Author SHA1 Message Date
arys 99d169ec67 fix: stop the no-answer retry from racing the dialplan hangup + fix event catalog validation
Two bugs found via a live test with two real registered browser softphones:

1. retry_escalation_no_answer() re-resolved the client channel via
   _resolve_handoff_channel() -> a live AMI CoreShowChannels round-trip that
   can take ~10s. The new mvpcc-transfer dialplan wait window (MusicOnHold,
   also ~10s, added to give the backend time to redirect before the final
   Hangup) was consistently LOST to this exact same duration: the backend's
   AMI Redirect fired against a channel the dialplan had already hung up
   ('Channel does not exist: PJSIP/...', confirmed in escalation timeline).
   Fixed by reusing the actively-maintained AsteriskCallLinkRow.channel_name
   directly (unchanged for a PJSIP channel across Redirect between contexts
   of the same call) instead of re-discovering it, falling back to the slow
   path only if that field is empty.

2. VoiceEventIn.event_type is a pydantic Literal restricted to 6 legacy
   values (call.started/ivr.completed/...). None of the Phase 2 event
   catalog names (AgentReserved/AgentRinging/AgentNoAnswer/AgentConnected/
   TransferCompleted/TransferFailed) were ever in it, so every single
   _emit_escalation_event() call has been failing with 422 since Phase 2
   shipped (swallowed silently by the broad except there) - confirmed by
   calling app._emit_voice_event() directly against the running service.
   Extended the Literal to include all six.
2026-08-31 01:03:47 +05:00
didar 9fdaa9472f Revert "fix: let voice early-plan turns answer from the FAQ knowledge base"
deploy / deploy (push) Successful in 31s
This reverts commit 8ced7a59e3.
2026-08-31 00:54:38 +05:00
arystanbek 1dc37d2764 Merge pull request 'fix: treat AST_CAUSE no-route/unallocated as a no-answer retry outcome' (#10) from fix/no-route-retry-cause into main
deploy / deploy (push) Successful in 32s
2026-08-30 19:43:43 +00:00
arys 8382dfa9ba fix: treat AST_CAUSE no-route/unallocated as a no-answer retry outcome
process_agent_dial_outcome only recognized hangup causes 17/18/19/21/34/38.
When the reserved agent's AOR has zero registered contacts (e.g. the
softphone dropped, or nobody ever registered), Asterisk immediately
hangs up with cause 3 (no route to destination) instead of running a
Dial() long enough to produce a DialEnd/NOANSWER at all - so the retry
listener silently ignored it and the escalation was left dangling in
'ringing' status (the agent itself still got released via the
call-ended fallback path, but no retry to the next agent was ever
attempted and the escalation record never reflects the failure).

Added causes 1 (unallocated number), 3 (no route), 20 (subscriber
absent), 22 (number changed) alongside the existing set.
2026-08-31 00:43:20 +05:00
didar 8ced7a59e3 fix: let voice early-plan turns answer from the FAQ knowledge base
deploy / deploy (push) Successful in 32s
The speculative "early plan" turn (computed on partial ASR, before the
caller finishes talking) can win the race and get spoken as the actual
reply, but it unconditionally skipped KB search and answered common
questions (schedule/address/price/status/problem) with a hardcoded
clarifying question even when the FAQ already had the answer.

KB search is a cheap in-memory lexical scan over a DB-cached row set,
so it fits the early-plan latency budget unlike a real LLM call. Now
early-plan runs it and, on a match, answers from the KB snippet
(intent resolved via normalize_intent) instead of guessing a generic
clarifying question; with no match it falls back to the prior
behavior unchanged. operator_request is unaffected.
2026-08-31 00:38:06 +05:00
didar d2438b6954 feat: canonical intent taxonomy for AI operator (kb_answer -> intent_code)
deploy / deploy (push) Successful in 30s
Centralizes fixed control intents and adds a data-driven intent_code
field on kb_articles so many phrasings of the same FAQ question
resolve to one stable code (e.g. VOUCHER_ACTIVATION) instead of a
free-form, unvalidated string the LLM invented on the fly.

