AI Document Question Answering Project

"""Answer a question by returning the most relevant source sentence."""

from __future__ import annotations

from dataclasses import dataclass
import re
from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.metrics.pairwise import cosine_similarity


DOCUMENTS = {
    "attendance": "The attendance requirement is at least 80 percent of scheduled sessions. Approved exceptions are reviewed by the training team.",
    "projects": "Portfolio projects are reviewed during designated mentoring sessions. Feedback focuses on correctness, documentation, and presentation.",
    "support": "Technical questions can be submitted through the published support channel. Response time depends on course schedule and issue complexity.",
}


@dataclass(frozen=True)
class Answer:
    text: str
    source: str | None
    score: float


def build_sentence_index(documents: dict[str, str] = DOCUMENTS):
    rows = []
    for source, text in documents.items():
        for sentence in re.split(r"(?<=[.!?])\s+", text):
            if sentence.strip():
                rows.append((source, sentence.strip()))
    vectorizer = TfidfVectorizer(stop_words="english", ngram_range=(1, 2))
    matrix = vectorizer.fit_transform([row[1] for row in rows])
    return rows, vectorizer, matrix


def answer_question(question: str, threshold: float = 0.12) -> Answer:
    rows, vectorizer, matrix = build_sentence_index()
    scores = cosine_similarity(vectorizer.transform([question]), matrix)[0]
    index = int(scores.argmax())
    score = float(scores[index])
    if score < threshold:
        return Answer("The supplied documents do not contain a confident answer.", None, score)
    return Answer(rows[index][1], rows[index][0], score)


def main() -> None:
    for question in ("What attendance is required?", "When are projects reviewed?", "What is the parking fee?"):
        print(question, "->", answer_question(question))


if __name__ == "__main__":
    main()