Update app.py

app.py

@@ -61,41 +61,23 @@ USE_HOSTED_COHERE = bool(COHERE_API_KEY and _HAS_COHERE)
 # Larger output (Cohere + HF fallback)
 MAX_NEW_TOKENS = int(os.getenv("MAX_NEW_TOKENS", "2048"))
 
-# ---------- System Master (two-phase, LLM-only behavior) ----------
+# ---------- Fixed System Preamble for Medical Guardrails ----------
 SYSTEM_MASTER = """
 SYSTEM ROLE (fixed, always on)
-You are ClarityOps, a medical analytics system that interacts only via this chat.
+You are ClarityOps, a medical analytics assistant.
 
 Absolute rules:
-- Use ONLY information provided in this conversation (scenario text + uploaded files).
+- Use ONLY information provided in this conversation (user messages + 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”).
+- Show all calculations explicitly when computing capacity and cost.
 - 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 ----------
@@ -132,8 +114,8 @@ def _sanitize_text(s: str) -> str:
         return s
     return re2.sub(r'[\p{C}--[\n\t]]+', '', s)
 
-def _history_to_prompt(message, history):
-    parts = [f"System: {SYSTEM_MASTER}"]
+def _history_to_prompt(message, history, system_text):
+    parts = [f"System: {system_text}"]
     for u, a in _iter_user_assistant(history):
         if u: parts.append(f"User: {u}")
         if a: parts.append(f"Assistant: {a}")
@@ -142,12 +124,12 @@ def _history_to_prompt(message, history):
     return "\n".join(parts)
 
 # ---------- Cohere first ----------
-def cohere_chat(message, history):
+def cohere_chat(message, history, system_text=SYSTEM_MASTER):
     if not USE_HOSTED_COHERE:
         return None
     try:
         client = cohere.Client(api_key=COHERE_API_KEY)
-        prompt = _history_to_prompt(message, history)
+        prompt = _history_to_prompt(message, history, system_text)
         resp = client.chat(
             model="command-r7b-12-2024",
             message=prompt,
@@ -190,8 +172,8 @@ def load_local_model():
         mdl.config.eos_token_id = tok.eos_token_id
     return mdl, tok
 
-def build_inputs(tokenizer, message, history):
-    msgs = [{"role": "system", "content": SYSTEM_MASTER}]
+def build_inputs(tokenizer, message, history, system_text=SYSTEM_MASTER):
+    msgs = [{"role": "system", "content": system_text}]
     for u, a in _iter_user_assistant(history):
         if u: msgs.append({"role": "user", "content": u})
         if a: msgs.append({"role": "assistant", "content": a})
@@ -243,26 +225,63 @@ def _mdsi_block():
         "outcomes_summary": outcomes
     }, indent=2)
 
-# ---------- Core chat logic (two-phase) ----------
-def clarityops_reply(user_msg, history, tz, uploaded_files_paths, awaiting_answers=False):
+# ---------- Scenario auto-detection ----------
+_SCENARIO_HEADINGS = [
+    "context", "background", "scenario", "case study",
+    "data inputs", "inputs", "evaluation questions", "questions",
+    "recommendations", "deployment strategy", "next steps", "assumptions"
+]
+_SCENARIO_KEYWORDS = [
+    "diabetes", "screening", "metabolic", "prevalence", "capacity",
+    "cost", "startup", "ongoing", "clinical", "a1c", "mmhg", "bmi",
+    "cholesterol", "settlements", "program", "mobile", "ops", "throughput"
+]
+
+def _looks_like_scenario(text: str, uploaded_paths) -> bool:
+    if not text:
+        return False
+    t = text.strip()
+    low = t.lower()
+
+    # 1) Length + structure signals
+    if len(t) >= 450 and any(h in low for h in _SCENARIO_HEADINGS):
+        return True
+
+    # 2) Strong clinical/ops vocabulary density
+    kw_hits = sum(1 for k in _SCENARIO_KEYWORDS if k in low)
+    if len(t) >= 350 and kw_hits >= 4:
+        return True
+
+    # 3) Table/percent/metrics hints
+    if re.search(r"\b\d{2,4}\b", low) and re.search(r"%|\bmmhg\b|\bbmi\b|\ba1c\b", low):
+        if len(t) >= 300:
+            return True
+
+    # 4) Files attached (CSV/PDF/DOCX) + domain keywords
+    if uploaded_paths and kw_hits >= 2:
+        return True
+
+    return False
+
+# ---------- Core chat logic (auto scenario) ----------
+def clarityops_reply(user_msg, history, tz, uploaded_files_paths, mode="chat"):
     """
-    awaiting_answers:
-      - False: Phase 1 -> generate clarification questions and WAIT
-      - True:  Phase 2 -> consume clarifications and produce structured analysis
+    mode: "chat" (default) or "awaiting_answers"
+    Returns: (updated_history, updated_mode)
     """
     try:
-        log_event("user_message", None, {"sizes": {"chars": len(user_msg or "")}})
+        log_event("user_message", None, {"sizes": {"chars": len(user_msg or "")}, "mode": mode})
 
