app.py

import gradio as gr

from src.config import ModelSpec, RunConfig
from src.orchestrator import Orchestrator
from src.tasks import TaskContext
from src.utils import pretty_json


DEFAULT_MODELS = [
    "openai/gpt-oss-20b",          # OpenAI open-weight :contentReference[oaicite:4]{index=4}
    "deepseek-ai/DeepSeek-R1",     # DeepSeek :contentReference[oaicite:5]{index=5}
    "google/gemma-2-2b-it",        # Gemma (open) :contentReference[oaicite:6]{index=6}
    "google/gemma-3-270m",         # Gemma 3 (multimodal family) :contentReference[oaicite:7]{index=7}
]


def run_pipeline(
    req_text, img,
    analyzer_model, analyzer_backend, analyzer_temp, analyzer_tokens,
    refactor_model, refactor_backend, refactor_temp, refactor_tokens,
    critic_model, critic_backend, critic_temp, critic_tokens,
):
    if not req_text or not req_text.strip():
        return "Вставь текст требований.", "", "", "{}"

    cfg = RunConfig(
        analyzer=ModelSpec(
            model_id=analyzer_model,
            backend=analyzer_backend,
            temperature=analyzer_temp,
            max_new_tokens=analyzer_tokens,
            system_prompt="Отвечай по-русски. Будь точным, не выдумывай факты.",
        ),
        refactor=ModelSpec(
            model_id=refactor_model,
            backend=refactor_backend,
            temperature=refactor_temp,
            max_new_tokens=refactor_tokens,
            system_prompt="Отвечай по-русски. Не добавляй требований из головы.",
        ),
        critic=ModelSpec(
            model_id=critic_model,
            backend=critic_backend,
            temperature=critic_temp,
            max_new_tokens=critic_tokens,
            system_prompt="Отвечай по-русски. Будь строгим ревьюером.",
        ),
    )

    ctx = TaskContext(name="requirements_analysis_and_refactor", language="ru")
    orch = Orchestrator(cfg, ctx=ctx)

    trace = orch.run(req_text, image=img)
    d = trace.to_dict()

    return (
        trace.analyzer.output,
        trace.refactor.output,
        trace.critic.output,
        pretty_json(d),
    )


with gr.Blocks(title="Multi-agent Requirements Lab") as demo:
    gr.Markdown(
        "# Multi-agent Requirements Lab (text + diagram)\n"
        "LLM-агенты для анализа/рефакторинга требований + извлечение контекста из фото диаграммы.\n\n"
        "**Для работы через HF Inference API добавь `HF_TOKEN` в Settings → Secrets.**"
    )

    with gr.Row():
        req_text = gr.Textbox(label="Текст требований", lines=14, placeholder="Вставь требования…")

    with gr.Row():
        img = gr.Image(label="Фото/скрин диаграммы (опционально)", type="pil")

    def agent_controls(title: str, default_model: str):
        with gr.Accordion(title, open=False):
            model = gr.Dropdown(DEFAULT_MODELS, value=default_model, label="Model ID")
            backend = gr.Radio(["hf_inference_api"], value="hf_inference_api", label="Backend")
            temp = gr.Slider(0.0, 1.5, value=0.2, step=0.05, label="Temperature")
            tokens = gr.Slider(64, 2000, value=600, step=32, label="Max new tokens")
        return model, backend, temp, tokens

    analyzer_model, analyzer_backend, analyzer_temp, analyzer_tokens = agent_controls("Analyzer agent", "openai/gpt-oss-20b")
    refactor_model, refactor_backend, refactor_temp, refactor_tokens = agent_controls("Refactor agent", "google/gemma-2-2b-it")
    critic_model, critic_backend, critic_temp, critic_tokens = agent_controls("Critic agent", "deepseek-ai/DeepSeek-R1")

    run_btn = gr.Button("Run", variant="primary")

    out_an = gr.Markdown(label="Analyzer output")
    out_rf = gr.Markdown(label="Refactor output")
    out_cr = gr.Markdown(label="Critic output")

    with gr.Accordion("Trace (JSON)", open=False):
        trace_json = gr.Code(language="json")

    run_btn.click(
        fn=run_pipeline,
        inputs=[
            req_text, img,
            analyzer_model, analyzer_backend, analyzer_temp, analyzer_tokens,
            refactor_model, refactor_backend, refactor_temp, refactor_tokens,
            critic_model, critic_backend, critic_temp, critic_tokens,
        ],
        outputs=[out_an, out_rf, out_cr, trace_json],
    )

if __name__ == "__main__":
    demo.launch()