config/ai_features was not the switch it was documented to be. Three paths
spent money with it off, and one path read it wrong, so per-system opt-outs
did not opt anything out.
- Correlation in the ingest pipeline tested the raw global flag instead of the
per-system resolution. With a system opted out, extraction was skipped but
the no-scenes fallback still correlated the call with empty tags, taking the
thin/recency path and attaching it to whatever incident was most recent on
that system. The opt-out did not disable correlation, it disabled good
correlation and left the worst kind running. (#75)
- Transcript correction ran on every transcribed call gated only by an env var,
spending Gemini tokens and a Places lookup per proposed location. An
"STT-only" window was never STT-only and its cost could not be attributed.
Now behind transcript_correction_enabled. (#76)
- _run_extraction_pipeline and the vocabulary learner, both reachable from
PATCH /calls/{id}/transcript, checked no flags at all. (#76, #81)
The flag resolver now lives in feature_flags.resolve_flags() rather than as a
local helper in upload.py. Three copies of that logic is how #75 happened.
PATCH /calls/{id}/transcript now refuses with 409 when correlation is off.
That route wipes tags, severity, location, units, embedding and unlinks the
call from every incident before queueing re-extraction. Gating extraction
alone would have made it destructive-only in the standing flags-off
configuration: the call left blank and orphaned forever, with the route still
answering 200. The wipe and the rebuild are one transaction in intent, so it
refuses before the first write.
Also: the summarizer's stale-incident sweep is no longer behind
summaries_enabled. It is pure Firestore with no model call in it, and gating
it meant nothing auto-resolved while AI was off - so every incident stayed
active forever and the candidate set every correlation reads kept growing.
transcript_correction_enabled is documented as NOT a pure cost lever. The
corrector is also the noise gate that sets not_speech; with it off, recogniser
noise reaches extraction as a real transcript, comes back thin, and
auto-attaches. Never open an evaluation window with correction off and
correlation on.
14 tests added covering flag precedence, both pipeline paths, the 409, the
correction gate and the summarizer no-op. Suite: 264 passed.
Refs #75, #76, #81, #45.
Correction existed, but as a line in intelligence.py's EXTRACTION_PROMPT --
which put it in the wrong place twice over. The same model call that extracted
units, location and severity emitted the correction afterwards, so extraction
reasoned over text already known to be wrong; and it sat behind
correlation_enabled, so during a cost-controlled STT-only window nothing was
ever corrected at all. That is the normal state during development.
internal/transcript_correction.py is now its own pass, between the degenerate
filter and the Firestore write. It receives an already-produced transcript plus
a reference list, so unlike a Whisper prompt it has no series to extend -- the
distinction that keeps vocabulary out of the recogniser's prompt, where an
enumerated ten-code list once made it hallucinate ten-code runs.
Reference data is merged from the talkgroup and the system, TALKGROUP FIRST. A
system spanning several counties can have a talkgroup covering one
municipality, and that municipality's streets must not be buried under a
county-wide list. A single-municipality system is the degenerate case: populate
the system level and every talkgroup inherits it. Area context is now SET --
municipality, county, roads, landmarks, on both scopes -- rather than guessed
from talkgroup names, which is what vocabulary_learner did and which is close
to useless across multiple counties.
Segments are corrected too, not just the joined text. extract_scenes builds its
prompt from numbered segments whenever there is more than one, so a correction
that only fixed the transcript would have been discarded on exactly the
multi-transmission calls carrying the most content. Alignment is enforced: an
array of the wrong length or type is dropped whole, because scenes map back to
transmissions by index and a shifted array would misattribute audio silently.
Whisper is also retried once on degenerate output. Call e49ea32c produced a
56-word ten-code counting run on one attempt and ordinary speech on the next --
same clip, same temperature=0 -- so a hallucination is a coin-flip, and
discarding on the first bad roll threw away a recoverable transcript.
Two things found on the way:
PUT /systems/{id} wiped ten_codes on every save. The systems form sends only
{name, type, config}, and model_dump() wrote every omitted field as its default
over the top. Now exclude_unset. area_context would have been the next victim,
which is why it gets its own route alongside ten-codes rather than a field on
that payload.
