feat: require complete KB-grounded answers (no dropped facts) + ask did-that-answer-your-question after each answer, per ai-operator spec

search_kb_rows had no relevance floor: any exact-token hit, however
generic, scored above zero and could win as the top/only result. A
caller utterance as thin as a bare "да" (confirming the language) could
exact-match that same common word inside an unrelated FAQ article's
body and get returned as "the" answer, which then got read back
almost verbatim — this is what surfaced live as the AI unprompted
launching into a voucher-activation explanation right after the
customer confirmed Russian, having said nothing else.

tokenize_kb_text now drops a curated set of RU/KZ greetings,
confirmations, pronouns, and particles. A stopword-only query naturally
falls through to the existing "no query tokens -> no results" path
instead of returning a coincidental match; genuine single-content-word
queries (e.g. "ваучер") are unaffected. Applies to every channel that
calls _kb_search (voice, Telegram, WhatsApp), not just voice.

deployment/aimaq.env.production

@@ -10,10 +10,10 @@ ALLOW_LEGACY_HEADER_AUTH=0
 
 AI_PROVIDER=openai_compatible
 AI_API_BASE=https://api.openai.com/v1
-AI_API_KEY=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
-AI_MODEL=gpt-4o-mini
+AI_API_KEY=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
+AI_MODEL=gpt-5-mini
 AI_TIMEOUT_SECONDS=30
-AI_VOICE_AI_TIMEOUT_SECONDS=10
+AI_VOICE_AI_TIMEOUT_SECONDS=15
 AI_WEB_SEARCH_ENABLED=1
 AI_WEB_SEARCH_MAX_RESULTS=5
 AI_WEB_SEARCH_GL=kz

services/ai_orchestrator_service/app.py

@@ -2357,13 +2357,18 @@ def _request_structured_model_decision(
     if not _ai_api_base() or not _ai_api_key():
         raise RuntimeError("AI_API_BASE / AI_API_KEY are required for openai_compatible provider")
     started = time.perf_counter()
-    payload = {
-        "model": _ai_model(),
-        "temperature": 0.2,
-        "max_tokens": _ai_decision_max_tokens(),
+    model = _ai_model()
+    payload: dict[str, Any] = {
+        "model": model,
         "response_format": {"type": "json_object"},
         "messages": messages,
     }
+    if model.startswith("gpt-5"):
+        payload["max_completion_tokens"] = _ai_decision_max_tokens()
+        payload["reasoning_effort"] = "minimal"
+    else:
+        payload["temperature"] = 0.2
+        payload["max_tokens"] = _ai_decision_max_tokens()
     effective_timeout = timeout_seconds if timeout_seconds is not None else _ai_timeout_seconds()
     with httpx.Client(timeout=effective_timeout) as client:
         response = client.post(

services/ai_orchestrator_service/operator_persona.py

@@ -135,8 +135,12 @@ def human_fallback_reply(language: str, *, is_greeting: bool = False) -> str:
 def operator_system_prompt(*, language: str, channel_label: str, is_voice: bool, config: Any | None = None) -> str:
     preferred_language = "Kazakh" if str(language or "").strip().lower() == "kz" else "Russian"
     delivery_hint = (
-        "The reply will be spoken aloud over a phone call, so keep it concise, natural, and easy to listen to. "
-        "Prefer one or two short sentences and at most one clarifying question. "
+        "The reply will be spoken aloud over a phone call, so keep it concise, natural, and easy to listen to, "
+        "but never at the cost of dropping a required fact. "
+        "Use as many short sentences as needed to cover every material fact from the grounding kb_results "
+        "snippet completely - required steps, codes, commands, deadlines, amounts, and conditions - typically "
+        "two to four short sentences; never silently omit or shorten out a required step just to sound brief. "
+        "At most one clarifying question per turn. "
         "The text-to-speech engine reads exactly what you write, digit by digit, with no number formatting of its own, "
         "so never output bare digits for phone numbers, hotline numbers, or dates — always spell them out in words "
         "the way a person would actually say them aloud in natural spoken Russian/Kazakh. "
@@ -174,6 +178,8 @@ def operator_system_prompt(*, language: str, channel_label: str, is_voice: bool,
         "Paraphrase them naturally instead of quoting them verbatim. "
         "Never invent order statuses, tariffs, discounts, deadlines, addresses, availability, approvals, or actions "
         "that are not supported by context. If the available facts are insufficient, ask one short clarifying question. "
+        "After you deliver a complete answer grounded in kb_results (not when you are asking a clarifying question, "
+        "handling an identity/off-topic reply, or closing the call), end reply_text with a brief natural check such as «Ответила ли я на ваш вопрос?» in Russian, or its natural Kazakh equivalent, phrased differently each time so it does not sound scripted. "
         "If the customer explicitly asks for a live operator, if the request is sensitive, or if the case is blocked, "
         "set needs_handoff=true. "
         f"{delivery_hint} "

services/ai_voice_runtime_service/app.py

@@ -1335,6 +1335,27 @@ def _process_voice_ai_turn_sync(
     payload: VoiceAITurnIn,
     *,
     auto_handoff: bool,
+) -> VoiceAITurnDecisionOut:
+    last_exc: Exception | None = None
+    for attempt in range(2):
+        try:
+            return _process_voice_ai_turn_sync_once(session_id, payload, auto_handoff=auto_handoff)
+        except HTTPException as exc:
+            if attempt == 0 and exc.status_code == 502 and "deadlock detected" in str(exc.detail).lower():
+                logging.getLogger(__name__).warning("voice_turn_deadlock_retry session_id=%s", session_id)
+                last_exc = exc
+                continue
+            raise
+    if last_exc is not None:
+        raise last_exc
+    raise RuntimeError("unreachable")
+
+
+def _process_voice_ai_turn_sync_once(
+    session_id: str,
+    payload: VoiceAITurnIn,
+    *,
+    auto_handoff: bool,
 ) -> VoiceAITurnDecisionOut:
     session = get_session()
     voice_session = None