- services/shared/intents.py: CONTROL_INTENTS + normalize_intent()
- kb_articles.intent_code column (ORM + dev/sqlite runtime compat +
  migrations/sql/0034_* for postgres/sqlite)
- kb_service CRUD exposes intent_code
- orchestrator surfaces intent_code to the LLM and validates its
  intent output against control intents + the KB codes shown that turn
- voice.py: _voice_early_intent_bucket renamed to _voice_ack_topic_bucket
  to stop it being conflated with the canonical FAQ intent
2026-08-31 00:17:51 +05:00
arystanbek 3826f5704a Merge pull request 'fix: track escalation for legacy AI voice handoff path' (#9) from fix/handoff-escalation-tracking into main
deploy / deploy (push) Successful in 31s
2026-08-30 19:08:20 +00:00
16 changed files with 272 additions and 10 deletions
@@ -0,0 +1,2 @@
ALTER TABLE kb_articles ADD COLUMN IF NOT EXISTS intent_code TEXT;
CREATE INDEX IF NOT EXISTS idx_kb_articles_intent_code ON kb_articles(intent_code);
@@ -0,0 +1,2 @@
ALTER TABLE kb_articles ADD COLUMN intent_code TEXT;
CREATE INDEX IF NOT EXISTS idx_kb_articles_intent_code ON kb_articles(intent_code);
+17 -4
View File
@@ -9,7 +9,7 @@ import os
import re
import time
from threading import Lock
from typing import Any
from typing import Any, Iterable
import httpx
from fastapi import Depends, FastAPI, HTTPException, Query
@@ -23,6 +23,7 @@ from services.shared.ai_context_summary import (
update_context_summary_from_assistant_turn,
update_context_summary_from_user_turn,
)
from services.shared.intents import normalize_intent
from services.shared.kb_localization import normalize_kb_language
from services.shared.kb_search import search_kb_rows
from services.shared.models import (
@@ -2315,6 +2316,7 @@ def _openai_prompt(
"article_id": article.article_id,
"title": article.title,
"snippet": _article_snippet(article),
"intent_code": getattr(article, "intent_code", None),
}
for article in kb_results
]
@@ -2410,10 +2412,15 @@ def _openai_compatible_decision(
)
def _sanitize_decision(raw: dict[str, Any], *, fallback_language: str) -> dict[str, Any]:
def _sanitize_decision(
raw: dict[str, Any],
*,
fallback_language: str,
known_topic_codes: Iterable[str] = (),
) -> dict[str, Any]:
decision = {
"language": str(raw.get("language") or fallback_language or "ru"),
"intent": str(raw.get("intent") or "unknown"),
"intent": normalize_intent(raw.get("intent"), known_topic_codes=known_topic_codes),
"reply_text": str(raw.get("reply_text") or "").strip(),
"extracted_name": str(raw.get("extracted_name") or "").strip() or None,
"confidence": float(raw.get("confidence") or 0.0),
@@ -2611,7 +2618,13 @@ def _decide_reply(
raw["_model"] = _ai_model()
raw["_latency_ms"] = 1
raw["_finish_reason"] = "stop"
return _sanitize_decision(raw, fallback_language=language)
return _sanitize_decision(
raw,
fallback_language=language,
known_topic_codes=[
code for article in kb_results if (code := getattr(article, "intent_code", None))
],
)
def _update_ai_session_context_summary_from_user_turn(
@@ -178,5 +178,10 @@ def operator_system_prompt(*, language: str, channel_label: str, is_voice: bool,
"set needs_handoff=true. "
f"{delivery_hint} "
"Return only a JSON object with keys: language, intent, reply_text, extracted_name, confidence, needs_handoff, "
"handoff_reason, case_action, kb_refs. case_action must be one of none, close, escalate, keep_open."
"handoff_reason, case_action, kb_refs. case_action must be one of none, close, escalate, keep_open. "
"For `intent`: if your reply is grounded in one of the provided kb_results, set intent to that snippet's "
"intent_code exactly as given (do not translate, reformat, or invent your own code). If no kb_results were "
"used, use one of these fixed values as appropriate: identity_question, handoff_request, sensitive_request, "
"resolution_confirmed, clarification, kb_answer, unknown. Never invent a new intent value outside of these "
"two sources — the platform discards anything else."