         # 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)], awaiting_answers
+            return history + [(user_msg, ans)], mode
 
         # 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)], awaiting_answers
+            return history + [(user_msg, ans)], mode
 
         # Ingest uploads
         if uploaded_files_paths:
@@ -279,20 +298,9 @@ def clarityops_reply(user_msg, history, tz, uploaded_files_paths, awaiting_answe
         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))], awaiting_answers
-
-        # Pull CSV summaries (NEW) and format a concise block for the prompt
-        csv_summaries = _session_rag.get_csv_summaries()
-        csv_block = ""
-        if csv_summaries:
-            # Cap to 3 latest digests to keep prompt lean
-            lines = ["Uploaded Data Summaries (auto-parsed)"]
-            for s in csv_summaries[:3]:
-                digest = s.get("digest") or ""
-                lines.append(f"- {digest}")
-            csv_block = "\n" + "\n".join(lines)
-
-        # Session retrieval to enrich the system preamble
+                return history + [(user_msg, "Here are the column names from your most recent CSV upload:\n\n- " + "\n- ".join(cols))], mode
+
+        # Session retrieval & context
         session_snips = "\n---\n".join(_session_rag.retrieve(
             "diabetes screening Indigenous Métis mobile program cost throughput outcomes logistics",
             k=6
@@ -308,68 +316,104 @@ def clarityops_reply(user_msg, history, tz, uploaded_files_paths, awaiting_answe
         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 or "")) > 0 else ""
-        # Inject the CSV summary digests into the system preamble so both phases see them
-        scenario_plus_csv = scenario_block + (("\n\n" + csv_block) if csv_block else "")
-
         system_preamble = build_system_preamble(
             snapshot=snapshot,
             policy_context=policy_context,
             computed_numbers=computed,
-            scenario_text=scenario_plus_csv + (f"\n\nExecutive Pre-Computed Blocks:\n{mdsi_extra}" if mdsi_extra else ""),
+            scenario_text=scenario_block + (f"\n\nExecutive Pre-Computed Blocks:\n{mdsi_extra}" if mdsi_extra else ""),
             session_snips=session_snips
         )
 
-        # Phase directive
-        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:
+        # -------- Auto-routing --------
+        if mode == "awaiting_answers":
+            # Any reply now triggers Phase 2
             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\nClarification answers from user:\n" + (safe_in or "<none>") + phase_directive
+
+            out = cohere_chat(augmented_user, history)
+            if not out:
+                model, tokenizer = load_local_model()
+                inputs = build_inputs(tokenizer, augmented_user, history)
+                out = local_generate(model, tokenizer, inputs, max_new_tokens=MAX_NEW_TOKENS)
+
+            if isinstance(out, str):
+                for tag in ("Assistant:", "System:", "User:"):
+                    if out.startswith(tag):
+                        out = out[len(tag):].strip()
+            out = _sanitize_text(out)
+
+            safe_out, blocked_out, reason_out = safety_filter(out, mode="output")
+            if blocked_out:
+                safe_out = refusal_reply(reason_out)
+
+            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": False
+            })
+            return history + [(user_msg, safe_out)], "chat"
+
+        # Normal chat — unless it looks like a scenario
+        if not _looks_like_scenario(safe_in, uploaded_files_paths):
+            normal_user = SYSTEM_MASTER + "\n\n" + system_preamble + "\n\nUser message:\n" + safe_in
+            out = cohere_chat(normal_user, history)
+            if not out:
+                model, tokenizer = load_local_model()
+                inputs = build_inputs(tokenizer, normal_user, history)
+                out = local_generate(model, tokenizer, inputs, max_new_tokens=MAX_NEW_TOKENS)
+
+            if isinstance(out, str):
+                for tag in ("Assistant:", "System:", "User:"):
+                    if out.startswith(tag):
+                        out = out[len(tag):].strip()
+            out = _sanitize_text(out)
+
+            safe_out, blocked_out, reason_out = safety_filter(out, mode="output")
+            if blocked_out:
+                safe_out = refusal_reply(reason_out)
+
+            log_event("assistant_reply", None, {
+                **hash_summary("prompt", normal_user if not PERSIST_CONTENT else ""),
+                **hash_summary("reply", safe_out if not PERSIST_CONTENT else ""),
+                "awaiting_next_phase": False
+            })
+            return history + [(user_msg, safe_out)], "chat"
+
+        # Scenario detected -> Phase 1
+        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"
+        )
+        augmented_user = SYSTEM_MASTER + "\n\n" + system_preamble + "\n\nUser scenario:\n" + safe_in + phase_directive
 
