Compare commits
12
Commits
| Author | SHA1 | Date | |
|---|---|---|---|
|
|
9f19750ea6 | ||
|
|
433b35d2ba | ||
|
|
badfe28823 | ||
|
|
737bdf0576 | ||
|
|
b9e7524817 | ||
|
|
e972cace4a | ||
|
|
969d175a67 | ||
|
|
731b54bed9 | ||
|
|
3c642e2946 | ||
|
|
f0a88d401c | ||
|
|
8eac32caf5 | ||
|
|
6e82ee8579 |
@@ -89,7 +89,17 @@ class Settings(BaseSettings):
|
||||
embedding_cross_tg_threshold: float = 0.85 # cross-TG path: same dept + 2+ shared units
|
||||
location_proximity_km: float = 0.5 # radius for location-proximity matching
|
||||
geocode_max_km: float = 40.0 # reject geocode results farther than this from the node
|
||||
incident_auto_resolve_minutes: int = 90 # auto-resolve after N minutes with no new calls
|
||||
incident_auto_resolve_minutes: int = 90 # auto-resolve after N minutes with no new calls (major / unknown severity)
|
||||
# Most jobs never say 10-8 on the air (replay of 09-22, server-26#170: ~5 of
|
||||
# ~25 real incidents had an audible clear), so the quiet timer IS the close
|
||||
# for most of them, and one 90-minute timer kept a lockout or a plate check
|
||||
# "active" on the portal an hour after it ended. Scaled by the incident's
|
||||
# severity instead, and made provisional: a timer-closed incident stays
|
||||
# reopenable for incident_reopen_window_minutes, so a long quiet job that
|
||||
# comes back on the air rejoins its own incident rather than splitting.
|
||||
incident_auto_resolve_minutes_routine: int = 30 # routine / minor
|
||||
incident_auto_resolve_minutes_moderate: int = 60
|
||||
incident_reopen_window_minutes: int = 90 # since last substantive call
|
||||
unit_continuity_max_idle_minutes: int = 20 # unit-continuity path: skip if incident idle > this
|
||||
recorrelation_scan_minutes: int = 60 # re-examine orphaned calls ended within this window
|
||||
tg_fast_path_idle_minutes: int = 90 # fast path: max minutes since incident last updated
|
||||
|
||||
@@ -224,6 +224,38 @@ def _normalize_unit(unit: str) -> str:
|
||||
return key or unit.strip().lower()
|
||||
|
||||
|
||||
def _after_close(inc: dict, now: datetime) -> bool:
|
||||
try:
|
||||
closed = datetime.fromisoformat(str(inc.get("resolved_at") or "").replace("Z", "+00:00"))
|
||||
except ValueError:
|
||||
return True
|
||||
if closed.tzinfo is None:
|
||||
closed = closed.replace(tzinfo=timezone.utc)
|
||||
return now > closed
|
||||
|
||||
|
||||
def _is_trackable_unit(unit: str) -> bool:
|
||||
"""
|
||||
Whether a unit is concrete enough to hold an incident open until it clears.
|
||||
|
||||
Extraction lists everything that sounds like a unit — "Desk", "Central",
|
||||
"Division", "sergeant", "unknown", and plate phonetics ("John Henry
|
||||
Zebra"). None of those ever transmit a 10-8, so while they sat in
|
||||
units_active the all-clear gate below could never pass: in the first
|
||||
replay (server-26#170, 09-22 10:00-12:00) 0 of 19 incidents resolved on
|
||||
a clear and every one had such a name in units_active. A real radio unit
|
||||
ID carries a number ("45-9", "11-Adam 2", "Whitestone 1", "E-14"), so
|
||||
only those gate resolution. The others are still kept in `units` and
|
||||
still match for correlation.
|
||||
"""
|
||||
if _TEN_CODE_RE.match((unit or "").strip()):
|
||||
return False # "10-8" read back as a unit ID is the status, not a unit
|
||||
return any(ch.isdigit() for ch in unit or "")
|
||||
|
||||
|
||||
_TEN_CODE_RE = re.compile(r"^10[\s-]?\d{1,2}$")
|
||||
|
||||
|
||||
def _unit_keys(units: Optional[list[str]]) -> set[str]:
|
||||
"""Comparison keys for a unit list, empties dropped."""