)
+16 -3
View File
@@ -1359,7 +1359,13 @@ def _voice_v2_enabled(metadata: dict[str, Any] | None = None) -> bool:
return _voice_policy_mode() in {"v2_fast_conversational", "v2_streaming_duplex"}
def _voice_early_intent_bucket(text: str) -> str:
def _voice_ack_topic_bucket(text: str) -> str:
"""Coarse keyword heuristic used only to pick an ack phrase / early clarifying
question (see _voice_ack_kind_for_intent and _voice_early_plan) while the real,
KB-grounded decision is still in flight. This is NOT the canonical FAQ intent
(see services.shared.intents) and must never be echoed back as the final
decision's `intent` value.
"""
normalized = " ".join(str(text or "").strip().lower().split())
if not normalized:
return "unknown"
@@ -1509,7 +1515,7 @@ def _voice_v2_metadata(
if not _voice_v2_enabled(request_metadata):
return {}
payload = request_metadata if isinstance(request_metadata, dict) else {}
early_intent = _voice_early_intent_bucket(transcript_text)
early_intent = _voice_ack_topic_bucket(transcript_text)
metadata: dict[str, Any] = {
"voice_v2_enabled": True,
"early_intent": early_intent,
@@ -1668,6 +1674,7 @@ def _voice_llm_prompt_messages(
"article_id": article.article_id,
"title": article.title,
"snippet": app._article_snippet(article, limit=240),
"intent_code": getattr(article, "intent_code", None),
}
for article in kb_results[:3]
]
@@ -1767,7 +1774,13 @@ def _voice_llm_decision(
)
except Exception:
return None
decision = app._sanitize_decision(raw, fallback_language=language)
decision = app._sanitize_decision(
raw,
fallback_language=language,
known_topic_codes=[
code for article in kb_results if (code := getattr(article, "intent_code", None))
],
)
if not str(decision.get("reply_text") or "").strip():
decision["reply_text"] = _voice_generic_prompt(language)
if decision.get("needs_handoff") and not decision.get("handoff_reason"):
@@ -1233,7 +1233,7 @@ def process_recording_ready(
_NO_ANSWER_DIAL_STATUSES = {"NOANSWER", "BUSY", "CANCEL", "CHANUNAVAIL", "CONGESTION"}
_NO_ANSWER_HANGUP_CAUSES = {"17", "18", "19", "21", "34", "38"}
_NO_ANSWER_HANGUP_CAUSES = {"1", "3", "17", "18", "19", "20", "21", "22", "34", "38"}
def process_agent_dial_outcome(session, row: AsteriskEventLogRow, payload: dict[str, Any]) -> None:
+5 -1
View File
@@ -938,7 +938,11 @@ def retry_escalation_no_answer(session, *, call_id: str, dial_outcome: str) -> N
)
attempted_ids = json.loads(escalation.attempted_agent_ids_json or "[]")
channel = _resolve_handoff_channel(session, link)
