Update app.py

app.py

@@ -1,39 +1,181 @@
+# Zawiyah AI Collective — Humanizer AI (Style 4 + Detector)
+# ----------------------------------------------------------
+# Requirements:
+#   pip install gradio openai
+# Environment:
+#   export OPENAI_API_KEY="sk-..."
+# Optional:
+#   export OPENAI_BASE_URL="https://api.openai.com/v1"   # if using default, you can skip
+
+import os
+import json
 import gradio as gr
-import random
+from openai import OpenAI
+
+# --- CONFIG ---
+MODEL_HUMANIZER = "gpt-5"          # use your strongest model name here
+MODEL_DETECTOR  = "gpt-5"          # same model is fine for classification
+TEMPERATURE_REWRITE = 0.6          # a bit of variation, still controlled
+MAX_TOKENS_REWRITE  = 1200         # adjust per your quotas
+
+# --- CLIENT ---
+def get_client():
+    # Allows custom base (useful for proxies / gateways), otherwise defaults to api.openai.com
+    base_url = os.environ.get("OPENAI_BASE_URL", None)
+    if base_url:
+        return OpenAI(api_key=os.environ.get("OPENAI_API_KEY"), base_url=base_url)
+    return OpenAI(api_key=os.environ.get("OPENAI_API_KEY"))
+
+# --- PROMPTS ---
+SYSTEM_HUMANIZER = (
+    "You are Zawiyah AI Humanizer. Rewrite text into neutral, balanced human English.\n"
+    "Goals: preserve meaning; improve clarity; vary sentence length; prefer concrete verbs; reduce filler.\n"
+    "Keep citations/URLs; maintain lists; remove obvious LLM tells (e.g., 'as an AI', 'in conclusion' unless necessary).\n"
+    "Avoid plagiarism; do not invent facts; keep numbers and named entities unchanged unless the user text is ambiguous.\n"
+    "Tone: natural, calm, and human—no corporate buzzwords, no robotic cadence. Keep it concise but not terse."
+)
 
-# Dummy logic (replace with your model later)
-def humanize_and_detect(text):
-    # Simulated refined output
-    refined = f"This is the humanized version of your text:\n\n{text}"
+USER_HUMANIZER_TEMPLATE = (
+    "Rewrite the text below into natural, balanced human English suited for a wide audience.\n"
+    "Constraints:\n"
+    "- Preserve all factual content, names, figures, and citations.\n"
+    "- Remove repetitive phrasing and overly formal constructions.\n"
+    "- Prefer contractions where natural (I'm, we'll) and mix short & long sentences.\n"
+    "- Keep formatting if present (bullet lists, headings) but tidy it up.\n\n"
+    "Text:\n```text\n{src}\n```"
+)
 
-    # Simulated AI-likeness score
-    score = random.uniform(0.3, 0.95)
-    ai_score = f"AI-likeness score: {score:.2f} ({'Likely Human' if score < 0.5 else 'Likely AI'})"
+SYSTEM_DETECTOR = (
+    "You are an AI-likeness detector. Given ORIGINAL and REWRITTEN text, estimate the likelihood the text "
+    "was produced by an AI writing model. Return a **strict JSON** object with keys:\n"
+    '{\n'
+    '  "ai_likeness_score": float in [0,1],  // higher = more likely AI\n'
+    '  "verdict": "Likely Human" | "Unclear" | "Likely AI",\n'
+    '  "rationale": string (1-2 concise reasons),\n'
+    '  "signals": [ up to 5 short phrases e.g., "uniform sentence length", "hedging", "template phrases" ]\n'
+    '}\n'
+    "Focus on stylistic signals (cadence, repetition, template markers), not topic.\n"
+    "Be conservative—when uncertain, push score toward 0.5 and verdict 'Unclear'."
+)
 
-    return refined, ai_score
+USER_DETECTOR_TEMPLATE = (
+    "ORIGINAL:\n```text\n{orig}\n```\n\n"
+    "REWRITTEN:\n```text\n{rew}\n```\n\n"
+    "Return only the JSON. No prose."
+)
 
