Update app.py

app.py

@@ -1,141 +1,122 @@
+import os
 from fastapi import FastAPI, HTTPException
-from fastapi.middleware.cors import CORSMiddleware
-from pydantic import BaseModel, Field
-from typing import Literal
+from pydantic import BaseModel
 from llama_cpp import Llama
-import re
+from contextlib import asynccontextmanager
 
-# ==================== MODEL CONFIG ====================
+# =========================
+# MODEL CONFIG
+# =========================
 MODEL_REPO = "bartowski/Phi-3.1-mini-4k-instruct-GGUF"
 MODEL_FILE = "Phi-3.1-mini-4k-instruct-IQ2_M.gguf"
 
-print("🚀 Loading Phi-3.1 Mini (Fast Humanizer)...")
-llm = Llama.from_pretrained(
-    repo_id=MODEL_REPO,
-    filename=MODEL_FILE,
-    n_threads=4,
-    n_ctx=2048,        # Smaller context = faster
-    n_batch=256,
-    n_gpu_layers=0,
-    verbose=False,
-)
-print("✅ Model loaded")
-
-# ==================== FASTAPI ====================
-app = FastAPI(title="Fast Humanizer")
-
-app.add_middleware(
-    CORSMiddleware,
-    allow_origins=["*"],
-    allow_methods=["*"],
-    allow_headers=["*"],
-)
-
-# ==================== REQUEST ====================
+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
+# =========================
+@asynccontextmanager
+async def lifespan(app: FastAPI):
+    global llm
+    llm = Llama(
+        model_path=MODEL_FILE,
+        n_ctx=N_CTX,
+        n_threads=N_THREADS,
+        n_batch=N_BATCH,
+        verbose=False,
+    )
+    yield
+
+app = FastAPI(title="AI Humanizer", lifespan=lifespan)
+
+# =========================
+# INPUT SCHEMA
+# =========================
 class HumanizeRequest(BaseModel):
-    text: str = Field(..., min_length=1, max_length=500)
-    style: Literal["professional", "casual", "academic", "marketing"] = "professional"
-
-# ==================== STYLE PROMPTS (SHORT & EFFECTIVE) ====================
-STYLE_PROMPTS = {
-    "professional":
-        "Rewrite this text in a clear, polished, professional tone. "
-        "Make it sound natural and confident. Output ONLY the rewritten text:\n",
-
-    "casual":
-        "Rewrite this text in a friendly, casual, human way. "
-        "Use natural phrasing and contractions. Output ONLY the rewritten text:\n",
-
-    "academic":
-        "Rewrite this text in a formal academic style. "
-        "Use clear structure and precise language. Output ONLY the rewritten text:\n",
-
-    "marketing":
-        "Rewrite this text as persuasive marketing copy. "
-        "Make it engaging and benefit-focused. Output ONLY the rewritten text:\n",
+    text: str
+    section: str  # abstract | introduction | methodology | results | discussion
+    author_notes: str | None = None
+
+# =========================
+# SECTION-AWARE STYLE MAP
+# =========================
+SECTION_STYLE = {
+    "abstract": "Write concisely, densely, and objectively. Avoid narrative tone.",
+    "introduction": "Use contextual and motivational academic tone. Explain relevance.",
+    "methodology": "Be procedural, precise, and restrained. No persuasion.",
+    "results": "Be cautious, numerical, and observational. No strong claims.",
+    "discussion": "Be interpretive, reflective, and analytical."
 }
 
-# ==================== OUTPUT CLEANER ====================
-def clean_output(text: str) -> str:
-    text = re.sub(r'<\|.*?\|>', '', text)
-    text = re.sub(r'\s+', ' ', text)
-    text = text.strip()
-
-    lines = [l.strip() for l in text.split("\n") if len(l.strip()) > 10]
-    return lines[-1] if lines else text
-
-# ==================== FALLBACK HUMANIZER ====================
-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, casual in replacements:
-        result = re.sub(formal, casual, result, flags=re.IGNORECASE)
-
-    result = re.sub(r"\b(do not|does not|did not|cannot|will not|is not|are not)\b",
-                    lambda m: m.group(1).replace(" ", "'").replace("cannot", "can't"),
-                    result,
-                    flags=re.IGNORECASE)
-    return result
-
-# ==================== ENDPOINT ====================
-@app.post("/api/humanize")
-async def humanize(req: HumanizeRequest):
-    text = req.text.strip()
-    style = req.style
-
-    prompt = (
-        f"<|user|>\n"
-        f"{STYLE_PROMPTS[style]}"
-        f"{text}\n"
-        f"<|end|>\n"
-        f"<|assistant|>\n"
+# =========================
+# PROMPT BUILDER
+# =========================
+def build_prompt(text: str, section: str, author_notes: str | None):
+    style = SECTION_STYLE.get(section.lower(), "Use formal academic tone.")
+
+    notes_block = (
+        f"\nAuthor context (do not invent new reasoning):\n{author_notes}\n"
+        if author_notes else ""
+    )
+
+    return f"""
+You are an academic writing assistant.
+
+GOAL:
+Restore natural human authorship signals while preserving formal academic language.
+
+NON-NEGOTIABLE RULES:
+- Preserve all technical meaning, claims, numbers, and citations
+- Do NOT add new information
+- Do NOT remove uncertainty
+- Do NOT invent justifications
+- Do NOT introduce grammar or punctuation errors
+
+STYLE GUIDANCE:
+{style}
+
+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 and highlight existing author reasoning or constraints
+- Use controlled lexical variation without replacing technical terms
+
+{notes_block}
+
+FINAL CHECK:
+Ensure the text sounds like a researcher explaining their own work.
+
+TEXT TO HUMANIZE:
+{text}
+
+OUTPUT:
+Return ONLY the humanized text.
+""".strip()
+
+# =========================
+# API ENDPOINT
+# =========================
+@app.post("/humanize")
+def humanize(req: HumanizeRequest):
+    if not req.text.strip():
+        raise HTTPException(status_code=400, detail="Input text is empty")
+
+    prompt = build_prompt(req.text, req.section, req.author_notes)
+
+    response = llm(
+        prompt,
+        max_tokens=512,
+        temperature=0.4,
+        top_p=0.9,
+        repetition_penalty=1.1,
+        stop=["<|end|>"]
     )
 
-    try:
-        output = llm(
-            prompt,
-            max_tokens=180,        # FAST
-            temperature=0.7,
-            top_p=0.9,
-            top_k=40,
-            repeat_penalty=1.1,
-            stop=["<|end|>", "<|user|>"],
-            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,
-            "style": style,
-            "humanized": cleaned,
-            "success": True
-        }
-
-    except Exception as e:
-        print("❌ Model error:", e)
-        return {
-            "original": text,
-            "style": style,
-            "humanized": fallback_humanize(text),
-            "success": False
-        }
-
-@app.get("/")
-def health():
-    return {"status": "ok", "model": MODEL_FILE}
+    output = response["choices"][0]["text"].strip()
+    return {"humanized_text": output}