import os
import pandas as pd
import gradio as gr
from huggingface_hub import hf_hub_download
from llama_cpp import Llama                         # GGUF inference on CPU

# ---------- model loading (done once at startup) ----------
MODEL_REPO  = "TheBloke/phi-2-GGUF"                 # fully open 2.7 B model
MODEL_FILE  = "phi-2.Q4_K_M.gguf"                  # 4‑bit, 3.5 GB RAM
CTX_SIZE    = 2048                                  # ample for prompt+answer

model_path = hf_hub_download(repo_id=MODEL_REPO, filename=MODEL_FILE)
llm = Llama(model_path=model_path,
            n_ctx=CTX_SIZE,
            n_threads=os.cpu_count() or 2)          # use all CPUs

# ---------- analysis + generation ----------
def analyze_ads(file):
    df = pd.read_csv(file.name)

    req = {"headline","description","impressions","CTR","form_opens","spend"}
    if not req.issubset(df.columns):
        return f"Missing columns: {', '.join(req - set(df.columns))}"

    # numeric conversions
    for col in ["impressions","CTR","form_opens","spend"]:
        df[col] = pd.to_numeric(df[col], errors="coerce")
    df = df.dropna()

    df["engagement_rate"]     = df["form_opens"] / df["impressions"]
    df["CPC"]                 = df["spend"] / (df["CTR"] * df["impressions"]).replace(0, pd.NA)
    df["cost_per_form_open"]  = df["spend"] / df["form_opens"].replace(0, pd.NA)

    top  = df.sort_values("CTR", ascending=False).head(3)
    worst = df.sort_values("CTR").head(3)

    def rows_to_text(sub):
        out = ""
        for _, r in sub.iterrows():
            out += (f"Headline: {r.headline}\n"
                    f"Description: {r.description}\n"
                    f"Imp: {int(r.impressions)}, CTR: {r.CTR:.3f}, "
                    f"Form Opens: {int(r.form_opens)}, ER: {r.engagement_rate:.3f}\n"
                    f"Spend: ${r.spend:.2f}, CPC: ${r.CPC:.2f}, CPF: ${r.cost_per_form_open:.2f}\n\n")
        return out

    prompt = (
        "You are a senior digital marketer.\n"
        "Analyse the high‑ and low‑performing ads below and deliver:\n"
        "1. Key patterns of winners.\n"
        "2. Weak points of losers.\n"
        "3. Three actionable creative improvements.\n\n"
        f"--- HIGH CTR ADS ---\n{rows_to_text(top)}"
        f"--- LOW CTR ADS ---\n{rows_to_text(worst)}"
    )

    # generate (stream=False -> returns dict)
    answer = llm(prompt, max_tokens=320, temperature=0.7, top_p=0.9)["choices"][0]["text"]
    return answer.strip()

# ---------- Gradio UI ----------
demo = gr.Interface(
    fn=analyze_ads,
    inputs=gr.File(label="CSV with: headline, description, impressions, CTR, form_opens, spend"),
    outputs=gr.Textbox(label="AI‑generated analysis & recommendations"),
    title="Ad Performance Analyzer (Phi‑2 4‑bit, CPU‑only)",
    description="Upload your ad data and get actionable insights without paid APIs."
)

if __name__ == "__main__":
    demo.launch()