Update app.py

app.py

@@ -50,6 +50,43 @@ USE_HOSTED_COHERE = bool(COHERE_API_KEY and _HAS_COHERE)
 
 MAX_NEW_TOKENS = int(os.getenv("MAX_NEW_TOKENS", "512"))
 
+# ---------- System Master (two-phase, LLM-only behavior) ----------
+SYSTEM_MASTER = """
+SYSTEM ROLE (fixed, always on)
+You are ClarityOps, a medical analytics system that interacts only via this chat.
+
+Absolute rules:
+- Use ONLY information provided in this conversation (scenario text + uploaded files).
+- Never invent data. If something required is missing after clarifications, output the literal token: INSUFFICIENT_DATA.
+- Always run in TWO PHASES:
+  Phase 1: Ask up to 5 concise clarification questions, grouped by category (Prioritization, Capacity, Cost, Clinical, Recommendations). Then STOP and WAIT.
+  Phase 2: After answers are provided, produce the final structured analysis exactly in the required format.
+
+Core behavior:
+- Read and synthesize any user-uploaded files (e.g., CSV/XLSX/PDF) relevant to the scenario.
+- Prefer analytics/longitudinal recommendations (risk targeting, follow-up, clustering) over generic ops advice.
+- Show all calculations explicitly for capacity and costs (e.g., “6 teams × 8 clients/day × 60 days = 2,880”).
+- Use correct clinical units and plausible ranges.
+- Include a brief “Provenance” section mapping each key output to scenario text, files, and/or clarified answers.
+
+Medical guardrails (always apply):
+- Units: BP in mmHg, A1c in %, BMI in kg/m², Total Cholesterol in mmol/L (or as provided), Percentages in %.
+- Plausible ranges: A1c 3–20 %, SBP 60–250 mmHg, DBP 30–150 mmHg, BMI 10–70 kg/m², Total Chol 2–12 mmol/L.
+- Privacy: avoid PHI; aggregate only; apply small-cell suppression where cohort < 10 (describe at a higher level).
+- When data includes mixed or ambiguous indicators, ask to confirm preferred indicators (e.g., obesity/metabolic syndrome vs self-reported diabetes).
+
+Formatting hard rules:
+- Phase 1 output MUST include the header line: “Clarification Questions”
+- Phase 2 output MUST include the header line: “Structured Analysis”
+- Phase 2 MUST follow this exact section order:
+  1. Prioritization
+  2. Capacity
+  3. Cost
+  4. Clinical Benefits
+  5. ClarityOps Top 3 Recommendations
+  (Include a short Provenance block at the end.)
+""".strip()
+
 # ---------- Helpers ----------
 def pick_dtype_and_map():
     if torch.cuda.is_available():
@@ -86,7 +123,12 @@ def _iter_user_assistant(history):
             yield u, a
 
 def _history_to_prompt(message, history):
+    """
+    Build a simple chat-style prompt INCLUDING the System Master preamble.
+    """
     parts = []
+    # system master always first
+    parts.append(f"System: {SYSTEM_MASTER}")
     for u, a in _iter_user_assistant(history):
         if u: parts.append(f"User: {u}")
         if a: parts.append(f"Assistant: {a}")
@@ -151,6 +193,8 @@ def load_local_model():
 
 def build_inputs(tokenizer, message, history):
     msgs = []
+    # Always inject system master into the chat template, if supported
+    msgs.append({"role": "system", "content": SYSTEM_MASTER})
     for u, a in _iter_user_assistant(history):
         if u: msgs.append({"role": "user", "content": u})
         if a: msgs.append({"role": "assistant", "content": a})
@@ -204,8 +248,13 @@ def _mdsi_block():
         "outcomes_summary": outcomes
     }, indent=2)
 
-# ---------- Core chat logic (Cohere-first with fallback) ----------
-def clarityops_reply(user_msg, history, tz, uploaded_files_paths):
+# ---------- Core chat logic with two-phase behavior ----------
+def clarityops_reply(user_msg, history, tz, uploaded_files_paths, awaiting_answers=False):
+    """
+    awaiting_answers:
+      - False: Phase 1 mode -> generate clarification questions and WAIT
+      - True: Phase 2 mode  -> consume clarifications and produce structured analysis
+    """
     try:
         log_event("user_message", None, {"sizes": {"chars": len(user_msg or "")}})
 
@@ -213,12 +262,12 @@ def clarityops_reply(user_msg, history, tz, uploaded_files_paths):
         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)]
+            return history + [(user_msg, ans)], awaiting_answers
 
         # 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)]
+            return history + [(user_msg, ans)], awaiting_answers
 
         # Debug slash command: /diag
         if (safe_in or "").strip().lower().startswith("/diag"):
@@ -232,14 +281,14 @@ def clarityops_reply(user_msg, history, tz, uploaded_files_paths):
                     "Sample retrieved snippets:",
                     *(sample or ["<no snippets>"])
                 ]
-                return history + [(user_msg, "\n\n".join(msg))]
+                return history + [(user_msg, "\n\n".join(msg))], awaiting_answers
             except Exception as e:
-                return history + [(user_msg, f"Diag error: {e}")]
+                return history + [(user_msg, f"Diag error: {e}")], awaiting_answers
 