-# Minimal Gradio interface
-with gr.Blocks() as demo:
+# --- CORE FUNCTIONS ---
+def run_humanizer(text: str) -> str:
+    """Rewrite into neutral, balanced human English."""
+    if not text or not text.strip():
+        return "Please paste some text above."
+    client = get_client()
+    resp = client.chat.completions.create(
+        model=MODEL_HUMANIZER,
+        temperature=TEMPERATURE_REWRITE,
+        max_tokens=MAX_TOKENS_REWRITE,
+        messages=[
+            {"role": "system", "content": SYSTEM_HUMANIZER},
+            {"role": "user",   "content": USER_HUMANIZER_TEMPLATE.format(src=text)}
+        ]
+    )
+    return resp.choices[0].message.content.strip()
+
+def run_detector(original_text: str, rewritten_text: str) -> dict:
+    """Return dict: {ai_likeness_score, verdict, rationale, signals}"""
+    if not rewritten_text.strip():
+        return {
+            "ai_likeness_score": 0.5,
+            "verdict": "Unclear",
+            "rationale": "No rewritten content provided.",
+            "signals": []
+        }
+    client = get_client()
+    resp = client.chat.completions.create(
+        model=MODEL_DETECTOR,
+        temperature=0.0,
+        max_tokens=400,
+        messages=[
+            {"role": "system", "content": SYSTEM_DETECTOR},
+            {"role": "user",   "content": USER_DETECTOR_TEMPLATE.format(orig=original_text, rew=rewritten_text)}
+        ]
+    )
+    raw = resp.choices[0].message.content.strip()
+    # Safe JSON parsing
+    try:
+        data = json.loads(raw)
+        # Minimal validation
+        score = float(data.get("ai_likeness_score", 0.5))
+        verdict = str(data.get("verdict", "Unclear"))
+        rationale = str(data.get("rationale", ""))
+        signals = data.get("signals", [])
+        if not isinstance(signals, list):
+            signals = []
+        return {
+            "ai_likeness_score": max(0.0, min(1.0, score)),
+            "verdict": verdict,
+            "rationale": rationale,
+            "signals": signals[:5]
+        }
+    except Exception:
+        # Fallback: return raw content in rationale for debugging
+        return {
+            "ai_likeness_score": 0.5,
+            "verdict": "Unclear",
+            "rationale": f"Non-JSON response from model: {raw[:300]}",
+            "signals": []
+        }
+
+def humanize_and_detect_pipeline(text: str):
+    try:
+        rewritten = run_humanizer(text)
+        det = run_detector(text, rewritten)
+        # Pretty print detector for the right-hand panel
+        verdict_line = f"{det['verdict']} (score: {det['ai_likeness_score']:.2f})"
+        details = "• " + "; ".join(det["signals"]) if det["signals"] else ""
+        detector_out = f"{verdict_line}\n{det['rationale']}\n{details}"
+        return rewritten, detector_out
+    except Exception as e:
+        return (
+            "⚠️ Error during processing. Please check your API key / quota.",
+            f"Error: {type(e).__name__}: {str(e)}"
+        )
+
+# --- UI ---
+custom_css = """
+#title {text-align:center; font-size:32px; font-weight:800;}
+.gradio-container {max-width:880px !important; margin:auto;}
+.section-box {border:1px solid #e6e6e6; padding:18px; border-radius:16px; background:#fafafa;}
+#footer {text-align:center; font-size:12px; color:#666; margin-top:12px;}
+"""
+
+with gr.Blocks(css=custom_css, theme=gr.themes.Soft()) as demo:
     gr.Markdown(
         """
-        # 🕌 Zawiyah AI Collective – Humanizer AI  
+        <div id="title">🕌 Zawiyah AI Collective — <span style='color:#8E0000'>Humanizer AI</span></div>
+        <p style="text-align:center; font-size:16px; margin-top:-8px">
         Refine your writing & detect AI-likeness. Simple, clean, effective.
-
-        ---
+        </p>
+        <hr>
         """
     )
 
-    input_text = gr.Textbox(label="✍️ Paste Your Text Here", placeholder="Enter text...", lines=6)
-
-    submit_btn = gr.Button("🚀 Refine & Detect")
+    with gr.Column(elem_classes="section-box"):
+        input_text = gr.Textbox(label="✍️ Paste Your Text Here", placeholder="Enter text…", lines=10)
+        submit_btn = gr.Button("🚀 Refine & Detect", size="lg")
 
-    output_refined = gr.Textbox(label="🧠 Humanized Output", lines=6)
-    ai_score = gr.Textbox(label="🤖 AI Detection Result")
+    with gr.Row():
+        with gr.Column(elem_classes="section-box"):
+            output_refined = gr.Textbox(label="🧠 Humanized Output", lines=12)
+        with gr.Column(elem_classes="section-box"):
+            ai_score = gr.Textbox(label="🤖 AI Detection Result", lines=12)
 
-    submit_btn.click(fn=humanize_and_detect,
+    submit_btn.click(fn=humanize_and_detect_pipeline,
                      inputs=input_text,
                      outputs=[output_refined, ai_score])
 
-demo.queue().launch()
-
+    gr.Markdown("<div id='footer'>Zawiyah AI Collective · Malaysia · v1.0 BETA · Built with Gradio</div>")
 
+if __name__ == "__main__":
+    # On Hugging Face Spaces, don't set share=True. Local dev: share=True if you want a public URL.
+    demo.queue().launch()