Update app.py

app.py

@@ -15,8 +15,10 @@ os.environ.setdefault("GRADIO_TEMP_DIR", "/data/gradio")
 os.environ.setdefault("GRADIO_CACHE_DIR", "/data/gradio")
 os.environ.pop("TRANSFORMERS_CACHE", None)
 for p in ["/data/.cache/huggingface/hub", "/data/gradio"]:
-    try: os.makedirs(p, exist_ok=True)
-    except Exception: pass
+    try:
+        os.makedirs(p, exist_ok=True)
+    except Exception:
+        pass
 
 # Optional timezone
 try:
@@ -42,31 +44,49 @@ 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")
+# ---------- Config ----------
+# Local fallback model (lightweight by default). You can override via env.
+MODEL_ID = os.getenv("MODEL_ID", "microsoft/Phi-3-mini-4k-instruct")
 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)
+COHERE_TIMEOUT_SEC = float(os.getenv("COHERE_TIMEOUT_SEC", "30"))
+
+MAX_NEW_TOKENS = int(os.getenv("MAX_NEW_TOKENS", "512"))  # faster defaults; adjust as needed
 
 # ---------- 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"}
+    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"
+        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 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
+        if match(last_user):
+            return True
     return False
 
 def _iter_user_assistant(history):
+    # history is a list of (user, assistant) tuples (Chatbot default format)
     for item in (history or []):
         if isinstance(item, (list, tuple)):
             u = item[0] if len(item) > 0 else ""
@@ -82,28 +102,36 @@ def _history_to_prompt(message, history):
     parts.append("Assistant:")
     return "\n".join(parts)
 
-# ---------- Cohere path ----------
+# ---------- Cohere path (default first; fallback to local on failure) ----------
 _co_client = None
 if USE_HOSTED_COHERE:
-    _co_client = cohere.Client(api_key=COHERE_API_KEY)
+    _co_client = cohere.Client(api_key=COHERE_API_KEY, timeout=COHERE_TIMEOUT_SEC)
 
 def cohere_chat(message, history):
+    """
+    Returns text on success, or None to signal fallback to local model.
+    """
+    if not _co_client:
+        return None
     try:
         prompt = _history_to_prompt(message, history)
         resp = _co_client.chat(
             model="command-r7b-12-2024",
             message=prompt,
             temperature=0.3,
-            max_tokens=900,
+            max_tokens=MAX_NEW_TOKENS,
         )
-        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}"
+        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 None
+    except Exception:
+        return None
 
-# ---------- Local model (with accelerate fallback) ----------
+# ---------- Local model (accelerate-safe fallback) ----------
 @lru_cache(maxsize=1)
 def load_local_model():
     if not HF_TOKEN:
@@ -111,9 +139,10 @@ def load_local_model():
     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,
+        MODEL_ID, token=HF_TOKEN, use_fast=True, model_max_length=8192,
+        padding_side="left", trust_remote_code=True,
     )
-    # Try device_map path (needs accelerate). Fallback to manual .to(device) if it fails.
+    # Try device_map (needs accelerate); fallback to manual .to(device) if it fails.
     try:
         mdl = AutoModelForCausalLM.from_pretrained(
             MODEL_ID, token=HF_TOKEN, device_map=device_map,
@@ -130,19 +159,25 @@ def load_local_model():
     return mdl, tok
 
 def build_inputs(tokenizer, message, history):
+    # Convert tuple history to chat template input for HF models
     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")
+    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):
+def local_generate(model, tokenizer, input_ids, max_new_tokens=MAX_NEW_TOKENS):
     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,
+            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()
@@ -153,6 +188,7 @@ def _load_snapshot(path=SNAPSHOT_PATH):
         with open(path, "r", encoding="utf-8") as f:
             return json.load(f)
     except Exception:
+        # Safe fallback if no snapshot present
         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,
@@ -164,7 +200,7 @@ def _load_snapshot(path=SNAPSHOT_PATH):
 