         # Ingest uploads: returns chunks + artifacts
         if uploaded_files_paths:
             ing = extract_text_from_files(uploaded_files_paths)
-            chunks = ing.get("chunks", []) if isinstance(ing, dict) else (inf or [])
+            chunks = ing.get("chunks", []) if isinstance(ing, dict) else (ing or [])
             artifacts = ing.get("artifacts", []) if isinstance(ing, dict) else []
             if chunks:
                 _session_rag.add_docs(chunks)
@@ -251,17 +300,7 @@ def clarityops_reply(user_msg, history, tz, uploaded_files_paths):
         if re.search(r"\b(columns?|headers?)\b", (safe_in or "").lower()):
             cols = _session_rag.get_latest_csv_columns()
             if cols:
-                return history + [(user_msg, "Here are the column names from your most recent CSV upload:\n\n- " + "\n- ".join(cols))]
-
-        # Heuristic: scenario mode nudge if a long case study was pasted
-        plain = (safe_in or "").strip().lower()
-        looks_like_case = ("background" in plain and "objective" in plain) or ("case study" in plain)
-        if looks_like_case and len(plain) > 600:
-            safe_in += (
-                "\n\nPlease analyze the scenario above using the Expected Output Format: "
-                "produce structured recommendations, estimates and assumptions, include tables and bullet points, "
-                "and explicitly state how uploaded files (CSV/docs) influenced your estimates."
-            )
+                return history + [(user_msg, "Here are the column names from your most recent CSV upload:\n\n- " + "\n- ".join(cols))], awaiting_answers
 
         # Retrieve from session uploads (text chunks)
         session_snips = "\n---\n".join(_session_rag.retrieve(
@@ -280,7 +319,7 @@ def clarityops_reply(user_msg, history, tz, uploaded_files_paths):
         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 ""
 
-        scenario_block = safe_in if len(safe_in) > 400 else ""
+        scenario_block = safe_in if len((safe_in or "")) > 0 else ""
         system_preamble = build_system_preamble(
             snapshot=snapshot,
             policy_context=policy_context,
@@ -289,7 +328,22 @@ def clarityops_reply(user_msg, history, tz, uploaded_files_paths):
             session_snips=session_snips
         )
 
-        augmented_user = system_preamble + "\n\nUser question or request:\n" + safe_in
+        # Phase-specific instruction appended to the user content
+        if not awaiting_answers:
+            phase_directive = (
+                "\n\n[INSTRUCTION TO MODEL]\n"
+                "Produce **Phase 1** only: output a header 'Clarification Questions' and ask up to 5 concise, grouped questions "
+                "(Prioritization, Capacity, Cost, Clinical, Recommendations). Then STOP and WAIT.\n"
+            )
+        else:
+            phase_directive = (
+                "\n\n[INSTRUCTION TO MODEL]\n"
+                "Produce **Phase 2** only: output a header 'Structured Analysis' and follow the exact section order "
+                "(Prioritization, Capacity, Cost, Clinical Benefits, ClarityOps Top 3 Recommendations). "
+                "Use uploaded files + the user's latest answers as authoritative. Show calculations, units, and a brief Provenance.\n"
+            )
+
+        augmented_user = SYSTEM_MASTER + "\n\n" + system_preamble + "\n\nUser message:\n" + safe_in + phase_directive
 
         # Cohere first
         out = cohere_chat(augmented_user, history)
@@ -312,20 +366,30 @@ def clarityops_reply(user_msg, history, tz, uploaded_files_paths):
         if blocked_out:
             safe_out = refusal_reply(reason_out)
 
+        # Decide next state:
+        # If we just asked clarifications, set awaiting_answers=True.
+        # If we just produced structured analysis, set awaiting_answers=False.
+        new_awaiting = awaiting_answers
+        if not awaiting_answers and "clarification questions" in safe_out.lower():
+            new_awaiting = True
+        elif awaiting_answers and "structured analysis" in safe_out.lower():
+            new_awaiting = False
+
         # Audit (content-free fingerprints)
         log_event("assistant_reply", None, {
             **hash_summary("prompt", augmented_user if not PERSIST_CONTENT else ""),
             **hash_summary("reply", safe_out if not PERSIST_CONTENT else ""),
+            "awaiting_next_phase": new_awaiting
         })
 
-        return history + [(user_msg, safe_out)]
+        return history + [(user_msg, safe_out)], new_awaiting
     except Exception as e:
         err = f"Error: {e}"
         try:
             traceback.print_exc()
         except Exception:
             pass
-        return history + [(user_msg, err)]
+        return history + [(user_msg, err)], awaiting_answers
 
