"""Redac — a local privacy gateway.

Two entry points, one redaction core:
  - Text:  paste confidential text -> detect PII -> reversibly redact.
  - Image: upload a document/ID image -> a local vision model extracts the
           fields -> the extracted text is redacted the same way.

The redacted text is what you would safely hand to a downstream LLM; the
mapping stays local so any answer can be rehydrated. All compute runs
in-Space (vision model on ZeroGPU). No external API calls.
"""

import gradio as gr

from redac import (
    detect_entities,
    redact,
    rehydrate,
    extract_text_from_image,
    DEFAULT_LABELS,
)
from redac.detect import gliner_available

import inspect

# show_copy_button exists in Gradio 5.x; in 6.x copying is the default and the
# kwarg was removed. Pass it only when supported so the app runs on both.
_COPY = (
    {"show_copy_button": True}
    if "show_copy_button" in inspect.signature(gr.Textbox.__init__).parameters
    else {}
)

_NER_NOTE = (
    "" if gliner_available()
    else "\n\n> ⚠️ GLiNER not installed: running **regex-only** (structured "
    "IDs like email/phone/IBAN). Names, addresses and other free-text PII "
    "need GLiNER, which is available on the Space."
)

EXAMPLE = (
    "Patient John A. Doe, DOB 1985-04-12, was admitted on 2026-06-01. "
    "Contact: john.doe@example.com, +49 151 23456789. "
    "Insurance ID 123-45-6789, IBAN DE89370400440532013000."
)


def _redact_text(text, labels, threshold):
    entities = detect_entities(text, labels=labels or DEFAULT_LABELS, threshold=threshold)
    redacted, mapping = redact(text, entities)
    table = [[e.label, e.text, f"{e.score:.2f}", e.source] for e in entities]
    summary = f"{len(entities)} PII span(s) redacted into {len(mapping)} placeholder(s)."
    return redacted, table, summary, mapping


def run_text(text, labels, threshold):
    return _redact_text(text, labels, threshold)


def run_image(image, labels, threshold):
    extracted = extract_text_from_image(image)
    if not extracted:
        return "", "", [], "No image / nothing extracted.", {}
    redacted, table, summary, mapping = _redact_text(extracted, labels, threshold)
    return extracted, redacted, table, summary, mapping


def do_rehydrate(redacted_text, mapping):
    if not mapping:
        return "Run a redaction first."
    return rehydrate(redacted_text, mapping)


with gr.Blocks(title="Redac", theme=gr.themes.Soft()) as demo:
    gr.Markdown(
        "# 🖍️ Redac\n"
        "**A local privacy gateway.** Upload a document or paste text, Redac "
        "extracts and redacts the PII *locally*, and you copy the safe output to "
        "use with any LLM. Raw values never leave: the mapping stays local so you "
        "can rehydrate answers yourself." + _NER_NOTE
    )

    with gr.Tabs():
        # --- Text tab --------------------------------------------------------
        with gr.Tab("Text"):
            t_map = gr.State({})
            with gr.Row():
                with gr.Column():
                    t_in = gr.Textbox(
                        label="Confidential text",
                        placeholder="Paste a document, message, or record...",
                        lines=10,
                        value=EXAMPLE,
                    )
                    t_labels = gr.Dropdown(
                        label="PII types", choices=DEFAULT_LABELS,
                        value=DEFAULT_LABELS, multiselect=True,
                    )
                    t_thr = gr.Slider(0.1, 0.9, value=0.45, step=0.05, label="Threshold")
                    t_btn = gr.Button("Redact", variant="primary")
                with gr.Column():
                    t_out = gr.Textbox(
                        label="✅ Safe output — copy this into your LLM (no raw PII)",
                        lines=10,
                        **_COPY,
                    )
                    t_sum = gr.Markdown()
                    t_tab = gr.Dataframe(
                        headers=["type", "value", "score", "source"],
                        label="Detected PII", wrap=True,
                    )
                    with gr.Accordion("Rehydrate (local only)", open=False):
                        t_reh_btn = gr.Button("Restore original values")
                        t_reh = gr.Textbox(label="Rehydrated", lines=6)

            t_btn.click(run_text, [t_in, t_labels, t_thr], [t_out, t_tab, t_sum, t_map])
            t_reh_btn.click(do_rehydrate, [t_out, t_map], [t_reh])

        # --- Image tab -------------------------------------------------------
        with gr.Tab("Image"):
            gr.Markdown(
                "Upload a document or ID image. A local vision model "
                "(MiniCPM-V-4.5) extracts the fields, then Redac redacts them."
            )
            i_map = gr.State({})
            with gr.Row():
                with gr.Column():
                    i_in = gr.Image(label="Document / ID image", type="pil")
                    i_labels = gr.Dropdown(
                        label="PII types", choices=DEFAULT_LABELS,
                        value=DEFAULT_LABELS, multiselect=True,
                    )
                    i_thr = gr.Slider(0.1, 0.9, value=0.45, step=0.05, label="Threshold")
                    i_btn = gr.Button("Extract & redact", variant="primary")
                with gr.Column():
                    i_extracted = gr.Textbox(
                        label="🔒 Extracted fields (raw — stays local, never sent)",
                        lines=8,
                    )
                    i_out = gr.Textbox(
                        label="✅ Safe output — copy this into your LLM (no raw PII)",
                        lines=8,
                        **_COPY,
                    )
                    i_sum = gr.Markdown()
                    i_tab = gr.Dataframe(
                        headers=["type", "value", "score", "source"],
                        label="Detected PII", wrap=True,
                    )
                    with gr.Accordion("Rehydrate (local only)", open=False):
                        i_reh_btn = gr.Button("Restore original values")
                        i_reh = gr.Textbox(label="Rehydrated", lines=6)

            i_btn.click(
                run_image, [i_in, i_labels, i_thr],
                [i_extracted, i_out, i_tab, i_sum, i_map],
            )
            i_reh_btn.click(do_rehydrate, [i_out, i_map], [i_reh])


if __name__ == "__main__":
    import os

    port = int(os.environ.get("GRADIO_SERVER_PORT", "7860"))
    demo.launch(server_name="0.0.0.0", server_port=port)