# Use the already-known, actively-maintained channel name directly instead of
# _resolve_handoff_channel()'s live AMI CoreShowChannels re-discovery: that
# round-trip can take ~10s, which races (and loses) against the dialplan's
# own short MusicOnHold-then-hangup wait window for this exact retry path.
channel = str(link.channel_name or "").strip() or _resolve_handoff_channel(session, link)
required_skills = json.loads(escalation.required_skills_json or "[]")
next_agent = _reserve_routing_agent(
call_id=call_id,
+8
View File
@@ -40,6 +40,7 @@ def _article_out(row: KBArticleRow) -> KBArticleOut:
article_id=row.article_id,
category_id=row.category_id,
article_group_id=resolve_article_group_id(row.article_id, row.article_group_id),
intent_code=row.intent_code,
language=normalize_kb_language(row.language),
title=row.title,
body=row.body,
@@ -49,6 +50,10 @@ def _article_out(row: KBArticleRow) -> KBArticleOut:
)
def _normalize_intent_code(value: str | None) -> str | None:
return str(value or "").strip().upper() or None
def _article_group_expr():
return func.coalesce(KBArticleRow.article_group_id, KBArticleRow.article_id)
@@ -133,6 +138,7 @@ def create_article(
article_id=article_id,
category_id=payload.category_id,
article_group_id=article_group_id,
intent_code=_normalize_intent_code(payload.intent_code),
language=language,
title=payload.title,
body=payload.body,
@@ -196,6 +202,8 @@ def update_article(
if "article_group_id" in data:
row.article_group_id = target_group_id
if "intent_code" in data:
row.intent_code = _normalize_intent_code(data["intent_code"])
if "language" in data:
row.language = target_language
if "title" in data:
+39
View File
@@ -0,0 +1,39 @@
from __future__ import annotations
from typing import Iterable
# Fixed, code-level conversational signals — not FAQ topics. These strings are
# already used as intent literals across ai_orchestrator_service/app.py,
# voice.py, and asserted directly in tests; centralized here rather than
# renamed so every call site validates against the same set.
CONTROL_INTENTS: frozenset[str] = frozenset(
{
"identity_question",
"handoff_request",
"sensitive_request",
"resolution_confirmed",
"clarification",
"kb_answer",
"unknown",
}
)
UNKNOWN_INTENT = "unknown"
def normalize_intent(raw: str | None, *, known_topic_codes: Iterable[str] = ()) -> str:
"""Validate a model-produced intent against control intents and KB topic codes.
`known_topic_codes` are the `intent_code` values of the KB articles actually
shown to the model for this turn anything else the model invents collapses
to UNKNOWN_INTENT rather than being trusted verbatim.
"""
candidate = str(raw or "").strip()
if not candidate:
return UNKNOWN_INTENT
if candidate in CONTROL_INTENTS:
return candidate
normalized_topic_codes = {str(code or "").strip().upper() for code in known_topic_codes if str(code or "").strip()}
if candidate.upper() in normalized_topic_codes:
return candidate.upper()
return UNKNOWN_INTENT
+1
View File
@@ -25,6 +25,7 @@ class KBSearchRow(Protocol):
title: str
body: str
tags_json: str
intent_code: str | None
T = TypeVar("T", bound=KBSearchRow)
+8
View File
@@ -363,6 +363,12 @@ class VoiceEventIn(BaseModel):
"recording.ready",
"call.connected",
"call.transferred",
"AgentReserved",
"AgentRinging",
"AgentNoAnswer",
"AgentConnected",
"TransferCompleted",
"TransferFailed",
]
call_id: str
interaction_id: str | None = None
@@ -983,6 +989,7 @@ class KBCategoryOut(KBCategoryCreate):
class KBArticleCreate(BaseModel):
category_id: str
article_group_id: str | None = None
intent_code: str | None = None
language: str = "ru"
title: str
body: str
@@ -991,6 +998,7 @@ class KBArticleCreate(BaseModel):
class KBArticleUpdate(BaseModel):
article_group_id: str | None = None
intent_code: str | None = None
language: str | None = None
title: str | None = None