Closes server-26#36.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Three AI dependency failures in one night (retired Gemini model IDs,
depleted Gemini balance, unpayable OpenAI account) each surfaced only
as a single ERROR log line that nobody was watching. Add
app/internal/ai_health.py, a shared in-memory registry that
transcription.py and llm_correlator.py report into on every call
(success and failure), distinguishing permanent conditions (dead
model, dead billing) which alert immediately from transient ones
(rate limits, network blips) which only alert after they persist.
Alerts POST once per degradation episode and once on recovery to an
optional Discord webhook (AI_ALERT_WEBHOOK_URL), reusing alerter.py's
httpx pattern. State is exposed unauthenticated at GET /health/ai
alongside the existing /health.
Closeslogan/server-26#14
Transcription is the top of the pipeline and it fails soft: any exception logs
a WARNING, returns None, and upload.py carries on. That is the right behaviour
for a network blip and exactly the wrong behaviour for an unpayable account,
because with no transcript there is no extraction, no correlation and no
incident -- the system keeps accepting calls and quietly stores empty ones,
which looks like quiet radio traffic rather than an outage.
This is the third instance of the same failure mode today. The Gemini
correlator was down first on a retired model ID and then on a depleted
balance, and in both cases the only signal was a per-call WARNING that read as
noise. The OpenAI balance is low enough that this one is a matter of when.
Billing-shaped errors (insufficient_quota, billing, credit, quota exceeded)
now log once at ERROR, name what is dead downstream, and link the top-up page.
Everything else keeps the existing per-call WARNING.
No new environment variables, so CI deploys this without an ansible run.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Two independent sources of garbage in the AI pipeline, both visible in the
2026-08-16 correlation dump.
1. Hallucinated transcripts. The Whisper prompt opened with an enumerated run
of ten-codes: 10-4, 10-23, 10-20, 10-97 and so on. Whisper treats prompt
text as preceding transcript, so on noisy or silent audio it continued the
series, emitting transcripts that count upward from 10-4 to 10-99. The
existing no_speech_prob filter could not catch these: the model is highly
confident in text it invented by continuing a pattern.
The prompt no longer contains a series to extend, and _is_degenerate()
rejects the three shapes this failure takes: ascending ten-code runs, one
phrase looping, and near-identical segments across a whole recording.
Verified against 13 transcripts from production: all four known
hallucinations rejected, all nine real ones kept, including terse traffic
containing legitimate codes.
2. Duplicate recordings. node-002 and node-PI-2 both cover TG 9048 and both
uploaded the same transmissions, ~1.1s apart. Nine pairs appeared in one
dump. Each was transcribed, billed and correlated twice, and the resulting
incident listed two units where there was one.
Canonical selection is by earliest started_at, tie-broken on call_id, NOT
by upload order: upload order varies with encode time and network latency,
so it would make the authoritative recording non-deterministic. Call
documents are created from MQTT call_start before uploads arrive, so both
nodes independently reach the same verdict. The loser keeps its audio (it
may be the cleaner capture) but is excluded from STT, correlation, the
re-correlation sweep and the orphan debug view.
Also fixes _sync_transcribe returning a bare None when OPENAI_API_KEY is
missing, where the caller unpacks two values. A missing key surfaced as a
misleading "Transcription failed" instead of the real warning.
Adds tests/test_dedup.py (15 cases). dedup.py reaches Firestore through an
injected callable so it stays importable without firebase-admin present.
- correlator: unit_overlap on dispatch channels now applies content
divergence check when the call has geocoded coords but the incident
doesn't; previously this gap caused unrelated calls to merge into
stale incidents (e.g. patrol officer at a second scene 70 min later)
- STT: switch default model from gpt-4o-transcribe to whisper-1, which
faithfully transcribes all exchanges in multi-PTT recordings; gpt-4o
was silently dropping utterances, starving the correlation engine
- STT: remove vocabulary from the Whisper prompt; whisper-1 echoes
prompted terms into noise/silence, skewing extracted incident data;
vocabulary context is now applied exclusively in the GPT extraction
step (build_gpt_vocab_block) where it is used as reference only
Upload 404 warning doc_set(merge=True) in upload.py — creates doc if missing
MQTT call_end 404 error doc_set(merge=True) in mqtt_handler.py — same root cause
Transcription 404 (saving transcript to nonexistent doc) doc_set(merge=True) in transcription.py
Transcription ADC credentials error Explicit service_account.Credentials from gcp-key.json in _sync_transcribe — same pattern as storage.py
Logan Cusano