 # ---------- Theme & CSS ----------
 theme = gr.themes.Soft(primary_hue="teal", neutral_hue="slate", radius_size=gr.themes.sizes.radius_lg)
@@ -336,13 +400,36 @@ 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; }
+
+/* Centered handshake overlay */
+#handshake-overlay {
+  position: absolute;
+  z-index: 50;
+  top: 50%;
+  left: 50%;
+  transform: translate(-50%, -50%);
+  background: rgba(13, 148, 136, 0.92);
+  color: #fff;
+  padding: 18px 22px;
+  border-radius: 14px;
+  font-size: 16px;
+  max-width: 720px;
+  text-align: center;
+  box-shadow: 0 10px 24px rgba(0,0,0,0.2);
+}
+#handshake-overlay.hidden { display: none; }
+#chat-container { position: relative; }
 """
 
 # ---------- UI (single window; uploads at bottom) ----------
 with gr.Blocks(theme=theme, css=custom_css, analytics_enabled=False) as demo:
     gr.Markdown("# ClarityOps Augmented Decision AI")
 
-    chat = gr.Chatbot(label="", show_label=False, height=700)
+    with gr.Column(elem_id="chat-container"):
+        chat = gr.Chatbot(label="", show_label=False, height=700)
+        handshake = gr.HTML(
+            value='<div id="handshake-overlay">ClarityOps loaded. Paste your scenario and attach files. I’ll ask up to 5 clarifications, then produce the structured analysis</div>'
+        )
 
     with gr.Row():
         uploads = gr.Files(
@@ -354,7 +441,7 @@ with gr.Blocks(theme=theme, css=custom_css, analytics_enabled=False) as demo:
         msg = gr.Textbox(
             label="",
             show_label=False,
-            placeholder="Type a message… (paste scenarios here too; ClarityOps will adapt)",
+            placeholder="Paste your scenario here (attach files below). ClarityOps will ask clarifications first.",
             scale=10
         )
         send = gr.Button("Send", scale=1)
@@ -362,6 +449,7 @@ with gr.Blocks(theme=theme, css=custom_css, analytics_enabled=False) as demo:
 
     state_history = gr.State(value=[])
     state_uploaded = gr.State(value=[])
+    state_awaiting = gr.State(value=False)  # False = Phase 1 next; True = awaiting answers for Phase 2
 
     def _store_uploads(files, current):
         paths = []
@@ -371,12 +459,14 @@ with gr.Blocks(theme=theme, css=custom_css, analytics_enabled=False) as demo:
 
     uploads.change(fn=_store_uploads, inputs=[uploads, state_uploaded], outputs=state_uploaded)
 
-    def _on_send(user_msg, history, up_paths):
+    def _on_send(user_msg, history, up_paths, awaiting):
+        # Hide handshake on first interaction by returning a class change
+        hide_overlay_js = gr.update(value='<div id="handshake-overlay" class="hidden"></div>')
         try:
             if not user_msg or not user_msg.strip():
-                return history, "", history
-            new_history = clarityops_reply(user_msg.strip(), history or [], None, up_paths or [])
-            return new_history, "", new_history
+                return history, "", history, awaiting, hide_overlay_js
+            new_history, new_awaiting = clarityops_reply(user_msg.strip(), history or [], None, up_paths or [], awaiting_answers=awaiting)
+            return new_history, "", new_history, new_awaiting, hide_overlay_js
         except Exception as e:
             err = f"Error: {e}"
             try:
@@ -384,23 +474,24 @@ with gr.Blocks(theme=theme, css=custom_css, analytics_enabled=False) as demo:
             except Exception:
                 pass
             new_hist = (history or []) + [(user_msg or "", err)]
-            return new_hist, "", new_hist
+            return new_hist, "", new_hist, awaiting, hide_overlay_js
 
-    send.click(_on_send, inputs=[msg, state_history, state_uploaded],
-               outputs=[chat, msg, state_history],
+    send.click(_on_send, inputs=[msg, state_history, state_uploaded, state_awaiting],
+               outputs=[chat, msg, state_history, state_awaiting, handshake],
                concurrency_limit=2, queue=True)
 
-    msg.submit(_on_send, inputs=[msg, state_history, state_uploaded],
-               outputs=[chat, msg, state_history],
+    msg.submit(_on_send, inputs=[msg, state_history, state_uploaded, state_awaiting],
+               outputs=[chat, msg, state_history, state_awaiting, handshake],
                concurrency_limit=2, queue=True)
 
-    clear.click(lambda: ([], "", []), None, [chat, msg, state_history])
+    def _on_clear():
+        # Reset everything, show handshake again
+        return [], "", [], False, '<div id="handshake-overlay">ClarityOps loaded. Paste your scenario and attach files. I’ll ask up to 5 clarifications, then produce the structured analysis</div>'
+
+    clear.click(_on_clear, None, [chat, msg, state_history, state_awaiting, handshake])
 
 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)
 
 
-
-
-