body: str | None = None
+8
View File
@@ -583,6 +583,7 @@ def _apply_runtime_schema_compatibility() -> None:
if "kb_articles" in table_names:
columns = _table_columns(inspector, "kb_articles")
_add_column_if_missing(conn, columns, "kb_articles", "article_group_id", "VARCHAR(64)")
_add_column_if_missing(conn, columns, "kb_articles", "intent_code", "VARCHAR(64)")
_add_column_if_missing(conn, columns, "kb_articles", "language", "VARCHAR(8) DEFAULT 'ru'")
if "language" in columns:
conn.execute(
@@ -619,6 +620,13 @@ def _apply_runtime_schema_compatibility() -> None:
"ON kb_articles(language)"
)
)
if "idx_kb_articles_intent_code" not in indexes:
conn.execute(
text(
"CREATE INDEX IF NOT EXISTS idx_kb_articles_intent_code "
"ON kb_articles(intent_code)"
)
)
if "sales_automation_tasks" in table_names:
columns = _table_columns(inspector, "sales_automation_tasks")
+1
View File
@@ -700,6 +700,7 @@ class KBArticleRow(Base):
article_id: Mapped[str] = mapped_column(String(64), unique=True, index=True)
category_id: Mapped[str] = mapped_column(String(64), index=True)
article_group_id: Mapped[str | None] = mapped_column(String(64), nullable=True, index=True)
intent_code: Mapped[str | None] = mapped_column(String(64), nullable=True, index=True)
language: Mapped[str] = mapped_column(String(8), index=True, default="ru")
title: Mapped[str] = mapped_column(String(512), index=True)
body: Mapped[str] = mapped_column(Text)
+84
View File
@@ -414,6 +414,7 @@ def seed_kb_article(
*,
language: str = "ru",
article_group_id: str | None = None,
intent_code: str | None = None,
) -> dict[str, str]:
session = get_session()
try:
@@ -434,6 +435,7 @@ def seed_kb_article(
article_id=article_id,
category_id=category_id,
article_group_id=resolved_group_id,
intent_code=intent_code,
language=language,
title=title,
body=body,
@@ -1347,6 +1349,88 @@ def test_voice_llm_guarded_decision_uses_operator_style_without_ai_or_kb(monkeyp
assert "Do not say or imply that you are an AI" in system_prompt
def test_voice_llm_decision_echoes_kb_article_intent_code(monkeypatch):
monkeypatch.setenv("AI_VOICE_POLICY_MODE", "llm_guarded")
monkeypatch.setattr(ai_module, "_ai_provider", lambda: "openai_compatible")
def _fake_structured(messages, **kwargs):
del messages
return {
"language": "ru",
"intent": "voucher_activation",
"reply_text": "Подтвердите СМС с номера 1414 командой 21*1, затем завершите активацию в eGov.",
"confidence": 0.9,
"needs_handoff": False,
"handoff_reason": None,
"case_action": "keep_open",
"kb_refs": ["kba_voucher_1"],
"_model": "gpt-test",
"_latency_ms": 30,
"_finish_reason": "stop",
}
monkeypatch.setattr(ai_module, "_request_structured_model_decision", _fake_structured)
kb_article = SimpleNamespace(
article_id="kba_voucher_1",
title="Активация ваучера",
body="Подтвердите СМС 1414 командой 21*1, затем перейдите по ссылке и завершите в eGov Mobile.",
intent_code="VOUCHER_ACTIVATION",
)
decision = voice_module._voice_decision(
language="ru",
customer=None,
interaction=SimpleNamespace(interaction_id="int_voice_voucher", customer_id=None, status="new", queue_id="que_voice", subject="voucher"),
transcript_text="Что делать с СМС от 1414?",
transcript_window=[SimpleNamespace(speaker="caller", text="Что делать с СМС от 1414?", sequence_no=1, source_type="voice_asr", barge_in_interrupted=False, created_at=utc_now_iso())],
kb_results=[kb_article],
disclosure_required=False,
)
assert decision["intent"] == "VOUCHER_ACTIVATION"
def test_voice_llm_decision_rejects_invented_intent_not_in_kb_results(monkeypatch):
monkeypatch.setenv("AI_VOICE_POLICY_MODE", "llm_guarded")
monkeypatch.setattr(ai_module, "_ai_provider", lambda: "openai_compatible")
def _fake_structured(messages, **kwargs):
del messages
return {
"language": "ru",
"intent": "totally_made_up_intent",
"reply_text": "Подтвердите СМС с номера 1414 командой 21*1.",