|
||||
return {k for k in (_normalize_unit(u) for u in (units or [])) if k}
|
||||
@@ -596,7 +628,14 @@ def _incident_at_capacity(inc: dict, now: datetime) -> Optional[str]:
|
||||
it has and still auto-resolves on the normal idle sweep. It just stops
|
||||
being a candidate, so the next call opens a fresh incident.
|
||||
"""
|
||||
call_count = len(inc.get("call_ids") or [])
|
||||
# Content-free replies ("10-4", "Cut.", "Copy") are what filled the cap:
|
||||
# the 09-22 bridge MVA hit 40 calls in 32 minutes with ~40% of them thin,
|
||||
# split in half, and its second half took a different job's title. Only
|
||||
# substantive calls count; incidents written before this field existed
|
||||
# fall back to the raw count.
|
||||
call_count = inc.get("substantive_call_count")
|
||||
if call_count is None:
|
||||
call_count = len(inc.get("call_ids") or [])
|
||||
if call_count >= settings.incident_max_calls:
|
||||
return f"call_cap:{call_count}"
|
||||
span = _incident_span_minutes(inc, now)
|
||||
@@ -829,8 +868,16 @@ async def _build_context(
|
||||
# the whole collection rather than being unable to correlate at all.
|
||||
if org_id is not None:
|
||||
all_active = await fstore.collection_list("incidents", status="active", org_id=org_id)
|
||||
reopenable = await fstore.collection_list("incidents", status="resolved", reopenable=True, org_id=org_id)
|
||||
else:
|
||||
all_active = await fstore.collection_list("incidents", status="active")
|
||||
reopenable = await fstore.collection_list("incidents", status="resolved", reopenable=True)
|
||||
# A timer-closed incident is provisional (summarizer._resolve_stale_incidents):
|
||||
# inside its reopen window it is still a candidate, and linking a call to
|
||||
# it reopens it (_update_incident). The fast path's own recency gate
|
||||
# (tg_fast_path_idle_minutes) still applies to it like any other candidate.
|
||||
reopen_window = timedelta(minutes=settings.incident_reopen_window_minutes)
|
||||
all_active += [inc for inc in reopenable if _idle_gate_minutes(inc, now) <= reopen_window.total_seconds() / 60]
|
||||
# Incidents past the hard caps are removed from the candidate pool here, so
|
||||
# neither the rules engine nor the LLM tier (which reads ctx["recent"] /
|
||||
# ctx["all_active"]) can propose linking into one.
|
||||
@@ -1910,11 +1957,19 @@ def _apply_unit_clearance(inc: dict, cleared: list[str]) -> tuple[list[str], lis
|
||||
"""
|
||||
units_active = list(inc.get("units_active") or [])
|
||||
units_cleared = list(inc.get("units_cleared") or [])
|
||||
for u in cleared:
|
||||
if u in units_active:
|
||||
units_active.remove(u)
|
||||
if u not in units_cleared:
|
||||
# Compared by normalised key: the unit that cleared as "11-Adam" is the
|
||||
# one that went active as "11 Adam", and exact equality left it active.
|
||||
# Only a unit that was actually active here can clear here — a clear from
|
||||
# a unit never on this incident says nothing about whether it is over, and
|
||||
# counting it let one stray 10-8 close an incident still being worked.
|
||||
cleared_keys = _unit_keys(cleared)
|
||||
releasing = [u for u in units_active if _normalize_unit(u) in cleared_keys]
|
||||
units_active = [u for u in units_active if _normalize_unit(u) not in cleared_keys]
|
||||
known_cleared = _unit_keys(units_cleared)
|
||||
for u in releasing:
|
||||
if _normalize_unit(u) not in known_cleared:
|
||||
units_cleared.append(u)
|
||||
known_cleared.add(_normalize_unit(u))
|
||||
auto_resolved = bool(units_cleared) and not units_active
|
||||
return units_active, units_cleared, auto_resolved
|
||||
|
||||
@@ -1984,7 +2039,8 @@ async def _update_incident(
|
||||
incident_id = inc["incident_id"]
|
||||
|
||||
call_ids = list(inc.get("call_ids") or [])
|
||||
if call_id not in call_ids:
|
||||
is_new_call = call_id not in call_ids
|
||||
if is_new_call:
|
||||
call_ids.append(call_id)
|
||||
|
||||
talkgroup_ids = list(inc.get("talkgroup_ids") or [])
|
||||
@@ -2009,9 +2065,11 @@ async def _update_incident(
|
||||
# units_active = units currently on scene; units_cleared = units back in service
|
||||
units_active = list(inc.get("units_active") or [])
|
||||
units_cleared = list(inc.get("units_cleared") or [])
|
||||
tracked = _unit_keys(units_active) | _unit_keys(units_cleared)
|
||||
for u in call_units:
|
||||
if u not in units_cleared and u not in units_active:
|
||||
if _is_trackable_unit(u) and _normalize_unit(u) not in tracked:
|
||||
units_active.append(u)
|
||||
tracked.add(_normalize_unit(u))
|
||||
inc_with_active_update = {**inc, "units_active": units_active, "units_cleared": units_cleared}
|
||||
units_active, units_cleared, _ = _apply_unit_clearance(inc_with_active_update, cleared_units or [])
|
||||
|
||||
@@ -2054,6 +2112,11 @@ async def _update_incident(
|
||||
# thin traffic rides along without extending its life.
|
||||
if refresh_activity:
|
||||
updates["updated_at"] = _floor_at_started_at(inc, now).isoformat()
|
||||
if is_new_call: # a second scene of the same call is not a second call
|
||||
updates["substantive_call_count"] = (
|
||||
inc.get("substantive_call_count")
|
||||
if inc.get("substantive_call_count") is not None else len(inc.get("call_ids") or [])
|
||||
) + 1
|
||||
else:
|
||||
updates["last_thin_at"] = now.isoformat()
|
||||
# Update incident type when a re-classified call provides a concrete type.
|
||||
@@ -2071,6 +2134,15 @@ async def _update_incident(
|
||||
# Signal-based auto-resolve: every tracked unit has cleared, none still active.
|
||||
# Requires at least one unit to have explicitly signalled back-in-service so we
|
||||
# don't fire on incidents where units were never tracked (no unit mentions at all).
|
||||
if inc.get("status") == "resolved" and refresh_activity and _after_close(inc, now):
|
||||
# A timer close was provisional and a related, substantive call arrived
|
||||
# after it. A thin "10-4" rides along without reopening (it would not
|
||||
# refresh updated_at, so the next sweep would just close it again),
|
||||
# and neither does a sweep link of a call from before the close.
|
||||
updates.update({"status": "active", "resolved_at": None, "resolved_via": None,
|
||||
"reopenable": False, "reopened_count": (inc.get("reopened_count") or 0) + 1})
|
||||
logger.info(f"Correlator: reopened timer-closed incident {incident_id} (call {call_id})")
|
||||
|
||||
if units_cleared and not units_active:
|
||||
updates["status"] = "resolved"
|
||||
updates["resolved_at"] = now.isoformat()
|
||||
@@ -2139,11 +2211,12 @@ async def _create_incident(
|
||||
**location_fields,
|
||||
"location_mentions": [location] if location else [],
|
||||
"call_ids": [call_id],
|
||||
"substantive_call_count": 1,
|
||||
"talkgroup_ids": [str(talkgroup_id)] if talkgroup_id is not None else [],
|
||||
"system_ids": [system_id] if system_id else [],
|
||||
"tags": tags + ["auto-generated"],
|
||||
"units": call_units,
|
||||
"units_active": list(call_units),
|
||||
"units_active": [u for u in call_units if _is_trackable_unit(u)],
|
||||
"units_cleared": [],
|
||||
"vehicles": call_vehicles,
|
||||
"srcaddrs": [call_srcaddr] if call_srcaddr else [],
|
||||
@@ -2315,6 +2388,10 @@ async def _find_cross_system_parent(
|
||||
best_score = 0.0
|
||||
|
||||
for inc in recent:
