Spaces:
Sleeping
Sleeping
| \ | |
| import os, re, json | |
| from functools import lru_cache | |
| import gradio as gr | |
| import torch | |
| os.environ.setdefault("HF_HOME", "/data/.cache/huggingface") | |
| os.environ.setdefault("HF_HUB_CACHE", "/data/.cache/huggingface/hub") | |
| os.environ.setdefault("GRADIO_TEMP_DIR", "/data/gradio") | |
| os.environ.setdefault("GRADIO_CACHE_DIR", "/data/gradio") | |
| for p in ["/data/.cache/huggingface/hub", "/data/gradio"]: | |
| try: os.makedirs(p, exist_ok=True) | |
| except Exception: pass | |
| try: | |
| from zoneinfo import ZoneInfo | |
| except Exception: | |
| ZoneInfo = None | |
| try: | |
| import cohere | |
| _HAS_COHERE = True | |
| except Exception: | |
| _HAS_COHERE = False | |
| from transformers import AutoTokenizer, AutoModelForCausalLM | |
| from huggingface_hub import login | |
| from safety import safety_filter, refusal_reply | |
| from retriever import init_retriever, retrieve_context | |
| from decision_math import compute_operational_numbers | |
| from prompt_templates import build_system_preamble | |
| from upload_ingest import extract_text_from_files | |
| from session_rag import SessionRAG | |
| from mdsi_analysis import capacity_projection, cost_estimate, outcomes_summary | |
| 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) | |
| 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 | |
| def _iter_user_assistant(history): | |
| for item in (history or []): | |
| if isinstance(item, (list, tuple)): | |
| u = item[0] if len(item) > 0 else "" | |
| a = item[1] if len(item) > 1 else "" | |
| yield u, a | |
| def _history_to_prompt(message, history): | |
| parts = [] | |
| for u, a in _iter_user_assistant(history): | |
| if u: parts.append(f"User: {u}") | |
| if a: parts.append(f"Assistant: {a}") | |
| parts.append(f"User: {message}") | |
| parts.append("Assistant:") | |
| return "\n".join(parts) | |
| _co_client = None | |
| if USE_HOSTED_COHERE: | |
| _co_client = cohere.Client(api_key=COHERE_API_KEY) | |
| def cohere_chat(message, history): | |
| try: | |
| prompt = _history_to_prompt(message, history) | |
| resp = _co_client.chat( | |
| model="command-r7b-12-2024", | |
| message=prompt, | |
| temperature=0.3, | |
| max_tokens=700, | |
| ) | |
| if hasattr(resp, "text") and resp.text: return resp.text.strip() | |
| if hasattr(resp, "reply") and resp.reply: return resp.reply.strip() | |
| if hasattr(resp, "generations") and resp.generations: return resp.generations[0].text.strip() | |
| return "Sorry, I couldn't parse the response from Cohere." | |
| except Exception as e: | |
| return f"Error calling Cohere API: {e}" | |
| 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=8192, 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 _iter_user_assistant(history): | |
| if u: msgs.append({"role": "user", "content": u}) | |
| if a: 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=900): | |
| 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() | |
| def _load_snapshot(path="snapshots/current.json"): | |
| try: | |
| with open(path, "r", encoding="utf-8") as f: | |
| return json.load(f) | |
| except Exception: | |
| return { | |
| "timestamp": None, "beds_total": 400, "staffed_ratio": 1.0, "occupied_pct": 0.97, | |
| "ed_census": 62, "ed_admits_waiting": 19, "avg_ed_wait_hours": 8, | |
| "discharge_ready_today": 11, "discharge_barriers": {"allied_health": 7, "placement": 4}, | |
| "rn_shortfall": {"med_ward_A": 1, "med_ward_B": 1}, | |
| "forecast_admits_next_24h": {"respiratory": 14, "other": 9}, | |
| "isolation_needs_waiting": {"contact": 3, "airborne": 1}, "telemetry_needed_waiting": 5 | |
| } | |
| # Init retriever & session RAG | |
| init_retriever() | |
| _session_rag = SessionRAG() | |
| def _mdsi_block() -> str: | |
| base_capacity = capacity_projection(18, 48, 6) | |
| cons_capacity = capacity_projection(12, 48, 6) | |
| opt_capacity = capacity_projection(24, 48, 6) | |
| cost_1200 = cost_estimate(1200, 74.0, 75000.0) | |
| outcomes = outcomes_summary() | |
| return json.dumps({ | |
| "capacity_projection": { | |
| "conservative": cons_capacity, "base": base_capacity, "optimistic": opt_capacity | |
| }, | |
| "cost_for_1200": cost_1200, | |
| "outcomes_summary": outcomes | |
| }, indent=2) | |
| def chat_fn(message, history, user_tz, uploaded_files, scenario_text): | |
| try: | |
| safe_in, blocked_in, reason_in = safety_filter(message, mode="input") | |