-        augmented_user = SYSTEM_MASTER + "\n\n" + system_preamble + "\n\nUser message:\n" + safe_in + phase_directive
-
-        # Call LLM
         out = cohere_chat(augmented_user, history)
         if not out:
             model, tokenizer = load_local_model()
             inputs = build_inputs(tokenizer, augmented_user, history)
             out = local_generate(model, tokenizer, inputs, max_new_tokens=MAX_NEW_TOKENS)
 
-        # Clean + sanitize
         if isinstance(out, str):
             for tag in ("Assistant:", "System:", "User:"):
                 if out.startswith(tag):
                     out = out[len(tag):].strip()
         out = _sanitize_text(out)
 
-        # Safety (output)
         safe_out, blocked_out, reason_out = safety_filter(out, mode="output")
         if blocked_out:
             safe_out = refusal_reply(reason_out)
 
-        # Flip phase state based on headers
-        new_awaiting = awaiting_answers
-        low = safe_out.lower()
-        if not awaiting_answers and "clarification questions" in low:
-            new_awaiting = True
-        elif awaiting_answers and "structured analysis" in low:
-            new_awaiting = False
-
         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
+            "awaiting_next_phase": True
         })
-
-        return history + [(user_msg, safe_out)], new_awaiting
+        return history + [(user_msg, safe_out)], "awaiting_answers"
 
     except Exception as e:
         err = f"Error: {e}"
@@ -377,7 +421,7 @@ def clarityops_reply(user_msg, history, tz, uploaded_files_paths, awaiting_answe
             traceback.print_exc()
         except Exception:
             pass
-        return history + [(user_msg, err)], awaiting_answers
+        return history + [(user_msg, err)], mode
 
 # ---------- Theme & CSS ----------
 theme = gr.themes.Soft(primary_hue="teal", neutral_hue="slate", radius_size=gr.themes.sizes.radius_lg)
@@ -410,13 +454,17 @@ with gr.Blocks(theme=theme, css=custom_css, analytics_enabled=False) as demo:
             gr.HTML("<h2>What can I help with?</h2>")
             with gr.Row(elem_classes="search-row"):
                 hero_msg = gr.Textbox(
-                    placeholder="Ask anything (paste scenarios here; you can attach files after)...",
+                    placeholder="Ask anything — paste a scenario (and attach files) to trigger structured analysis.",
                     show_label=False,
                     lines=1,
                     elem_classes="hero-box"
                 )
                 hero_send = gr.Button("➤", scale=0)
-            gr.Markdown('<div class="hint">ClarityOps will first ask up to 5 clarifications, then produce a structured analysis.</div>')
+            gr.Markdown(
+                '<div class="hint">Tip: Pasting a structured medical scenario (with sections like '
+                '<i>Context, Data Inputs, Evaluation Questions</i>) will auto-trigger clarifications first, '
+                'then the final analysis. CSVs are auto-summarized.</div>'
+            )
 
     # --- MAIN APP (hidden until first message) ---
     with gr.Column(elem_id="chat-container", visible=False) as app_wrap:
@@ -430,7 +478,7 @@ with gr.Blocks(theme=theme, css=custom_css, analytics_enabled=False) as demo:
             msg = gr.Textbox(
                 label="",
                 show_label=False,
-                placeholder="Continue here. Paste scenario details, add files below.",
+                placeholder="Chat freely… Paste a scenario to auto-start clarifications.",
                 scale=10
             )
             send = gr.Button("Send", scale=1)
@@ -439,8 +487,9 @@ with gr.Blocks(theme=theme, css=custom_css, analytics_enabled=False) as demo:
     # ---- State
     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
+    state_mode = gr.State(value="chat")  # "chat" or "awaiting_answers"
 