|
||||
# A timer-closed incident is in `recent` only so a related call can
|
||||
# reopen it; it must not be adopted as another agency's parent.
|
||||
if inc.get("status") != "active":
|
||||
continue
|
||||
# Only cross-system candidates
|
||||
if system_id in (inc.get("system_ids") or []):
|
||||
continue
|
||||
|
||||
@@ -67,7 +67,7 @@ Rules:
|
||||
- tags: describe WHAT happened, not WHERE. Specific, lowercase, hyphenated. Do not use location names, road names, talkgroup names, or place names as tags (wrong: "lower-macy's", "canvas-route-6", "route-202"; right: "suspect-search", "shoplifting", "vehicle-pursuit"). Do not repeat incident_type as a tag.
|
||||
- units: ONLY identifiers that appear verbatim in the transcript. Use speaker role inference to distinguish units being dispatched from units acknowledging — both should be included. Never infer or guess unit IDs not present in the text. If a unit ID format is given below, use it to recognise a unit spoken in a shortened or partial form (e.g. just the phonetic name alone) as the same unit — but still only extract what is actually said, never fabricate the full form.
|
||||
- Do not invent details not present in the transcript.
|
||||
- incident_type: FIRST decide whether this transmission has any incident behind it at all, using the same bar as the "routine" severity rule below — pure administrative/status traffic with nothing describable happening: post/unit check-ins, roll call, bare acknowledgements ("10-4", "copy", "received"), records/report exchanges, "show me admin"/"show me available", a status ten-code with no event attached. If it is administrative/status-only, return "unknown" — this applies on EVERY channel, including a police channel; do not let the channel default override it (server-26#138: forcing a channel default onto content-free chatter is what let radio housekeeping open incidents). Only once real event content is present, let the talkgroup channel be your primary signal for WHICH type. Use "fire" ONLY if the talkgroup is clearly a fire/rescue channel OR the transcript explicitly describes active fire, smoke, flames, or structure fire activation. Police or EMS referencing a fire scene → use "police" or "ems". When the channel is a police channel, a real event is present, and nothing in the transcript contradicts it, return "police". Reserve "other" for a real event that genuinely belongs to no emergency service (rail operations, public works, utility coordination) — not for administrative chatter, which is "unknown" per above regardless of channel. Also reserve "unknown" for transcripts too garbled to place at all.
|
||||
- incident_type: FIRST decide whether this transmission has any incident behind it at all, using the same bar as the "routine" severity rule below — pure administrative/status traffic with nothing describable happening: post/unit check-ins, roll call, bare acknowledgements ("10-4", "copy", "received"), records/report exchanges, "show me admin"/"show me available", a status ten-code with no event attached. If it is administrative/status-only, return "unknown" — this applies on EVERY channel, including a police channel; do not let the channel default override it (server-26#138: forcing a channel default onto content-free chatter is what let radio housekeeping open incidents). Only once real event content is present, let the talkgroup channel be your primary signal for WHICH type. Use "fire" ONLY if the talkgroup is clearly a fire/rescue channel OR the transcript explicitly describes active fire, smoke, flames, or structure fire activation. Police or EMS referencing a fire scene → use "police" or "ems". When the channel is a police channel, a real event is present, and nothing in the transcript contradicts it, return "police". Reserve "other" for a real event that genuinely belongs to no emergency service (rail operations, public works, utility coordination) — not for administrative chatter, which is "unknown" per above regardless of channel. Also reserve "unknown" for transcripts too garbled to place at all. A unit reporting its OWN activity is a real event, not status traffic: "on a stop" / traffic stop / car stop, "out with a vehicle", "put me out with a pedestrian/subject" — return "police", tag it (e.g. "traffic-stop", "pedestrian-assist"), severity at least "minor". The plate/license lookups for that stop belong to it.
|
||||
- severity: ALWAYS return one of the four values. Judge the underlying event, not how dramatic the words sound.
|
||||
"routine" — administrative/status traffic with no incident behind it: mileage and transport logging, radio checks, acknowledgements, shift changes, track block/power requests, records lookups.
