Update app.py

app.py

@@ -2,23 +2,31 @@ from fastapi import FastAPI, HTTPException
 from fastapi.middleware.cors import CORSMiddleware
 from pydantic import BaseModel
 from llama_cpp import Llama
-from huggingface_hub import hf_hub_download  # just to ensure import works
+from huggingface_hub import hf_hub_download  # keep to satisfy HF Spaces
 
-# GGUF MODEL
+# -----------------------------
+# MODEL CONFIG
+# -----------------------------
 MODEL_REPO = "bartowski/Phi-3.1-mini-4k-instruct-GGUF"
+
+# Use a lighter, faster quant. You can keep Q4_K_M if you want,
+# but IQ2_M will usually be faster on CPU.
 MODEL_FILE = "Phi-3.1-mini-4k-instruct-IQ2_M.gguf"
 
 print("Loading Phi-3.1 Mini GGUF model...")
 llm = Llama.from_pretrained(
     repo_id=MODEL_REPO,
     filename=MODEL_FILE,
-    n_threads=4,          # increase if you have more CPU cores
-    n_ctx=2048,           # 2048 is enough for humanization, smaller = faster
-    n_batch=256,          # larger batch = faster generation on CPU
+    n_threads=4,      # match your CPU cores
+    n_ctx=1024,       # enough for a few paragraphs, faster than 2048
+    n_batch=512,      # big speed-up
     verbose=False,
 )
 print("Model loaded successfully.")
 
+# -----------------------------
+# FASTAPI APP
+# -----------------------------
 app = FastAPI()
 
 app.add_middleware(
@@ -32,18 +40,21 @@ class RequestBody(BaseModel):
     text: str
 
 
-# -------------------------------
-# STRONG SINGLE-PASS HUMANIZE PROMPT
-# -------------------------------
-SYSTEM_PROMPT = """You are a rewriting engine, not a question-answering assistant.
+# Strong paraphrasing / humanizing prompt
+SYSTEM_PROMPT = """
+You are a professional paraphrasing engine.
 
-Your job:
-- Take the user's text and rewrite it in more natural, human-like English.
-- Keep the same intent, meaning, and sentence type (a question stays a question).
-- Do NOT answer the question, ask for information, or add new content.
+Rewrite the given text so that:
+- The wording and sentence structure are clearly different from the original.
+- The meaning, facts, and tone stay exactly the same.
+- You avoid copying long phrases from the input. Do not keep more than 5–6 consecutive words the same.
+- Vary sentence lengths and break or merge sentences when natural.
+- Do NOT answer questions or add new information.
 - Do NOT explain what you are doing.
-- Output only one rewritten version of the text and nothing else."""
+Output ONLY the rewritten text and nothing else.
+"""
 
+MAX_INPUT_CHARS = 3000  # safety limit so long essays don't kill the CPU
 
 
 @app.post("/api/humanize")
@@ -52,23 +63,33 @@ async def humanize(body: RequestBody):
     if not text:
         raise HTTPException(status_code=400, detail="Text cannot be empty")
 
-    # Phi-3.1-style instruction formatting
+    # truncate extremely long inputs for speed/stability
+    if len(text) > MAX_INPUT_CHARS:
+        text = text[:MAX_INPUT_CHARS]
+
+    # Simple instruct-style prompt (NO chat tags)
     prompt = (
-        f"<s>[SYSTEM]\n{SYSTEM_PROMPT}\n[/SYSTEM]\n"
-        f"[USER]\n{text}\n[/USER]\n"
-        "[ASSISTANT]\n"
+        SYSTEM_PROMPT
+        + "\n\nOriginal text:\n"
+        + text
+        + "\n\nRewritten text:\n"
     )
 
     try:
         output = llm(
             prompt,
-            max_tokens=260,   # ⬅ bump this up from 160
-            temperature=0.3,  # if you already set deterministic mode
-            top_p=1.0,
-            repeat_penalty=1.0,
-            stop=["</s>", "[USER]", "[SYSTEM]", "[ASSISTANT]", "[/ASSISTANT]"],
+            max_tokens=260,      # enough to fully rewrite a long paragraph
+            temperature=0.6,     # some variation → more human-like
+            top_p=0.9,
+            repeat_penalty=1.05,
+            # no chat stop tokens; model will stop on its own
         )
         result = output["choices"][0]["text"].strip()
+
+        if not result:
+            # this prevents your frontend from showing confusing "No output" forever
+            raise HTTPException(status_code=500, detail="Model returned empty output")
+
         return {"result": result}
 
     except Exception as e:
@@ -78,4 +99,4 @@ async def humanize(body: RequestBody):
 
 @app.get("/")
 def health():
-    return {"status": "ok"}
+    return {"status": "ok"}