Search_guided_GPT.r

Sys.setenv( # get an API key here: https://platform.openai.com/account/api-keys
  OPENAI_API_KEY = 'YOUR_API_KEY_HERE'
)


### Make a text "database" to search:
library(tm)
library(dplyr)
library(corpus)
library(rjson)
library(stringi)
library(reticulate)
library(openai)

# Check if python is available, some help setting this up is here:  https://rpubs.com/eR_ic/transfoRmers
reticulate::py_available()

# Importing 🤗 transformers into R session (source: https://rpubs.com/eR_ic/transfoRmers)
transformers <- reticulate::import("transformers")

# read the news articles, source of news: https://www.kaggle.com/competitions/learn-ai-bbc/data
BBC.News.Train <- read.csv("BBC_News_Train.csv")



# Search based on this example: http://www.dataperspective.info/2017/11/information-retrieval-document-search-using-vector-space-model-in-r.html

### Function to pick documents related to your query:
doc.picker <- function(query,docs){ 
  
  length <- nrow(docs)
  ds <- DataframeSource(docs)
  x <- Corpus(ds)
  
  my.corpus <- x
  #remove punctuation
  my.corpus <- tm_map(my.corpus, removePunctuation)
  
  #remove numbers, uppercase, additional spaces
  my.corpus <- tm_map(my.corpus, removeNumbers)
  my.corpus <- tm_map(my.corpus, content_transformer(tolower))
  my.corpus <- tm_map(my.corpus, stripWhitespace)
  
  #create document matrix in a format that is efficient
  term.doc.matrix.stm <- TermDocumentMatrix(my.corpus)
  
  #constructing the Vector Space Model
  get.tf.idf.weights <- function(tf.vec) {
    # Compute tfidf weights from term frequency vector
    n.docs <- length(tf.vec)
    doc.frequency <- length(tf.vec[tf.vec > 0])
    weights <- rep(0, length(tf.vec))
    weights[tf.vec > 0] <- (1 + log2(tf.vec[tf.vec > 0])) * log2(n.docs/doc.frequency)
    return(weights)
  }
  
  # normalized frequency of words in each document:
  tfidf.matrix <- t(apply(term.doc.matrix.stm, 1,
                          FUN = function(row) {get.tf.idf.weights(row)}))
  
  colnames(tfidf.matrix) <- colnames(term.doc.matrix.stm)
  
  
  tfidf.matrix <- scale(tfidf.matrix, center = FALSE,
                        scale = sqrt(colSums(tfidf.matrix^2)))
  
  
  # split Q from rest of documents:
  query.vector <- tfidf.matrix[,length ]
  tfidf.matrix <- tfidf.matrix[, 1:(length -1)]
  
  # score the documents:
  doc.scores <- t(query.vector) %*% tfidf.matrix
  
  # collect results
  results.df <- data.frame(doc = docs[1:(length-1),]$doc_id, score = t(doc.scores))
  # rank the docs:
  results.df <- results.df[order(results.df$score, decreasing = TRUE), ]
  # return the docs but ranked:
  results.df
}

# Perform a very poor news corpus search
news.search <- function(query,docs=BBC.News.Train){
  
  # reead the document corpus, append the search term:
  
  docs <- data.frame(doc_id = c(BBC.News.Train[,1],9999),
                     text = c(BBC.News.Train[,2],query),
                     dmeta1 = c(BBC.News.Train[,3],query),
                     stringsAsFactors = FALSE)
  
  
  # search all docs in the newslibrary:
  ranked.docs <-  doc.picker(query = query,texts = texts,docs=docs)
  
  docs[docs[,1] %in% ranked.docs[1:3,1],2]
}

# ask OpenAI's GPT a question, and provide it documents as context:
news.search.gpt <- function(query){
  
  question = query
  context = news.search(query = query)
  
  # this is what we will ask GPT 3.5:
  prompt = paste0("you are a chatbot anwsering a question about the news, 
                  you are provided with both a question, and for context some related news articles from the BBC, 
                  use the context provided in formulating an anwser. 
                  The user question is:",question,". 
                  The following information is the context for anwsering the question:")
  
  # here we mush together all the context:
  full.prompt <- paste(prompt,paste(context[1],context[2],context[3]))
  
  # Get an Awnser from OpenAI:
  anwser <- create_completion(
    model = "text-davinci-003",
    prompt = full.prompt,
    max_tokens = 500
  )
  
  list( GPT_response = anwser$choices, actual_context_provided = context)
  
}  


### Huggin Face implementation, this uses open source language models, not OpenAI:
news.seach.hf <- function(query){
  # Specify task
  reader <- transformers$pipeline(task = "question-answering", model = "deepset/minilm-uncased-squad2")
  searched <- news.search(query = query) 
  context <- paste0(searched[1],searched[2],searched[3])
  outputs <- reader(question = query, context = context)
  outputs
}


# get an awnser from a different AI model:
anwser <- news.seach.hf("Why did Alex Ferguson want Thierry Henry punished in 2004?")
anwser # o wow not great...

# Get an Answer from OpenAI with outr context:
news.search.gpt("Why did Alex Ferguson want Thierry Henry punished in 2004?") # context aware implementation

# Get an Answer from Vanilla GPT 3.5 without context:
create_completion(model = "text-davinci-003", prompt = "Why did Alex Ferguson want Thierry Henry punished in 2004?", max_tokens = 500)



# get an awnser from a different AI model:
anwser <- news.seach.hf("What kind of celebrity innitiated efforts where made to raise money for the asian christmas day tsunami aid effort?")
anwser # o wow not great...

# Get an Awnser from OpenAI with our context:
news.search.gpt("What kind of celebrity innitiated efforts where made to raise money for the asian christmas day tsunami aid effort?") # context aware implementation

# Get an Awnser from Vanilla GPT 3.5 without context:
create_completion(model = "text-davinci-003", prompt = "What kind of celebrity innitiated efforts where made to raise money for the asian christmas day tsunami aid effort?", max_tokens = 500)