|
||||
"minor" — a real but low-stakes call: lift assist, parking complaint, past-tense larceny report, noise complaint, welfare check.
|
||||
@@ -247,18 +247,34 @@ async def extract_scenes(
|
||||
f"Intelligence: call {call_id} — transcript too short for extraction "
|
||||
f"({len(transcript.split())} words), skipping"
|
||||
)
|
||||
cleared_unit = _short_clearance_unit(transcript)
|
||||
try:
|
||||
# Severity is still recorded: a five-word acknowledgement is genuinely
|
||||
# routine traffic, and downstream code treats a missing severity as
|
||||
# "not yet processed" rather than "nothing happened".
|
||||
await fstore.doc_set("calls", call_id, {
|
||||
updates = {
|
||||
"skip_reason": "transcript_too_short",
|
||||
"severity": "routine",
|
||||
"chatter_classifier_verdict": chatter_is_chatter,
|
||||
"chatter_classifier_reason": chatter_reason,
|
||||
})
|
||||
}
|
||||
if cleared_unit:
|
||||
updates["units"] = [cleared_unit]
|
||||
updates["cleared_units"] = [cleared_unit]
|
||||
await fstore.doc_set("calls", call_id, updates)
|
||||
except Exception:
|
||||
pass
|
||||
if cleared_unit:
|
||||
logger.info(f"Intelligence: call {call_id} — short clearance from {cleared_unit!r}")
|
||||
return [_clearance_scene(transcript, cleared_unit)]
|
||||
# "Adam 3 on a stop" is the whole report of a stop, and it is <=5
|
||||
# words: the self-initiated backstop has to run here too or the most
|
||||
# common phrasing never opens an incident.
|
||||
tags, typ, sev = _self_initiated_backstop(transcript, [], None, "routine", talkgroup_name)
|
||||
if tags:
|
||||
logger.info(f"Intelligence: call {call_id} — short self-initiated report {tags}")
|
||||
return [{**_clearance_scene(transcript, ""), "units": [], "cleared_units": [],
|
||||
"tags": tags, "incident_type": typ, "severity": sev}]
|
||||
return []
|
||||
|
||||
try:
|
||||
@@ -410,6 +426,10 @@ async def extract_scenes(
|
||||
transcript, segments, segment_indices, transcript_corrected
|
||||
)
|
||||
|
||||
tags, incident_type, severity = _self_initiated_backstop(
|
||||
scene_transcript or transcript, tags, incident_type, severity, talkgroup_name,
|
||||
)
