Update app.py

app.py

@@ -1,107 +1,94 @@
-import os
 from fastapi import FastAPI, HTTPException
-from pydantic import BaseModel
+from fastapi.middleware.cors import CORSMiddleware
+from pydantic import BaseModel, Field
+from typing import Literal, Optional
 from llama_cpp import Llama
-from contextlib import asynccontextmanager
-from huggingface_hub import hf_hub_download
+import re
 
-# ==================================================
-# MODEL CONFIGURATION
-# ==================================================
+# ==================== MODEL CONFIG ====================
 MODEL_REPO = "bartowski/Phi-3.1-mini-4k-instruct-GGUF"
 MODEL_FILE = "Phi-3.1-mini-4k-instruct-IQ2_M.gguf"
 
-N_THREADS = int(os.getenv("N_THREADS", "8"))
-N_CTX = int(os.getenv("N_CTX", "2048"))
-N_BATCH = int(os.getenv("N_BATCH", "256"))
-
-llm = None
-
-# ==================================================
-# FASTAPI LIFESPAN (MODEL LOAD)
-# ==================================================
-@asynccontextmanager
-async def lifespan(app: FastAPI):
-    global llm
-    try:
-        print("⏳ Downloading model...")
-        model_path = hf_hub_download(
-            repo_id=MODEL_REPO,
-            filename=MODEL_FILE,
-        )
-
-        print("✅ Model downloaded. Loading...")
-        llm = Llama(
-            model_path=model_path,
-            n_ctx=N_CTX,
-            n_threads=N_THREADS,
-            n_batch=N_BATCH,
-            verbose=False,
-        )
+print("🚀 Loading Phi-3.1 Mini (Human Authorship Restorer)...")
+llm = Llama.from_pretrained(
+    repo_id=MODEL_REPO,
+    filename=MODEL_FILE,
+    n_threads=4,
+    n_ctx=1024,      # safer for HF Spaces
+    n_batch=128,
+    n_gpu_layers=0,
+    verbose=False,
+)
+print("✅ Model loaded")
 
-        print("🚀 Model loaded successfully")
+# ==================== FASTAPI ====================
+app = FastAPI(title="AI Humanizer – Author Voice Restorer")
 
-    except Exception as e:
-        print("❌ Model load failed:", e)
-        llm = None
-
-    yield
-
-# ==================================================
-# APP INIT
-# ==================================================
-app = FastAPI(
-    title="AI Humanizer",
-    description="Academic-safe AI Humanizer",
-    version="1.0",
-    lifespan=lifespan
+app.add_middleware(
+    CORSMiddleware,
+    allow_origins=["*"],
+    allow_methods=["*"],
+    allow_headers=["*"],
 )
-@app.get("/")
-def root():
-    return {
-        "status": "ok",
-        "message": "AI Humanizer backend is running",
-        "endpoints": {
-            "humanize": "POST /humanize",
-            "auth": "GET /api/auth/verify",
-            "docs": "/docs"
-        }
-    }
 
-# ==================================================
-# DUMMY AUTH VERIFY (Frontend Fix)
-# ==================================================
-@app.get("/api/auth/verify")
-def verify_auth():
-    return {"authenticated": True}
-
-# ==================================================
-# INPUT SCHEMA
-# ==================================================
+# ==================== REQUEST ====================
 class HumanizeRequest(BaseModel):
-    text: str
-    section: str  # abstract | introduction | methodology | results | discussion
-    author_notes: str | None = None
-
-# ==================================================
-# SECTION STYLE MAP
-# ==================================================
-SECTION_STYLE = {
-    "abstract": "Write concisely and densely with objective academic tone.",
-    "introduction": "Use contextual, motivational academic tone.",
-    "methodology": "Be procedural, precise, and restrained.",
-    "results": "Be cautious, numerical, and observational.",
-    "discussion": "Be interpretive, reflective, and analytical."
+    text: str = Field(..., min_length=1, max_length=3000)
+    section: Literal[
+        "abstract",
+        "introduction",
+        "methodology",
+        "results",
+        "discussion"
+    ]
+    author_notes: Optional[str] = None
+
+# ==================== SECTION-AWARE STYLE ====================
+SECTION_GUIDANCE = {
+    "abstract":
+        "Write concisely and densely. Maintain objective academic tone.",
+    "introduction":
+        "Provide context and motivation. Sound like a researcher framing a problem.",
+    "methodology":
+        "Be procedural, precise, and restrained. No persuasive language.",
+    "results":
+        "Be cautious, observational, and factual. Avoid strong claims.",
+    "discussion":
+        "Be interpretive and reflective. Explain implications carefully."
 }
 