 # ---------- Init retrieval engines ----------
 init_retriever()
-_session_rag = SessionRAG()  # in-memory only
+_session_rag = SessionRAG()  # in-memory only; lazy-loads embeddings
 
 # ---------- Executive pre-compute (MDSi block) ----------
 def _mdsi_block():
@@ -179,17 +215,25 @@ def _mdsi_block():
         "outcomes_summary": outcomes
     }, indent=2)
 
-# ---------- Core chat logic ----------
+# ---------- Core chat logic (Cohere-first with fallback) ----------
 def clarityops_reply(user_msg, history, tz, uploaded_files_paths):
+    """
+    - user_msg: latest message text
+    - history: list[(user, assistant)]
+    - tz: timezone str (unused but kept for future features)
+    - uploaded_files_paths: list[str] absolute paths of uploaded files
+    """
     try:
         # Audit (content-free)
         log_event("user_message", None, {"sizes": {"chars": len(user_msg or "")}})
 
+        # Safety (input)
         safe_in, blocked_in, reason_in = safety_filter(user_msg, mode="input")
         if blocked_in:
             ans = refusal_reply(reason_in)
             return history + [(user_msg, ans)]
 
+        # Identity short-circuit
         if is_identity_query(safe_in, history):
             ans = "I am ClarityOps, your strategic decision making AI partner."
             return history + [(user_msg, ans)]
@@ -199,24 +243,25 @@ def clarityops_reply(user_msg, history, tz, uploaded_files_paths):
             items = extract_text_from_files(uploaded_files_paths)
             if items:
                 _session_rag.add_docs(items)
-                # Audit upload names & sizes only
                 log_event("uploads_added", None, {"count": len(items)})
 
         # Retrieve from session uploads
         session_snips = "\n---\n".join(_session_rag.retrieve(
-            "diabetes screening Indigenous Métis mobile program cost throughput outcomes logistics bed flow staffing discharge forecast", k=6
+            "diabetes screening Indigenous Métis mobile program cost throughput outcomes logistics bed flow staffing discharge forecast",
+            k=6
         ))
 
+        # Load daily snapshot + policies + computed ops numbers
         snapshot = _load_snapshot()
         policy_context = retrieve_context(
             "mobile diabetes screening Indigenous community outreach logistics referral pathways cultural safety data governance cost effectiveness outcomes bed management discharge acceleration ambulance offload"
         )
         computed = compute_operational_numbers(snapshot)
 
+        # Exec scenario detect (MDSi)
         user_lower = (safe_in or "").lower()
         mdsi_extra = _mdsi_block() if ("diabetes" in user_lower or "mdsi" in user_lower or "mobile screening" in user_lower) else ""
 
-        # Optionally include long scenario text; redact if persisting later (we don't persist by default)
         scenario_block = safe_in if len(safe_in) > 400 else ""
         system_preamble = build_system_preamble(
             snapshot=snapshot,
@@ -228,13 +273,16 @@ def clarityops_reply(user_msg, history, tz, uploaded_files_paths):
 
         augmented_user = system_preamble + "\n\nUser question or request:\n" + safe_in
 
-        # Generate
+        # --- Cohere first ---
+        out = None
         if USE_HOSTED_COHERE:
             out = cohere_chat(augmented_user, history)
-        else:
+
+        # --- Fallback to local HF model if Cohere not set or fails ---
+        if not out:
             model, tokenizer = load_local_model()
             inputs = build_inputs(tokenizer, augmented_user, history)
-            out = local_generate(model, tokenizer, inputs, max_new_tokens=900)
+            out = local_generate(model, tokenizer, inputs, max_new_tokens=MAX_NEW_TOKENS)
 
         # Tidy echoes
         if isinstance(out, str):
@@ -242,6 +290,7 @@ def clarityops_reply(user_msg, history, tz, uploaded_files_paths):
                 if out.startswith(tag):
                     out = out[len(tag):].strip()
 