|
||||
|
||||
processed.append({
|
||||
"tags": tags,
|
||||
"incident_type": incident_type,
|
||||
@@ -469,6 +489,118 @@ async def extract_scenes(
|
||||
return processed
|
||||
|
||||
|
||||
# Self-initiated activity: a unit putting itself "on a stop" or "out with" a
|
||||
# vehicle/pedestrian. Replay of 09-22 (server-26#170): every traffic stop on
|
||||
# the Ch 1 channel ("45 Adam on a stop, Eastbound Central Express", "CM2 on
|
||||
# the stop, southbound") came back untyped/untagged/routine from extraction —
|
||||
# read as status traffic — so the creation gate never opened an incident and
|
||||
# the stop was visible only in the archive. The prompt now says so too; this
|
||||
# is the deterministic backstop, because a tag is what the creation gate
|
||||
# counts as substance (incident_correlator.has_event_substance).
|
||||
# Only a unit's own "on a stop" self-report — a bare "traffic stop"/"car stop"
|
||||
# mention (a plate lookup on a records channel, a dispatcher's question) is
|
||||
# left to the prompt, and "pull over" is too common in non-stop traffic
|
||||
# ("Medic 2 pull over to the side") to trust (review of 433b35d).
|
||||
_SELF_INITIATED = (
|
||||
(re.compile(r"\bon (a|the) (traffic |car |vehicle |motor vehicle )?stop\b", re.IGNORECASE),
|
||||
"traffic-stop"),
|
||||
(re.compile(r"\b((put|show) me out with|out with (a|one) (pedestrian|vehicle|disabled|male|female|"
|
||||
r"subject|party|juvenile))\b", re.IGNORECASE),
|
||||
"self-initiated"),
|
||||
)
|
||||
_NEGATED = re.compile(r"\b(not|don't|dont|no|never)\s+(\S+\s+){0,5}$", re.IGNORECASE)
|
||||
# "Train 4 holding on the stop", "out with a disabled on the bridge": on rail,
|
||||
# bridge/tunnel, fire and EMS channels these phrases are operations, not a
|
||||
# police stop. Everywhere else — including "Ch 1 (Patched ...)", which is where
|
||||
# the stops actually are — the backstop applies.
|
||||
_NO_BACKSTOP_TG = re.compile(r"\b(mta|rail|railroad|train|transit|bridges? and tunnels|fire|ems|"
|
||||
r"rescue|ambulance|dpw|public works)\b", re.IGNORECASE)
|
||||
|
||||
|
||||
def _self_initiated_backstop(
|
||||
text: str, tags: list, incident_type: Optional[str], severity: str,
|
||||
talkgroup_name: Optional[str] = None,
|
||||
) -> tuple[list, Optional[str], str]:
|
||||
if talkgroup_name and _NO_BACKSTOP_TG.search(talkgroup_name):
|
||||
return tags, incident_type, severity
|
||||
for pattern, tag in _SELF_INITIATED:
|
||||
m = pattern.search(text or "")
|
||||
if not m or _NEGATED.search(text[: m.start()]):
|
||||
continue
|
||||
if tag not in tags:
|
||||
tags = [*tags, tag]
|
||||
incident_type = incident_type or "police"
|
||||
if severity == "routine":
|
||||
severity = "minor"
|
||||
return tags, incident_type, severity
|
||||
|
||||
|
||||
# "45-9, I'm clear." / "Vehicle 1, clear." / "Car 12 10-8" — a unit reporting
|
||||
# itself back in service is the one signal that ends an incident, and it is
|
||||
# almost always five words or fewer, which is exactly the population the
|
||||
# too-short skip above keeps away from GPT. In the first replay
|
||||
# (server-26#170, 09-22 10:00-12:00 ET) 25 transmissions said 10-8/clear and
|
||||
# 2 reached cleared_units. Rule-based on purpose: no model call, and only a
|
||||
# unit named BEFORE the status word counts, so "10-8, 10-8." or "CMT clear."
|
||||
# (no number) clears nobody rather than guessing.
|
||||
_CLEAR_WORD_RE = re.compile(
|
||||
r"\b(clear|10-?8|10-?98|back in service|in service|available)\b", re.IGNORECASE
|
||||
)
|
||||
_TEN_CODE_TOKEN_RE = re.compile(r"^10-?\d{1,2}$")
|
||||
_UNIT_PREFIX_WORDS = {"unit", "car", "vehicle", "engine", "ladder", "medic", "rescue", "post", "truck", "squad"}
|
||||
# A number after one of these is a place or a time, not a radio unit.
|
||||
_NOT_UNIT_PREFIX_WORDS = {"room", "route", "rt", "exit", "pole", "apartment", "apt", "floor",
