feat(chat): scenario-based training — choice of channel/Situation + realistic per-turn evaluation

Backend: - Channel/initiation now driven by a SCENARIO chosen at chat start, not baked into the persona: social (customer opens), f2f_call (seller must open, proactive), recontact (customer re-contacts after 1-3 months). - /chat/start accepts {scenario}; session stores scenario + internal{turns,score}. - persona_reply takes scenario + adapts tone; system-role transcript entries are fed to the persona as hidden scene notes. - JUDGE updated for realism: good response can WIN even in hard/tough-text scenarios; long/no-close chats (turns >~12) lose; pushy/ignoring-need loses. Efficiency rewarded. Frontend: - Scenario picker before chat (choose Social / Face-to-face-call / Re-contact). - Chat thread renders role=system as a centered time-lapse/scene note. - Choose-scenario i18n (EN+TH). Rebuilt dist.
2026-08-08 11:27:05 +07:00
parent 8670addcc3
commit bd6a7ffa32
27 changed files with 314 additions and 110 deletions
--- a/backend/app/api/chat_routes.py
+++ b/backend/app/api/chat_routes.py
@@ -27,6 +27,35 @@ def _sim(group, persona):
    return Simulator(llm)


+SCENARIOS = {
+    "social": {
+        "label": "Social Media",
+        "init": "customer",
+        "system_preamble": "💬 ช่องทาง ข้อความโซเชียล — ลูกค้าทักมาหาคุณก่อน (โทนสั้น ทักๆ ตามสไตล์แชท)",
+        "adapt": "You are chatting on a social-messaging app (LINE-style). Keep replies SHORT, casual, and quick — one line to a few lines. The customer opened the chat.",
+    },
+    "f2f_call": {
+        "label": "พบหน้า / โทรศัพท์",
+        "init": "seller",
+        "system_preamble": "📞 สถานการณ์ พบหน้าหรือโทรศัพท์ — คุณต้องเป็นฝ่ายเปิดการสนทนาเชิงรุกกับลีด (lead)",
+        "adapt": "This is a face-to-face or phone sales situation. You must sound natural and conversational like a live talk. The seller will open — you are the lead they are reaching out to.",
+    },
+    "recontact": {
+        "label": "ลูกค้ากลับมาติดต่อ (เคยได้ข้อมูล 1-3 เดือน)",
+        "init": "customer",
+        "system_preamble": "⏳ 1-3 เดือนผ่านไป... ลูกค้าคนนี้เคยได้รับข้อมูลสินค้าไปแล้ว ตอนนี้กลับมาติดต่อคุณอีกครั้ง (พร้อมตัดสินใจมากขึ้น)",
+        "adapt": "This customer researched you 1-3 months ago. They have already read about the product and are now more ready to decide. They re-contacted you on a messaging channel. Keep replies natural and fairly concise.",
+    },
+}
+
+
+def _scenario_config(scenario: str, persona: dict):
+    cfg = SCENARIOS.get(scenario, SCENARIOS["social"])
+    # OR with persona's own initiation preference: f2f/call forces seller-first unless persona
+    # is strongly customer-initiated (we let scenario win, but keep persona special handling).
+    return cfg, cfg["init"]
+
+
 def _get_ready_group(s, gid: str) -> dict:
    """Org-scoped group access for trainees + require ready status (IDOR defense)."""
    group = s["groups"].get_or_none(gid)
@@ -53,6 +82,12 @@ def start_session(gid: str, pid: str):
    if not persona:
        raise ApiError("persona not found", 404)
    actor = current_user()
+    # Scenario chosen by the trainee at chat start (not baked into the persona).
+    body = request.get_json(silent=True) or {}
+    scenario = (body.get("scenario") or "social").strip().lower()
+    if scenario not in ("social", "f2f_call", "recontact"):
+        scenario = "social"
+    scenario_meta, init_mode = _scenario_config(scenario, persona)
    # One-shot: reject if already finished this persona
    try:
        session = s["sessions"].create(
@@ -60,27 +95,35 @@ def start_session(gid: str, pid: str):
            persona_name=persona.get("name", "?"),
            persona_meta={
                "tier": persona.get("tier"),
-                "initiation_mode": persona.get("initiation_mode"),
+                "initiation_mode": init_mode,
                "channel": persona.get("channel"),
+                "scenario": scenario,
            },
        )
    except ValueError as exc:
        raise ApiError(str(exc), 400)