"confidence": 0.9,
"needs_handoff": False,
"handoff_reason": None,
"case_action": "keep_open",
"kb_refs": ["kba_voucher_2"],
"_model": "gpt-test",
"_latency_ms": 30,
"_finish_reason": "stop",
}
monkeypatch.setattr(ai_module, "_request_structured_model_decision", _fake_structured)
kb_article = SimpleNamespace(
article_id="kba_voucher_2",
title="Активация ваучера",
body="Подтвердите СМС 1414 командой 21*1.",
intent_code="VOUCHER_ACTIVATION",
)
decision = voice_module._voice_decision(
language="ru",
customer=None,
interaction=SimpleNamespace(interaction_id="int_voice_voucher_2", customer_id=None, status="new", queue_id="que_voice", subject="voucher"),
transcript_text="Куда отправлять 21*1?",
transcript_window=[SimpleNamespace(speaker="caller", text="Куда отправлять 21*1?", sequence_no=1, source_type="voice_asr", barge_in_interrupted=False, created_at=utc_now_iso())],
kb_results=[kb_article],
disclosure_required=False,
)
assert decision["intent"] == "unknown"
def test_voice_v2_fast_conversational_adds_ack_metadata_and_compacts_reply(monkeypatch):
monkeypatch.setenv("AI_VOICE_POLICY_MODE", "v2_fast_conversational")
monkeypatch.setattr(ai_module, "_ai_provider", lambda: "openai_compatible")
+34
View File
@@ -0,0 +1,34 @@
from services.shared.intents import CONTROL_INTENTS, UNKNOWN_INTENT, normalize_intent
def test_control_intent_passes_through_unchanged():
assert normalize_intent("handoff_request") == "handoff_request"
assert normalize_intent("kb_answer", known_topic_codes=["VOUCHER_ACTIVATION"]) == "kb_answer"
def test_matching_topic_code_passes_through_case_insensitively():
assert normalize_intent("voucher_activation", known_topic_codes=["VOUCHER_ACTIVATION"]) == "VOUCHER_ACTIVATION"
assert normalize_intent(" Voucher_Activation ", known_topic_codes=["voucher_activation"]) == "VOUCHER_ACTIVATION"
def test_unknown_topic_code_falls_back_to_unknown():
assert normalize_intent("made_up_intent", known_topic_codes=["VOUCHER_ACTIVATION"]) == UNKNOWN_INTENT
assert normalize_intent("voucher_activation", known_topic_codes=[]) == UNKNOWN_INTENT
def test_empty_or_missing_intent_falls_back_to_unknown():
assert normalize_intent(None) == UNKNOWN_INTENT
assert normalize_intent("") == UNKNOWN_INTENT
assert normalize_intent(" ") == UNKNOWN_INTENT
def test_control_intents_frozenset_matches_documented_values():
assert CONTROL_INTENTS == {
"identity_question",
"handoff_request",
"sensitive_request",
"resolution_confirmed",
"clarification",
"kb_answer",
"unknown",
}
+40
View File
@@ -40,6 +40,46 @@ def test_kb_lite_search():
assert len(search.json()) >= 1
def test_kb_article_intent_code_round_trips_through_create_get_update():
client = TestClient(kb_app)
headers = {"X-User": "analyst", "X-Role": "analyst"}
cat = client.post(
"/knowledge/categories",
json={"name": "Vouchers", "description": "Voucher help"},
headers=headers,
)
assert cat.status_code == 200
category_id = cat.json()["category_id"]
created = client.post(
"/knowledge/articles",
json={
"category_id": category_id,
"title": "Активация ваучера",
"body": "Подтвердите СМС 1414 командой 21*1.",
"tags": ["voucher"],
"intent_code": "voucher_activation",
},
headers=headers,
)
assert created.status_code == 200
assert created.json()["intent_code"] == "VOUCHER_ACTIVATION"
article_id = created.json()["article_id"]
fetched = client.get(f"/knowledge/articles/{article_id}")
assert fetched.status_code == 200
assert fetched.json()["intent_code"] == "VOUCHER_ACTIVATION"
updated = client.patch(
f"/knowledge/articles/{article_id}",
json={"intent_code": "voucher_activation_v2"},
headers=headers,
)
assert updated.status_code == 200
assert updated.json()["intent_code"] == "VOUCHER_ACTIVATION_V2"
def test_kb_lite_search_ranks_title_over_body_only_matches():
client = TestClient(kb_app)
headers = {"X-User": "analyst", "X-Role": "analyst"}