| if blocked_in: return refusal_reply(reason_in) | |
| if is_identity_query(safe_in, history): | |
| return "I am ClarityOps, your strategic decision making AI partner." | |
| # Ingest uploads | |
| filepaths = [f.name if hasattr(f, "name") else f for f in (uploaded_files or [])] | |
| if filepaths: | |
| items = extract_text_from_files(filepaths) | |
| if items: _session_rag.add_docs(items) | |
| # Retrieve snippets from session uploads | |
| session_snips = "\\n---\\n".join(_session_rag.retrieve( | |
| "diabetes screening Indigenous Métis mobile program cost throughput outcomes logistics", k=6 | |
| )) | |
| snapshot = _load_snapshot() | |
| policy_context = retrieve_context( | |
| "mobile diabetes screening Indigenous community outreach logistics referral pathways privacy cultural safety data governance cost effectiveness outcomes" | |
| ) | |
| computed = compute_operational_numbers(snapshot) | |
| mdsi_extra = _mdsi_block() if ("diabetes" in (scenario_text or "").lower() or "mdsi" in (scenario_text or "").lower()) else "" | |
| system_preamble = build_system_preamble( | |
| snapshot=snapshot, | |
| policy_context=policy_context, | |
| computed_numbers=computed, | |
| scenario_text=(scenario_text or "" ) + (f"\\n\\nExecutive Pre-Computed Blocks:\\n{mdsi_extra}" if mdsi_extra else ""), | |
| session_snips=session_snips | |
| ) | |
| augmented_user = system_preamble + "\\n\\nUser question or request:\\n" + safe_in | |
| if USE_HOSTED_COHERE: | |
| out = cohere_chat(augmented_user, history) | |
| else: | |
| model, tokenizer = load_local_model() | |
| inputs = build_inputs(tokenizer, augmented_user, history) | |
| out = local_generate(model, tokenizer, inputs, max_new_tokens=900) | |
| if isinstance(out, str): | |
| for tag in ("Assistant:", "System:", "User:"): | |
| if out.startswith(tag): out = out[len(tag):].strip() | |
| safe_out, blocked_out, reason_out = safety_filter(out, mode="output") | |
| if blocked_out: return refusal_reply(reason_out) | |
| return safe_out | |
| except Exception as e: | |
| return f"Error: {e}" | |
| theme = gr.themes.Soft(primary_hue="teal", neutral_hue="slate", radius_size=gr.themes.sizes.radius_lg) | |
| custom_css = """ | |
| :root { --brand-bg: #e6f7f8; --brand-accent: #0d9488; --brand-text: #0f172a; --brand-text-light: #ffffff; } | |
| .gradio-container { background: var(--brand-bg); } | |
| h1 { color: var(--brand-text); font-weight: 700; font-size: 28px !important; } | |
| .chatbot header, .chatbot .label, .chatbot .label-wrap, .chatbot .top, .chatbot .header, .chatbot > .wrap > header { display: none !important; } | |
| .message.user, .message.bot { background: var(--brand-accent) !important; color: var(--brand-text-light) !important; border-radius: 12px !important; padding: 8px 12px !important; } | |
| textarea, input, .gr-input { border-radius: 12px !important; } | |
| .examples, .examples .grid { display: flex !important; justify-content: center !important; text-align: center !important; } | |
| """ | |
| 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") | |
| hide_label_sink = gr.HTML(visible=False) | |
| demo.load(fn=lambda: "", inputs=None, outputs=hide_label_sink, js=""" | |
| () => { const sel = ['.chatbot header','.chatbot .label','.chatbot .label-wrap','.chatbot .top','.chatbot .header','.chatbot > .wrap > header']; | |
| sel.forEach(s => document.querySelectorAll(s).forEach(el => el.style.display = 'none')); return ""; } """) | |
| gr.Markdown("# ClarityOps Augmented Decision AI") | |
| uploads = gr.Files(label="Upload docs/images (PDF, DOCX, CSV, PNG, JPG)", file_types=["file"], file_count="multiple") | |
| scenario = gr.Textbox(label="Scenario Context (paste case studies or executive briefs here)", | |
| lines=10, placeholder="Paste scenario text...") | |
| gr.ChatInterface( | |
| fn=chat_fn, | |
| type="messages", | |
| additional_inputs=[tz_box, uploads, scenario], | |
| chatbot=gr.Chatbot(label="", show_label=False, type="messages", height=700), | |
| examples=[ | |
| ["What are the symptoms of hypertension?"], | |
| ["What are common drug interactions with aspirin?"], | |
| ["What are the warning signs of diabetes?"], | |
| ], | |
| cache_examples=False, | |
| ) | |
| if __name__ == "__main__": | |
| port = int(os.environ.get("PORT", "7860")) | |
| demo.launch(server_name="0.0.0.0", server_port=port, show_api=False, max_threads=8) | |