+    s["sessions"].update(
+        session["id"],
+        scenario=scenario,
+        internal={**(session.get("internal") or {}), "turns": 0, "score": 50, "signals": []},
+    )
    sim = _sim(group, persona)
-    # Seller-initiated: give the trainee an opening task (no persona message yet).
-    init_mode = persona.get("initiation_mode", "customer")
+    # Seed messages: system note(s) then customer opener for customer-first scenarios.
+    seeded = list(session.get("messages", []))
+    if scenario_meta.get("system_preamble"):
+        seeded.append({"role": "system", "text": scenario_meta["system_preamble"]})
    if init_mode == "customer":
-        # Customer opens: inject the persona's opener as the first message.
        opener = persona.get("opener") or "Hi, I saw your product and had a question."
-        s["sessions"].update(session["id"], messages=[{"role": "customer", "text": opener}])
-    else:
-        s["sessions"].update(
-            session["id"],
-            task="The customer did NOT message first. You must open the sale — start the "
-                 "conversation with this lead (e.g. introduce yourself and engage with interest).",
-        )
-    return jsonify({"session": s["sessions"].get(session["id"]), "initiation_mode": init_mode})
+        seeded.append({"role": "customer", "text": opener})
+    s["sessions"].update(session["id"], messages=seeded)
+    sess = s["sessions"].get(session["id"])
+    return jsonify({
+        "session": sess,
+        "initiation_mode": init_mode,
+        "scenario": scenario,
+        "scenario_meta": scenario_meta,
+    })


@chat_bp.post("/<gid>/personas/<pid>/chat/send")
@@ -106,6 +149,8 @@ def send_message(gid: str, pid: str):
        raise ApiError("session context missing", 404)
    messages = list(session.get("messages", []))
    messages.append({"role": "seller", "text": text})
+    scenario = session.get("scenario", "social") or "social"
+    adapt = SCENARIOS.get(scenario, SCENARIOS["social"]).get("adapt", "")

    sim = _sim(group, persona)
    try:
@@ -114,12 +159,19 @@ def send_message(gid: str, pid: str):
            sales_kit=group.get("sales_kit") or {},
            messages=messages,
            internal=session.get("internal", {}),
+            scenario=scenario,
+            scenario_adapt=adapt,
        )
    except LLMError as exc:
        raise ApiError(f"LLM error: {exc}", 500)
    messages.append({"role": "customer", "text": reply})

-    s["sessions"].update(session["id"], messages=messages)
+    # Per-turn tracking: count exchanged turns + nudge score down as the chat drags.
+    internal = session.get("internal", {}) or {}
+    internal.setdefault("turns", 0)
+    internal["turns"] = internal.get("turns", 0) + 1
+
+    s["sessions"].update(session["id"], messages=messages, internal=internal)
    return jsonify({"reply": reply, "messages": messages})


@@ -139,7 +191,9 @@ def finish_session(gid: str, pid: str):
    sim = _sim(group, persona)
    messages = session.get("messages", [])
    try:
-        verdict = sim.judge(persona=persona, messages=messages)
+        verdict = sim.judge(
+            persona=persona, messages=messages, internal=session.get("internal", {})
+        )
    except LLMError as exc:
        raise ApiError(f"LLM error: {exc}", 500)

--- a/backend/app/services/simulator.py
+++ b/backend/app/services/simulator.py
@@ -47,9 +47,18 @@ A sale is CLOSED only if BOTH:
     AND
  2. The customer verbally accepts the offer/price (in the final exchange).

