tests/test_llm_parsing.py

from signspeak.llm import (
    deterministic_speech_from_intent,
    enforce_intent_consistency,
    extract_json_object,
    normalize_llm_output,
    safe_json_loads,
)


def test_extract_json_object_from_markdown_fence():
    raw = """```json
{"subtitle": "I'm happy to see you!", "voice_instruction": "Speak warmly."}
```"""

    parsed = extract_json_object(raw)

    assert parsed["subtitle"] == "I'm happy to see you!"
    assert parsed["voice_instruction"] == "Speak warmly."


def test_extract_json_object_from_text_wrapper():
    raw = 'Result: {"subtitle": "Hello.", "voice_instruction": "Speak clearly."} Done.'

    parsed = extract_json_object(raw)

    assert parsed == {
        "subtitle": "Hello.",
        "voice_instruction": "Speak clearly.",
    }


def test_normalize_llm_output_blocks_json_subtitle():
    normalized = normalize_llm_output(
        {
            "subtitle": '{"subtitle": "Bad nested payload"}',
            "voice_instruction": "Speak warmly.",
        }
    )

    assert normalized["subtitle"] == "I want to say something."
    assert normalized["voice_instruction"] == "Speak warmly."


def test_safe_json_loads_falls_back_to_raw_text():
    parsed = safe_json_loads("not json")

    assert parsed["raw_input"] == "not json"
    assert "warning" in parsed


def test_deterministic_speech_maps_i_love_you():
    result = deterministic_speech_from_intent(
        {
            "detected_glosses": ["I", "LOVE", "YOU"],
            "detected_facial_expression": "neutral",
        }
    )

    assert result["subtitle"] == "I love you."


def test_deterministic_speech_does_not_hallucinate_empty_glosses():
    result = deterministic_speech_from_intent({"detected_glosses": []})

    assert result["subtitle"] == "No ASL words were detected yet."


def test_enforce_intent_consistency_corrects_wrong_love_subtitle():
    result = enforce_intent_consistency(
        {"detected_glosses": ["I", "LOVE", "YOU"]},
        {"subtitle": "I am happy to see you.", "voice_instruction": "Speak warmly."},
    )

    assert result["subtitle"] == "I love you."
    assert "consistency_warning" in result


def test_enforce_intent_consistency_corrects_wrong_where_subtitle():
    result = enforce_intent_consistency(
        {"detected_glosses": ["where"]},
        {"subtitle": "I am happy to see you.", "voice_instruction": "Speak warmly."},
    )

    assert result["subtitle"] == "Where?"
    assert result["raw_llm_subtitle"] == "I am happy to see you."