+        # Safety (output)
         safe_out, blocked_out, reason_out = safety_filter(out, mode="output")
         if blocked_out:
             safe_out = refusal_reply(reason_out)
@@ -261,20 +310,43 @@ theme = gr.themes.Soft(primary_hue="teal", neutral_hue="slate", radius_size=gr.t
 custom_css = """
 :root { --brand-bg: #e6f7f8; --brand-accent: #0d9488; --brand-text: #0f172a; --brand-text-light: #ffffff; }
 .gradio-container { background: var(--brand-bg); }
+
+/* Title */
 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; }
+
+/* Hide default Chatbot label */
+.chatbot header, .chatbot .label, .chatbot .label-wrap, .chatbot .top, .chatbot .header, .chatbot > .wrap > header {
+  display: none !important;
+}
+
+/* Chat bubbles */
+.message.user, .message.bot {
+  background: var(--brand-accent) !important;
+  color: var(--brand-text-light) !important;
+  border-radius: 12px !important;
+  padding: 8px 12px !important;
+}
+
+/* Inputs softer */
 textarea, input, .gr-input { border-radius: 12px !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")
+    demo.load(
+        lambda tz: tz,
+        inputs=[tz_box],
+        outputs=[tz_box],
+        js="() => Intl.DateTimeFormat().resolvedOptions().timeZone",
+    )
+
     gr.Markdown("# ClarityOps Augmented Decision AI")
 
+    # Main chat (tuple-format history)
     chat = gr.Chatbot(label="", show_label=False, height=700)
 
+    # Uploads above the input
     with gr.Row():
         uploads = gr.Files(
             label="Upload docs/images (PDF, DOCX, CSV, PNG, JPG)",
@@ -282,13 +354,20 @@ with gr.Blocks(theme=theme, css=custom_css) as demo:
         )
 
     with gr.Row():
-        msg = gr.Textbox(label="", show_label=False, placeholder="Type a message… (paste scenarios here too; ClarityOps will adapt)", scale=10)
+        msg = gr.Textbox(
+            label="",
+            show_label=False,
+            placeholder="Type a message… (paste scenarios here too; ClarityOps will adapt)",
+            scale=10
+        )
         send = gr.Button("Send", scale=1)
         clear = gr.Button("Clear chat", scale=1)
 
+    # State
     state_history = gr.State(value=[])
     state_uploaded = gr.State(value=[])
 
+    # Store uploaded file paths in state (persist through session)
     def _store_uploads(files, current):
         paths = []
         for f in (files or []):
@@ -297,17 +376,33 @@ with gr.Blocks(theme=theme, css=custom_css) as demo:
 
     uploads.change(fn=_store_uploads, inputs=[uploads, state_uploaded], outputs=state_uploaded)
 
+    # Send / Enter handlers
     def _on_send(user_msg, history, tz, up_paths):
         if not user_msg or not user_msg.strip():
             return history, "", history
         new_history = clarityops_reply(user_msg.strip(), history or [], tz, up_paths or [])
         return new_history, "", new_history
 
-    send.click(fn=_on_send, inputs=[msg, state_history, tz_box, state_uploaded], outputs=[chat, msg, state_history], queue=True)
-    msg.submit(fn=_on_send, inputs=[msg, state_history, tz_box, state_uploaded], outputs=[chat, msg, state_history], queue=True)
+    send.click(
+        fn=_on_send,
+        inputs=[msg, state_history, tz_box, state_uploaded],
+        outputs=[chat, msg, state_history],
+        queue=True,
+    )
+    msg.submit(
+        fn=_on_send,
+        inputs=[msg, state_history, tz_box, state_uploaded],
+        outputs=[chat, msg, state_history],
+        queue=True,
+    )
 
+    # Clear chat (keep uploads)
     clear.click(lambda: ([], "", []), None, [chat, msg, state_history])
 
+# Enable queue to avoid websocket timeouts on first call / heavy loads
+demo = demo.queue(concurrency_count=2, max_size=32)
+
 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)
+