|
||||
"building", "hours", "hour", "block", "lane", "highway", "interstate"}
|
||||
# The status word has to END the transmission: "clear the scene", "clear to
|
||||
# transport", "available for" are orders or plans, not a unit back in service.
|
||||
_TRAILING_OK = {"10-4", "thanks", "thank", "you", "k", "over", "now", "again", "from", "headquarters", "hq",
|
||||
"central", "dispatch"}
|
||||
|
||||
|
||||
def _short_clearance_unit(transcript: str) -> Optional[str]:
|
||||
text = (transcript or "").strip()
|
||||
if not text or "?" in text:
|
||||
return None # "45-9, are you clear?" asks; it doesn't report
|
||||
m = _CLEAR_WORD_RE.search(text)
|
||||
if not m:
|
||||
return None
|
||||
before = [t.strip(".,;:!") for t in text[: m.start()].split()]
|
||||
before = [t for t in before if t]
|
||||
after = [t.strip(".,;:!").lower() for t in text[m.end():].split()]
|
||||
if any(t and t not in _TRAILING_OK for t in after):
|
||||
return None
|
||||
if any(t.lower() in {"not", "is", "are", "negative", "you"} for t in before):
|
||||
return None # "not clear yet", "Is 45-9 clear", "you clear"
|
||||
for i, tok in enumerate(before[:4]):
|
||||
if not any(ch.isdigit() for ch in tok) or _TEN_CODE_TOKEN_RE.match(tok):
|
||||
continue
|
||||
prev = before[i - 1].lower() if i else ""
|
||||
if prev in _NOT_UNIT_PREFIX_WORDS:
|
||||
return None
|
||||
nxt = before[i + 1] if i + 1 < len(before) else ""
|
||||
if nxt.lower() in _NOT_UNIT_PREFIX_WORDS:
|
||||
return None # "1400 hours, clear"
|
||||
if nxt.isdigit():
|
||||
tok = f"{tok}-{nxt}" # "45 9 clear" is unit 45-9, not unit 45
|
||||
elif nxt.isalpha() and nxt[0].isupper() and nxt.lower() not in {"i'm", "im", "we're", "copy"}:
|
||||
tok = f"{tok} {nxt}" # "11 Adam, clear"
|
||||
if prev in _UNIT_PREFIX_WORDS:
|
||||
return f"{before[i - 1]} {tok}"
|
||||
return tok
|
||||
return None
|
||||
|
||||
|
||||
def _clearance_scene(transcript: str, unit: str) -> dict:
|
||||
"""A minimal scene for a rule-parsed clearance: the unit, and nothing that
|
||||
could make the incident-creation gate open a new incident for it."""
|
||||
return {
|
||||
"tags": [], "incident_type": None, "location": None, "location_coords": None,
|
||||
"resolved": False, "severity": "routine", "vehicles": [], "units": [unit],
|
||||
"cleared_units": [unit], "reassignment": False, "transcript": transcript,
|
||||
"transcript_corrected": None, "segment_indices": [], "embedding": None,
|
||||
}
|
||||
|
||||
|
||||
def _geo_dist_km(lat1: float, lon1: float, lat2: float, lon2: float) -> float:
|
||||
"""Haversine distance in km between two lat/lon points."""
|
||||
R = 6371.0
|
||||
|
||||
@@ -275,6 +275,12 @@ async def decide(call_id: str, ctx: dict) -> Optional[dict]:
|
||||
if ctx["is_thin_call"]:
|
||||
return None # thin calls have no transcript/units/coords to reason about
|
||||
|
||||
if _is_clearance_only(ctx):
|
||||
# "45-9, I'm clear." carries one fact: which unit is done. Only the
|
||||
# rules engine's unit match can say which incident that is; an LLM
|
||||
# link here would apply the clear to whatever incident it picked.
|
||||
return None
|
||||
|
||||
if not ctx["recent"]:
|
||||
return None # no incidents to correlate against — rules handles new-only
|
||||
|
||||
@@ -295,6 +301,14 @@ async def decide(call_id: str, ctx: dict) -> Optional[dict]:
|
||||
return None
|
||||
|
||||
|
||||
def _is_clearance_only(ctx: dict) -> bool:
|
||||
units = ctx.get("call_units") or []
|
||||
cleared = ctx.get("call_cleared") or []
|
||||
return bool(cleared) and set(units) <= set(cleared) and not (
|
||||
ctx.get("tags") or ctx.get("location") or ctx.get("call_vehicles") or ctx.get("incident_type")
|
||||
)
|
||||
|
||||
|
||||
_dead_models: set[str] = set()
|
||||
|
||||
|
||||
|
||||
@@ -142,15 +142,41 @@ async def _summarize_incident(inc: dict) -> None:
|
||||
await fstore.doc_set("incidents", incident_id, updates)
|
||||
|
||||
|
||||
def _auto_resolve_minutes(inc: dict) -> int:
|
||||
"""Quiet time before a timer close, by severity (see config: incident_auto_resolve_minutes_*)."""
