tests/test_eval_metrics.py

"""
tests/test_eval_metrics.py
Tests for evaluation metrics (no GPU required).
"""

import sys
sys.path.insert(0, "src")

import pytest
from ankahi.eval.metrics import compute_bleu, compute_chrf


class TestBleu:
    def test_perfect_score(self):
        hyps = ["the cat sat on the mat"]
        refs = ["the cat sat on the mat"]
        score = compute_bleu(hyps, refs)
        assert score > 90.0

    def test_zero_score(self):
        hyps = ["xyz xyz xyz xyz"]
        refs = ["the cat sat on the mat"]
        score = compute_bleu(hyps, refs)
        assert score == 0.0

    def test_partial_match(self):
        hyps = ["I want water please"]
        refs = ["I want some water"]
        score = compute_bleu(hyps, refs)
        assert 0.0 < score < 100.0

    def test_corpus_level(self):
        hyps = ["Mujhe abhi pani chahiye", "Mama jaldi yahan aao"]
        refs = ["Mujhe abhi paani chahiye", "Mummy jaldi yahan aao"]
        score = compute_bleu(hyps, refs)
        assert score > 0.0

    def test_empty_strings(self):
        hyps = [""]
        refs = ["something"]
        # sacrebleu handles empty gracefully
        score = compute_bleu(hyps, refs)
        assert score == 0.0


class TestChrF:
    def test_perfect_score(self):
        hyps = ["मुझे पानी चाहिए"]
        refs = ["मुझे पानी चाहिए"]
        score = compute_chrf(hyps, refs)
        assert score > 95.0

    def test_partial_indic(self):
        hyps = ["Mujhe pani chahiye"]
        refs = ["Mujhe paani chahida"]
        score = compute_chrf(hyps, refs)
        # chrF++ rewards character-level overlaps
        assert score > 30.0

    def test_corpus_level(self):
        hyps = ["Mama mujhe bhookh lagi hai", "I am tired please"]
        refs = ["Mummy mujhe bhookh lagi hai", "I am very tired"]
        score = compute_chrf(hyps, refs)
        assert score > 40.0

    def test_returns_float(self):
        score = compute_chrf(["test"], ["reference"])
        assert isinstance(score, float)


class TestPersonaModule:
    def test_all_personas_present(self):
        from ankahi.data.persona import PERSONAS
        assert len(PERSONAS) == 5
        for pid in ["ananya", "arjun", "priya", "rohan", "zara"]:
            assert pid in PERSONAS

    def test_persona_system_prompt_nonempty(self):
        from ankahi.data.persona import PERSONAS
        for pid, persona in PERSONAS.items():
            prompt = persona.to_system_prompt()
            assert len(prompt) > 50
            assert persona.name in prompt
            assert persona.city in prompt

    def test_persona_serialization(self):
        from ankahi.data.persona import PERSONAS, Persona
        import json
        arjun = PERSONAS["arjun"]
        d = arjun.to_dict()
        assert d["name"] == "Arjun"
        assert d["primary_language"] == "pa"
        assert isinstance(d["secondary_languages"], list)

    def test_arjun_language_mix(self):
        from ankahi.data.persona import PERSONAS
        arjun = PERSONAS["arjun"]
        assert arjun.primary_language == "pa"
        assert "hi" in arjun.secondary_languages
        assert "en" in arjun.secondary_languages

    def test_mediapipe_lora_rank_constraint(self):
        """Remind us that persona lora rank must be 4 or 8."""
        # This test documents the MediaPipe constraint
        MEDIAPIPE_SUPPORTED_RANKS = [4, 8]
        PERSONA_RANK = 8  # from stage2_persona.py
        assert PERSONA_RANK in MEDIAPIPE_SUPPORTED_RANKS, \
            f"Persona LoRA rank {PERSONA_RANK} is not supported by MediaPipe ({MEDIAPIPE_SUPPORTED_RANKS})"