+    # ---- Uploads
     def _store_uploads(files, current):
         paths = []
         for f in (files or []):
@@ -449,25 +498,27 @@ 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, awaiting):
+    # ---- Core send (used by both hero input and chat input)
+    def _on_send(user_msg, history, up_paths, mode):
         try:
             if not user_msg or not user_msg.strip():
-                return history, "", history, awaiting
-            new_history, new_awaiting = clarityops_reply(
-                user_msg.strip(), history or [], None, up_paths or [], awaiting_answers=awaiting
+                return history, "", history, mode
+            new_history, new_mode = clarityops_reply(
+                user_msg.strip(), history or [], None, up_paths or [], mode=mode
             )
-            return new_history, "", new_history, new_awaiting
+            return new_history, "", new_history, new_mode
         except Exception as e:
             err = f"Error: {e}"
             try: traceback.print_exc()
             except Exception: pass
             new_hist = (history or []) + [(user_msg or "", err)]
-            return new_hist, "", new_hist, awaiting
+            return new_hist, "", new_hist, mode
 
-    def _hero_start(user_msg, history, up_paths, awaiting):
-        chat_o, msg_o, hist_o, await_o = _on_send(user_msg, history, up_paths, awaiting)
+    # ---- Hero -> App transition + first send
+    def _hero_start(user_msg, history, up_paths, mode):
+        chat_o, msg_o, hist_o, mode_o = _on_send(user_msg, history, up_paths, mode)
         return (
-            chat_o, msg_o, hist_o, await_o,
+            chat_o, msg_o, hist_o, mode_o,
             gr.update(visible=False),   # hide hero
             gr.update(visible=True),    # show app
             ""                          # clear hero box
@@ -475,39 +526,36 @@ with gr.Blocks(theme=theme, css=custom_css, analytics_enabled=False) as demo:
 
     hero_send.click(
         _hero_start,
-        inputs=[hero_msg, state_history, state_uploaded, state_awaiting],
-        outputs=[chat, msg, state_history, state_awaiting, hero_wrap, app_wrap, hero_msg],
+        inputs=[hero_msg, state_history, state_uploaded, state_mode],
+        outputs=[chat, msg, state_history, state_mode, hero_wrap, app_wrap, hero_msg],
         concurrency_limit=2, queue=True
     )
     hero_msg.submit(
         _hero_start,
-        inputs=[hero_msg, state_history, state_uploaded, state_awaiting],
-        outputs=[chat, msg, state_history, state_awaiting, hero_wrap, app_wrap, hero_msg],
+        inputs=[hero_msg, state_history, state_uploaded, state_mode],
+        outputs=[chat, msg, state_history, state_mode, hero_wrap, app_wrap, hero_msg],
         concurrency_limit=2, queue=True
     )
 
-    send.click(_on_send, inputs=[msg, state_history, state_uploaded, state_awaiting],
-               outputs=[chat, msg, state_history, state_awaiting],
+    # ---- Normal chat interactions after hero is gone
+    send.click(_on_send, inputs=[msg, state_history, state_uploaded, state_mode],
+               outputs=[chat, msg, state_history, state_mode],
                concurrency_limit=2, queue=True)
-    msg.submit(_on_send, inputs=[msg, state_history, state_uploaded, state_awaiting],
-               outputs=[chat, msg, state_history, state_awaiting],
+    msg.submit(_on_send, inputs=[msg, state_history, state_uploaded, state_mode],
+               outputs=[chat, msg, state_history, state_mode],
                concurrency_limit=2, queue=True)
 
     def _on_clear():
+        # reset to fresh hero screen and chat mode
         return (
-            [], "", [], False,
+            [], "", [], "chat",
             gr.update(visible=True),   # show hero
             gr.update(visible=False),  # hide app
             ""                         # clear hero input
         )
 
-    clear.click(_on_clear, None, [chat, msg, state_history, state_awaiting, hero_wrap, app_wrap, hero_msg])
+    clear.click(_on_clear, None, [chat, msg, state_history, state_mode, hero_wrap, app_wrap, hero_msg])
 
 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)
-
-
-
-
-