evaluator.py

from sentence_transformers import SentenceTransformer
from sklearn.metrics.pairwise import cosine_similarity
from backend import run_llm

embed_model = SentenceTransformer("all-MiniLM-L6-v2")

def similarity_score(output: str, reference: str) -> float:
    if not reference.strip():
        return 0.0
    e1 = embed_model.encode([output])
    e2 = embed_model.encode([reference])
    return float(cosine_similarity(e1, e2)[0][0])

def conciseness_score(text: str) -> float:
    words = len(text.split())
    return max(0.0, 1 - (words / 300))

def llm_judge_score(output: str, goal="overall quality") -> float:
    try:
        judge_prompt = f"""
Score the following answer from 1 to 10 based on {goal}.
Only return a single number.

Answer:
{output}
"""
        score = run_llm(judge_prompt)
        return float(score.strip())
    except:
        return 5.0  # safe fallback

def final_score(output: str, reference: str = "") -> float:
    llm_score = llm_judge_score(output) / 10
    conc = conciseness_score(output)

    if reference.strip():
        sim = similarity_score(output, reference)
        score = 0.5 * llm_score + 0.3 * sim + 0.2 * conc
    else:
        score = 0.7 * llm_score + 0.3 * conc

    return round(score, 3)