-Otherwise it is LOST (or abandoned if the user ended early).
+REALISM RULES:
+- A good response can WIN even in a hard scenario (e.g. customer who 'texted wrong', 'changed their
+  mind', or has been silent). If the seller re-engages gently, re-qualifies the real need, and closes,
+  it's a WIN. Do NOT auto-fail on special cases — always reward genuinely skillful recovery.
+- LOST reflects the persona TYPICALLY losing (real-world >90% of such leads do not convert), but the
+  trainee's skill evaluation must remain fair: a strong close beats a weak one, always.
+- If the chat drags on many turns (or turns > ~12) without the seller reaching the pain or closing,
+  treat it as LOST due to failing to convert / the opportunity cooling (mirrors real leads going cold).
+- If the seller was pushy, rude, ignored the need, or mis-diagnosed the pain, mark LOST even if the
+  price was acceptable.

-Scoring (0-100): painResolution + trust + objectionHandling are the only factors.
+Scoring (0-100): painResolution + trust + objectionHandling + efficiency (fewer turns, higher).
 Return JSON:
 {
  "outcome": "won" | "lost",
@@ -76,14 +85,21 @@ class Simulator:
        sales_kit: dict[str, Any],
        messages: list[dict[str, str]],
        internal: dict[str, Any],
+        scenario: str = "social",
+        scenario_adapt: str = "",
    ) -> str:
        pains_txt = self._describe_pains(persona.get("pains", []))
+        adapt = scenario_adapt or {
+            "social": "Chat style: short, casual, quick social-messaging replies.",
+            "f2f_call": "Style: natural, conversational like a live face-to-face or phone talk.",
+            "recontact": "Style: casual messaging; you already know the product from 1-3 months ago.",
+        }.get(scenario, "")
        system = CHAT_SYSTEM.format(
            name=persona.get("name", "Customer"),
            tone=persona.get("communication_style", "natural, casual"),
            profession=persona.get("profession", "customer"),
            age_group=persona.get("age_group", "adult"),
-            channel=persona.get("channel", "facebook"),
+            channel=persona.get("channel", "facebook") + (f" ({scenario})" if scenario else ""),
            background=persona.get("background", ""),
            personality=persona.get("personality", ""),
            lifestyle=persona.get("lifestyle", ""),
@@ -96,7 +112,7 @@ class Simulator:
            init_mode="you contacted the seller first (customer-initiated)"
            if persona.get("initiation_mode") == "customer"
            else "the seller opened the sale to you (you are a lead)",
-            special_instr=self._special_instr(persona),
+            special_instr=self._special_instr(persona) + "\n" + adapt,
        )
        msgs = [{"role": "system", "content": system}]
        # send a compact recap of internal state to the persona ad
@@ -106,7 +122,13 @@ class Simulator:
            "content": "Internal state (for your role-play only): "
            + json.dumps(internal, ensure_ascii=False),
        })
-        msgs.extend(messages[-30:])  # context window
+        # Translate role 'system' transcript entries into a hidden system note for the LLM.
+        for m in messages[-30:]:
+            role = m.get("role")
+            if role == "system":
+                msgs.append({"role": "system", "content": f"[scene note from transcript]: {m.get('text')}"})
+            else:
+                msgs.append({"role": role, "content": m.get("text", "")})
        try:
            resp = self.llm.complete_conversation(msgs, temperature=0.7, max_tokens=400)
        except LLMError as exc:
@@ -125,6 +147,7 @@ class Simulator:
        *,
        persona: dict[str, Any],
        messages: list[dict[str, str]],
+        internal: dict[str, Any] | None = None,
    ) -> dict[str, Any]:
        persona_summary = json.dumps({
            "name": persona.get("name"),
@@ -136,7 +159,16 @@ class Simulator:
        transcript = "\n".join(
            f"{m.get('role')}: {m.get('text')}" for m in messages[-40:]
        )
-        user_prompt = f"PERSONA:\n{persona_summary}\n\nTRANSCRIPT:\n{transcript}"
+        state_note = ""
+        if internal:
+            try:
+                state_note = (
+                    "\n\nINTERNAL (hidden, for judging only): "
+                    f"turns={internal.get('turns', 0)}, score_trend={internal.get('score', 50)}"
+                )
+            except Exception:
+                state_note = ""
+        user_prompt = f"PERSONA:\n{persona_summary}\n\nTRANSCRIPT:\n{transcript}{state_note}"
        try:
            result = self.judge_llm.complete_json(
                JUDGE_SYSTEM, user_prompt, temperature=0.2, max_tokens=2000