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.
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
Phase 2 of the L1->L2 routing engine (Phase 1: MR!4).
- ami_loop() now also captures native AMI DialEnd/Hangup frames (not
only UserEvent), needed to detect that an escalated agent did not
answer. No dialplan change required - Redirect already routes the
client channel into an existing Dial()-based transfer context, so
Asterisk emits these events on its own; the listener just wasn't
reading them before.
- retry_escalation_no_answer(): on NOANSWER/BUSY/CANCEL/CHANUNAVAIL/
CONGESTION, releases the non-answering agent, excludes it, and
reserves+redirects to the next available agent via the routing
engine's existing exclude_agent_ids support. Exhausted pool marks
the escalation failed and leaves the call with the AI instead of
dropping the client (ТЗ §32).
- Agent status now actually moves through
RESERVED -> RINGING -> TALKING -> AFTER_CALL_WORK -> AVAILABLE
instead of staying stuck on RESERVED for the whole call; a new
acw_sweep_loop background thread (same pattern as the existing
failed_retry_loop) times out AFTER_CALL_WORK back to AVAILABLE.
- escalations gains attempt_count/real_agent_id/attempted_agent_ids_json
(migration 0033); fixes a latent bug where assigned_agent_id stored
the SIP extension instead of the real agent_id despite routing-service
already returning it in RoutingAgentReserveOut.
- Every transition now records an interaction timeline entry and
publishes the ТЗ §25 event catalog (AgentReserved/AgentRinging/
AgentNoAnswer/AgentConnected/TransferCompleted/TransferFailed)
through the existing emit_voice_event/EventOutboxRow idempotent path.
Not in this MR (see plan): SLA config, Callback, L3 (needs real
technical agents from the business), metrics.
Replaces the hardcoded single-extension redirect for AI->human call
escalation with a real Agent Pool + Routing Engine:
- agents/escalations/routing_rules tables (migration 0031), asterisk_call_links
gains tenant_id/current_level/required_skills_json/priority.
- services/routing_service/engine.py: level/tenant/skill filtered agent
selection with atomic (CAS) reservation, no double-booking.
- routing-service: /agents CRUD + /internal/routing/reserve-agent and
/internal/routing/release-agent.
- asterisk-bridge-service: voice_ai.request_handoff now uses the Routing
Engine automatically for any queue_code configured in
ASTERISK_QUEUE_LEVEL_MAP_JSON (all other queue_codes keep the existing
static ASTERISK_TRANSFER_TARGET_MAP_JSON behavior unchanged); new
POST /asterisk/live-calls/{call_id}/escalations entrypoint; agent is
released back to AVAILABLE and the escalation closed when the call ends.
Targets the Tele2 Kazgaz DID +77476456048 (from-tele2-kazgaz context) as the
first queue wired to real L2 routing instead of AI-only.
Known gap (documented in docs/architecture/l1-l2-routing-engine.md):
automatic no-answer retry-to-next-agent needs a small, separately reviewed
dialplan change and is left for a follow-up MR rather than guessed at blind.
Tests: services/routing_service/engine.py covered by
tests/test_routing_engine.py (selection filtering, atomic reservation,
release); existing test_asterisk_bridge_service.py and
test_routing_service_pg_counter.py suites still pass unmodified.
Two independent latency fixes for the voice-assistant reply pipeline,
both scoped to the parts of the flow that run regardless of whether
voice_v2 is enabled for a queue:
1. media_runtime._process_utterance: the v1/fallback turn path (used by
any queue not covered by AI_VOICE_V2_QUEUE_CODES) silently awaited the
full LLM decision with no audio playing at all, unlike the v2 path
which already has a decision-timeout ack. Give v1 the same behavior:
wait up to 600ms (_v1_ack_wait_seconds) for the decision, and if it's
still not ready, play a short "Секунду." filler via the existing
_emit_early_ack before the real reply, instead of leaving the caller
in silence for the full LLM+TTS round trip. Reuses the same ack
selection/playback code path v2 already exercises, so no new failure
modes - just an added timeout branch mirroring the existing v2 one.
2. ai_orchestrator_service._kb_search: every voice/chat turn re-ran a
full-table scan of kb_articles (all columns, including body text) and
rescored every row in Python, even though the KB rarely changes
mid-conversation. Added an in-process cache keyed by language, gated
on a cheap content fingerprint (row count + max id + max updated_at +
summed title/body/tags length, all computed server-side without
transferring the text columns). A fingerprint mismatch always
triggers a fresh fetch, so this can never serve stale results after
an insert/update/delete - unlike a naive TTL cache, which would have
been be wrong the moment a test (or a real KB edit) changed the table
within the cache window.
Note the first fingerprint design (count + max id + max updated_at
only) was insufficient: utc_now_iso() truncates to whole seconds and
SQLite reuses primary keys after a full-table delete, so two
different row sets written in the same wall-clock second could share
a fingerprint. Caught this via a real test failure
(test_ai_whatsapp_relaxed_kb_search_answers_phrase_query breaking
only when run after test_ai_orchestrator_service.py in the same
process) before it could reach production; the summed content-length
term closes the gap.
Added test_media_runtime_plays_filler_ack_when_v1_decision_is_slow
(asserts greeting -> ack -> reply delivery order when process_turn is
slow) and verified the KB cache against the full
test_ai_orchestrator_service.py + test_ai_whatsapp_orchestrator_service.py
suite plus a wider kb/orchestrator/whatsapp/telegram/voice-filtered run:
only the same pre-existing, already-documented failures remain (unrelated
sales_service test-isolation ordering, one known persona-prompt
assertion) - no new failures from either change.
Streaming the LLM decision itself (start speaking reply_text before the
full structured JSON response finishes generating) was scoped but
deliberately deferred: it needs incremental JSON parsing on top of SSE
streaming to detect when just the reply_text field is complete, shared
across both voice and text-channel decision paths - a separate,
higher-risk change that deserves its own PR and testing pass rather than
being bundled here.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
The AudioSocket/media_runtime playback pipeline already supports chunked
TTS streaming (voice_v2_streaming_tts), but every provider inherited the
base TTSProvider.synthesize_chunks(), which just called the blocking
synthesize() and yielded the entire finished audio as a single "chunk" -
so the caller waited for full-utterance synthesis before any playback
could start regardless of the flag.
ElevenLabs is the production default (AI_VOICE_TTS_PROVIDER=elevenlabs in
deployment/docker-compose.server.yml), so give it a real implementation
that POSTs to the /stream endpoint and yields audio as network chunks
arrive, instead of waiting for the whole response body. Chunk boundaries
are re-aligned to whole 16-bit PCM samples so a split sample at a network
read boundary can't corrupt playback. The full synthesized audio is still
written to the on-disk cache afterwards so repeat phrases stay fast and
skip the vendor call entirely, matching the existing synthesize() cache
behavior.
Added test_elevenlabs_tts_provider_streams_chunks_and_caches_full_audio to
cover: chunk splitting mid-sample gets re-aligned, all yielded chunks are
sample-aligned, the full audio round-trips through the cache, and a
cached synthesis is replayed without invoking the streaming endpoint
again.
Verified via tests/test_ai_voice_tts_provider.py (9/9 pass) and a wider
voice/tts-filtered run across the suite: the only failures present are
the same pre-existing, already-documented ones (sales_service test
cross-file isolation ordering, one known persona-prompt assertion) -
identical set to before this change, no new failures.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
didar
arys
arystanbek