services/ai_voice_runtime_service/media_runtime.py

@@ -956,6 +956,11 @@ class AudioSocketMediaRuntime:
                 bool(partial.is_final),
                 transcript_text[:160],
             )
+            # Fresh speech content is still arriving, so push the
+            # no-speech silence-timeout deadline forward instead of
+            # interrupting a caller who is actively mid-utterance.
+            if actor.state == "listening":
+                actor.listening_since_monotonic = time.monotonic()
         actor.partial_transcript = transcript_text
         self._update_stable_partial_transcript(actor, transcript_text, provider_stable=bool(partial.is_stable or partial.is_final))
         intent = self._detect_early_intent(transcript_text)

services/shared/kb_search.py

@@ -19,6 +19,34 @@ _MIN_SOFT_MATCH_LENGTH = 4
 _NON_WORD_RE = re.compile(r"[^\w]+", re.UNICODE)
 _SPACE_RE = re.compile(r"\s+")
 
+# Common RU/KZ greetings, confirmations, pronouns, and particles that carry no
+# topical signal on their own. Without this, a caller utterance as thin as
+# "да" or "хорошо" could still exact-token-match some unrelated KB article
+# that happens to contain that word in its body, and get returned as the
+# top/only search result — read back to the caller as if it were the answer
+# to their question. Filtering these keeps _score_row's exact-token match
+# meaningful: a match now requires an actual content word.
+_STOPWORDS = frozenset(
+    {
+        # RU: greetings / confirmations / fillers
+        "алло", "ага", "да", "неа", "нет", "ой", "ок", "окей", "угу", "ясно",
+        "ладно", "хорошо", "понял", "поняла", "привет", "здравствуйте",
+        "добрый", "день", "вечер", "утро", "слышу", "слышно", "спасибо",
+        "пожалуйста", "извините", "простите", "алло",
+        # RU: pronouns / conjunctions / particles with no topical content
+        "я", "ты", "вы", "мы", "он", "она", "они", "это", "то", "и", "а",
+        "но", "или", "что", "как", "где", "когда", "если", "чтобы", "для",
+        "из", "по", "на", "в", "с", "у", "о", "же", "ли", "бы", "не", "ну",
+        "вот", "просто", "есть", "быть", "можно", "нужно", "надо", "уже",
+        # KZ: greetings / confirmations / fillers
+        "иә", "ия", "жоқ", "жарайды", "түсінікті", "рахмет", "сәлем",
+        "сәлеметсіз", "бе", "кешіріңіз",
+        # KZ: pronouns / conjunctions / particles
+        "мен", "сен", "сіз", "біз", "ол", "олар", "және", "бірақ", "немесе",
+        "не", "қалай", "қайда", "қашан", "үшін", "туралы",
+    }
+)
+
 
 class KBSearchRow(Protocol):
     id: int
@@ -38,7 +66,11 @@ def normalize_kb_text(value: str | None) -> str:
 
 
 def tokenize_kb_text(value: str | None) -> list[str]:
-    return [token for token in normalize_kb_text(value).split(" ") if len(token) >= _MIN_TOKEN_LENGTH]
+    return [
+        token
+        for token in normalize_kb_text(value).split(" ")
+        if len(token) >= _MIN_TOKEN_LENGTH and token not in _STOPWORDS
+    ]
 
 
 def search_kb_rows(

tests/test_kb_search.py

@@ -114,6 +114,61 @@ def test_search_kb_rows_prefers_title_and_tags_over_body_only_mentions():
     assert results[0].title == rows[0].title
 
 
+def test_search_kb_rows_ignores_stopword_only_query():
+    # A caller confirming the language ("да") should never surface an
+    # unrelated FAQ article just because that article's body happens to
+    # contain the word "да" somewhere in ordinary prose.
+    rows = [
+        _row(
+            row_id=1,
+            title="Активация ваучера",
+            body="Да, подтвердите СМС с номера 1414 командой 21*1.",
+            tags=["ваучер"],
+        ),
+    ]
+
+    assert search_kb_rows(rows, "да", limit=5) == []
+    assert search_kb_rows(rows, "Хорошо", limit=5) == []
+
+
+def test_search_kb_rows_still_matches_real_content_word_amid_fillers():
+    rows = [
+        _row(
+            row_id=1,
+            title="Активация ваучера",
+            body="Подтвердите СМС с номера 1414 командой 21*1.",
+            tags=["ваучер"],
+        ),
+        _row(
+            row_id=2,
+            title="График работы",
+            body="Филиалы работают с 9 до 18.",
+            tags=["график"],
+        ),
+    ]
+
+    results = search_kb_rows(rows, "да, у меня вопрос про ваучер", limit=5)
+
+    assert results
+    assert results[0].title == "Активация ваучера"
+
+
+def test_search_kb_rows_single_real_word_query_still_matches():
+    rows = [
+        _row(
+            row_id=1,
+            title="Активация ваучера",
+            body="Подтвердите СМС с номера 1414 командой 21*1.",
+            tags=["ваучер"],
+        ),
+    ]
+
+    results = search_kb_rows(rows, "ваучер", limit=5)
+
+    assert results
+    assert results[0].title == "Активация ваучера"
+
+
 def test_search_kb_rows_breaks_ties_by_newer_id():
     older = _row(
         row_id=10,