents of that sheet anytime it is updated; (ii) use the following code to update the sheet with the output of the newsletter; (iii) make an R script that contains the source() command, createNewsletter(), and the write_sheet command then have that script executed as a cron job daily, for example.
gs4_auth(path = "your_auth_token")
write_sheet(
data = data.frame("content" = newsletter$newsletter$response),
sheet = "Sheet1", "your_sheet_share_link"
)Here is an example of the newsletter:
prompting strategies
Role and Goal-Based Constraints: These constraints narrow the AI’s response range, making it more appropriate and effective. Leveraging the AI’s pre-trained knowledge, they guide conversation within a specific persona.
Step-by-Step Instructions: Clarity and organization in instructions are crucial. It’s recommended to use simple, direct language and to break down complex problems into steps. This approach, including the “Chain of Thought” method, helps the AI follow and effectively respond to the user’s request.
Expertise and Pedagogy: The user’s knowledge and perspective play a vital role in guiding the AI. The user should have a clear vision of how the AI should respond and interact, especially in educational or pedagogical settings.
Constraints: Setting rules or conditions within prompts helps guide the AI’s behavior and makes its responses more predictable. This includes defining roles (like a tutor), limiting response lengths, and controlling the flow of conversation.
Personalization: Using prompts that solicit information and ask questions can help the AI adapt to different scenarios and provide more personalized responses.
Examples and Few-Shot Learning: Providing the AI with a few examples helps it understand and adapt to new tasks better than with zero-shot learning.
Asking for Specific Output: Experimenting with different types of outputs, such as images, charts, or documents, can leverage the AI’s capabilities.
Appeals to Emotion: Recent research suggests that adding emotional phrases to requests can improve the quality of AI responses. Different phrases may be effective in different contexts.
Testing and Feedback: It’s important to test prompts with various inputs and perspectives to ensure they are effective and helpful. Continuous tweaking based on feedback can improve the prompts further.
Sharing and Collaboration: Sharing structured prompts allows others to learn and apply them in different contexts, fostering a collaborative environment for AI use.