app.py

import os
import re
from datetime import datetime, timezone
from functools import lru_cache

import gradio as gr
import torch

# Timezone (Python 3.9+)
try:
    from zoneinfo import ZoneInfo
except Exception:
    ZoneInfo = None

# Cohere SDK
try:
    import cohere
    _HAS_COHERE = True
except Exception:
    _HAS_COHERE = False

from transformers import AutoTokenizer, AutoModelForCausalLM
from huggingface_hub import login, HfApi

# -------------------
# Config
# -------------------
MODEL_ID = os.getenv("MODEL_ID", "CohereLabs/c4ai-command-r7b-12-2024")
HF_TOKEN = os.getenv("HUGGINGFACE_HUB_TOKEN") or os.getenv("HF_TOKEN")
COHERE_API_KEY = os.getenv("COHERE_API_KEY")
USE_HOSTED_COHERE = bool(COHERE_API_KEY and _HAS_COHERE)

# -------------------
# Helpers
# -------------------
def pick_dtype_and_map():
    if torch.cuda.is_available():
        return torch.float16, "auto"
    if torch.backends.mps.is_available():
        return torch.float16, {"": "mps"}
    return torch.float32, "cpu"

def is_identity_query(message, history):
    patterns = [
        r"\bwho\s+are\s+you\b", r"\bwhat\s+are\s+you\b",
        r"\bwhat\s+is\s+your\s+name\b", r"\bwho\s+is\s+this\b",
        r"\bidentify\s+yourself\b", r"\btell\s+me\s+about\s+yourself\b",
        r"\bdescribe\s+yourself\b", r"\band\s+you\s*\?\b",
        r"\byour\s+name\b", r"\bwho\s+am\s+i\s+chatting\s+with\b"
    ]
    def match(t):
        return any(re.search(p, (t or "").strip().lower()) for p in patterns)
    if match(message):
        return True
    if history:
        last_user = history[-1][0] if isinstance(history[-1], (list, tuple)) else None
        if match(last_user):
            return True
    return False

# -------------------
# Cohere Hosted
# -------------------
_co_client = None
if USE_HOSTED_COHERE:
    _co_client = cohere.Client(api_key=COHERE_API_KEY)

def _cohere_parse(resp):
    if hasattr(resp, "output_text") and resp.output_text:
        return resp.output_text.strip()
    if getattr(resp, "message", None) and getattr(resp.message, "content", None):
        for p in resp.message.content:
            if hasattr(p, "text") and p.text:
                return p.text.strip()
    if hasattr(resp, "text") and resp.text:
        return resp.text.strip()
    return "Sorry, I couldn't parse the response from Cohere."

def cohere_chat(message, history):
    try:
        msgs = []
        for u, a in (history or []):
            msgs.append({"role": "user", "content": u})
            msgs.append({"role": "assistant", "content": a})
        msgs.append({"role": "user", "content": message})
        resp = _co_client.responses.create(
            model="command-r7b-12-2024",
            messages=msgs,
            temperature=0.3,
            max_tokens=350,
        )
        return _cohere_parse(resp)
    except Exception as e:
        return f"Error calling Cohere API: {e}"

# -------------------
# Local HF Model
# -------------------
@lru_cache(maxsize=1)
def load_local_model():
    if not HF_TOKEN:
        raise RuntimeError(
            "HUGGINGFACE_HUB_TOKEN is not set."
        )
    login(token=HF_TOKEN, add_to_git_credential=False)
    dtype, device_map = pick_dtype_and_map()
    tok = AutoTokenizer.from_pretrained(
        MODEL_ID,
        token=HF_TOKEN,
        use_fast=True,
        model_max_length=4096,
        padding_side="left",
        trust_remote_code=True,
    )
    mdl = AutoModelForCausalLM.from_pretrained(
        MODEL_ID,
        token=HF_TOKEN,
        device_map=device_map,
        low_cpu_mem_usage=True,
        torch_dtype=dtype,
        trust_remote_code=True,
    )
    if mdl.config.eos_token_id is None and tok.eos_token_id is not None:
        mdl.config.eos_token_id = tok.eos_token_id
    return mdl, tok

def build_inputs(tokenizer, message, history):
    msgs = []
    for u, a in (history or []):
        msgs.append({"role": "user", "content": u})
        msgs.append({"role": "assistant", "content": a})
    msgs.append({"role": "user", "content": message})
    return tokenizer.apply_chat_template(
        msgs, tokenize=True, add_generation_prompt=True, return_tensors="pt"
    )

def local_generate(model, tokenizer, input_ids, max_new_tokens=350):
    input_ids = input_ids.to(model.device)
    with torch.no_grad():
        out = model.generate(
            input_ids=input_ids,
            max_new_tokens=max_new_tokens,
            do_sample=True,
            temperature=0.3,
            top_p=0.9,
            repetition_penalty=1.15,
            pad_token_id=tokenizer.eos_token_id,
            eos_token_id=tokenizer.eos_token_id,
        )
    gen_only = out[0, input_ids.shape[-1]:]
    return tokenizer.decode(gen_only, skip_special_tokens=True).strip()

# -------------------
# Chat Function
# -------------------
def chat_fn(message, history, user_tz):
    try:
        if is_identity_query(message, history):
            return "I am ClarityOps, your strategic decision making AI partner."
        if USE_HOSTED_COHERE:
            return cohere_chat(message, history)
        model, tokenizer = load_local_model()
        inputs = build_inputs(tokenizer, message, history)
        return local_generate(model, tokenizer, inputs, max_new_tokens=350)
    except Exception as e:
        return f"Error: {e}"

# -------------------
# Theme & CSS
# -------------------
theme = gr.themes.Soft(
    primary_hue="teal",
    neutral_hue="slate",
    radius_size=gr.themes.sizes.radius_lg,
)

custom_css = """
:root {
  --brand-bg: #e6f7f8; /* soft medical teal */
  --brand-accent: #0d9488; /* teal-600 */
  --brand-text-light: #ffffff;
}

.gradio-container {
  background: var(--brand-bg);
}

h1 {
  color: #0f172a;
  font-weight: 700;
  font-size: 28px !important;
}

/* Both bot and user bubbles teal with white text */
.message.user, .message.bot {
  background: var(--brand-accent) !important;
  color: var(--brand-text-light) !important;
}
"""

# -------------------
# UI
# -------------------
with gr.Blocks(theme=theme, css=custom_css) as demo:
    tz_box = gr.Textbox(visible=False)
    demo.load(lambda tz: tz, inputs=[tz_box], outputs=[tz_box],
              js="() => Intl.DateTimeFormat().resolvedOptions().timeZone")

    gr.Markdown("# Medical Decision Support AI")

    gr.ChatInterface(
        fn=chat_fn,
        type="messages",
        additional_inputs=[tz_box],
        examples=[
            ["What are the symptoms of hypertension?", ""],
            ["What are common drug interactions with aspirin?", ""],
            ["What are the warning signs of diabetes?", ""],
        ],
        cache_examples=True,
    )

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