#!/usr/bin/env awk -f

# Step 1: Initialize stop words to filter out common noise
BEGIN {
    split("the and of to a in is that it on for as was with at by an be this are from or can your layout", words)
    for (i in words) stopwords[tolower(words[i])] = 1
    
    # Store the user's question from the external environment variable
    query = tolower(ENVIRON["QUERY"])
    patsplit(query, query_tokens, /[a-zA-Z0-9]+/)
    for (i in query_tokens) {
        if (!stopwords[query_tokens[i]]) {
            query_weights[query_tokens[i]]++
        }
    }
}

# Step 2: Chunk documents by empty lines (Paragraph mode)
BEGIN { RS = "" } 

{
    # Clean up formatting whitespace
    gsub(/[ \t\r\n]+/, " ")
    paragraph = $0
    score = 0
    
    # Tokenize the paragraph text
    patsplit(tolower(paragraph), p_tokens, /[a-zA-Z0-9]+/)
    
    # Count term frequencies in this paragraph
    delete tf
    for (i in p_tokens) {
        if (!stopwords[p_tokens[i]]) tf[p_tokens[i]]++
    }
    
    # Calculate text match score using a basic vector dot product
    for (token in query_weights) {
        if (token in tf) {
            score += query_weights[token] * tf[token]
        }
    }
    
    # If this paragraph scores higher than previous ones, save it
    if (score > max_score) {
        max_score = score
        best_paragraph = paragraph
    }
}

# Step 3: Format the context directly for llama.cpp
END {
    if (max_score == 0) {
        print "Context: (No highly relevant context found in documents.)"
    } else {
        print "Context: " best_paragraph
    }
    print "Question: " ENVIRON["QUERY"]
    print "Answer: "
}