-# ==================================================
-# PROMPT BUILDER
-# ==================================================
-def build_prompt(text: str, section: str, author_notes: str | None):
-    style = SECTION_STYLE.get(section.lower(), "Use formal academic tone.")
+# ==================== OUTPUT CLEANER ====================
+def clean_output(text: str) -> str:
+    text = re.sub(r'<\|.*?\|>', '', text)
+    text = re.sub(r'\s+', ' ', text)
+    return text.strip()
+
+# ==================== FALLBACK (UNCHANGED, SAFE) ====================
+def fallback_humanize(text: str) -> str:
+    replacements = [
+        ("utilize", "use"),
+        ("commence", "start"),
+        ("approximately", "about"),
+        ("therefore", "so"),
+        ("however", "but"),
+        ("in order to", "to"),
+        ("due to the fact that", "because"),
+        ("prior to", "before"),
+        ("subsequent to", "after"),
+    ]
+
+    result = text
+    for formal, simple in replacements:
+        result = re.sub(formal, simple, result, flags=re.IGNORECASE)
+
+    return result
+
+# ==================== PROMPT BUILDER ====================
+def build_prompt(text: str, section: str, author_notes: Optional[str]) -> str:
+    guidance = SECTION_GUIDANCE.get(section, "Use formal academic tone.")
 
     notes_block = (
-        f"\nAuthor context (do not invent reasoning):\n{author_notes}\n"
+        f"\nAuthor context (do NOT invent new reasoning):\n{author_notes}\n"
         if author_notes else ""
     )
 
@@ -116,62 +103,73 @@ NON-NEGOTIABLE RULES:
 - Do NOT add new information
 - Do NOT remove uncertainty
 - Do NOT invent justifications
+- Do NOT change terminology
 - Do NOT introduce grammar or punctuation errors
 
-STYLE GUIDANCE:
-{style}
+SECTION GUIDANCE:
+{guidance}
 
 HUMANIZATION RULES:
 - Vary sentence rhythm naturally (short / medium / long)
 - Reorder clauses where appropriate
-- Avoid overused academic fillers (e.g., "Moreover", "Furthermore", "It is important to note")
-- Prefer implicit transitions over explicit signposting
-- Preserve existing author reasoning and constraints
-- Use controlled lexical variation without replacing technical terms
+- Reduce overused academic fillers (e.g., "Moreover", "Furthermore")
+- Prefer implicit transitions
+- Preserve author intent and constraints
 
 {notes_block}
 
-FINAL CHECK:
-Ensure the text sounds like a researcher explaining their own work.
-
-TEXT TO HUMANIZE:
+TEXT:
 {text}
 
 OUTPUT:
-Return ONLY the humanized text.
+Return ONLY the revised text.
 """.strip()
 
-# ==================================================
-# HUMANIZE ENDPOINT
-# ==================================================
-@app.post("/humanize")
-def humanize(req: HumanizeRequest):
-    if llm is None:
-        raise HTTPException(
-            status_code=503,
-            detail="Model is still loading. Please try again in a few seconds."
-        )
-
-    if not req.text.strip():
-        raise HTTPException(status_code=400, detail="Input text is empty")
+# ==================== ENDPOINT ====================
+@app.post("/api/humanize")
+async def humanize(req: HumanizeRequest):
+    text = req.text.strip()
 
-    prompt = build_prompt(req.text, req.section, req.author_notes)
+    prompt = build_prompt(text, req.section, req.author_notes)
 
     try:
-        response = llm.create_completion(
-            prompt=prompt,
+        output = llm(
+            prompt,
             max_tokens=400,
-            temperature=0.4,
+            temperature=0.4,   # controlled, academic-safe
             top_p=0.9,
-            repeat_penalty=1.1,
+            top_k=40,
+            stop=["<|user|>", "<|end|>"],
+            echo=False,
         )
 
+        raw = output["choices"][0]["text"]
+        cleaned = clean_output(raw)
+
+        if not cleaned or cleaned.lower() == text.lower():
+            cleaned = fallback_humanize(text)
+
+        return {
+            "original": text,
+            "section": req.section,
+            "humanized": cleaned,
+            "success": True
+        }
+
     except Exception as e:
-        raise HTTPException(
-            status_code=500,
-            detail=f"Inference error: {str(e)}"
-        )
+        print("❌ Inference error:", e)
+        return {
+            "original": text,
+            "section": req.section,
+            "humanized": fallback_humanize(text),
+            "success": False
+        }
 
+# ==================== HEALTH ====================
+@app.get("/")
+def health():
     return {
-        "humanized_text": response["choices"][0]["text"].strip()
+        "status": "ok",
+        "model": MODEL_FILE,
+        "endpoint": "/api/humanize"
     }