|
||||
sev = (inc.get("severity") or "").lower()
|
||||
if sev in ("routine", "minor"):
|
||||
return settings.incident_auto_resolve_minutes_routine
|
||||
if sev == "moderate":
|
||||
return settings.incident_auto_resolve_minutes_moderate
|
||||
return settings.incident_auto_resolve_minutes
|
||||
|
||||
|
||||
async def _expire_reopen_windows(now) -> None:
|
||||
"""A timer-closed incident stops being reopenable once its window passes,
|
||||
so the correlator's reopenable pool stays bounded."""
|
||||
window = timedelta(minutes=settings.incident_reopen_window_minutes)
|
||||
for inc in await fstore.collection_list("incidents", status="resolved", reopenable=True):
|
||||
try:
|
||||
updated = datetime.fromisoformat(str(inc.get("updated_at", "")).replace("Z", "+00:00"))
|
||||
if updated.tzinfo is None:
|
||||
updated = updated.replace(tzinfo=timezone.utc)
|
||||
except ValueError:
|
||||
updated = None
|
||||
if updated is None or now - updated > window:
|
||||
await fstore.doc_set("incidents", inc["incident_id"], {"reopenable": False})
|
||||
|
||||
|
||||
async def _resolve_stale_incidents() -> None:
|
||||
"""Auto-resolve active incidents that have had no new calls for incident_auto_resolve_minutes."""
|
||||
"""Timer-close active incidents that have been quiet longer than their severity allows."""
|
||||
from app.internal import clock
|
||||
await _expire_reopen_windows(clock.now())
|
||||
all_active = await fstore.collection_list("incidents", status="active")
|
||||
if not all_active:
|
||||
return
|
||||
|
||||
from app.internal import clock
|
||||
now = clock.now()
|
||||
cutoff = timedelta(minutes=settings.incident_auto_resolve_minutes)
|
||||
count = 0
|
||||
|
||||
for inc in all_active:
|
||||
@@ -164,11 +190,12 @@ async def _resolve_stale_incidents() -> None:
|
||||
if updated_dt.tzinfo is None:
|
||||
updated_dt = updated_dt.replace(tzinfo=timezone.utc)
|
||||
idle_minutes = (now - updated_dt).total_seconds() / 60
|
||||
if idle_minutes > settings.incident_auto_resolve_minutes:
|
||||
if idle_minutes > _auto_resolve_minutes(inc):
|
||||
await fstore.doc_set("incidents", incident_id, {
|
||||
"status": "resolved",
|
||||
"resolved_at": now.isoformat(),
|
||||
"resolved_via": "idle_timeout",
|
||||
"reopenable": True,
|
||||
})
|
||||
from app.internal.incident_correlator import maybe_resolve_parent
|
||||
await maybe_resolve_parent(incident_id)
|
||||
|
||||
@@ -140,6 +140,7 @@ def _call_row(c: dict) -> dict:
|
||||
"cleared_units": c.get("cleared_units"),
|
||||
"location": c.get("location"),
|
||||
"skip_reason": c.get("skip_reason"),
|
||||
"srcaddr": c.get("srcaddr"),
|
||||
"corr_path": [p for p in paths if p] or ([c["corr_path"]] if c.get("corr_path") else []),
|
||||
"incident_ids": c.get("incident_ids") or [],
|
||||
}
|
||||
@@ -174,6 +175,7 @@ async def run_incidents(run_id: str, decoded: dict = Depends(require_admin_token
|
||||
"units_active": inc.get("units_active"),
|
||||
"units_cleared": inc.get("units_cleared"),
|
||||
"talkgroup_ids": inc.get("talkgroup_ids"),
|
||||
"srcaddrs": inc.get("srcaddrs"),
|
||||
"calls": rows,
|
||||
})
|
||||
orphans = sorted((r for r in by_id.values() if not r["incident_ids"]),
|
||||
|
||||
@@ -376,6 +376,7 @@ async def _run_extraction_pipeline(
|
||||
"status": "resolved",
|
||||
"resolved_at": clock.now().isoformat(),
|
||||
"resolved_via": "llm_closure",
|
||||
"reopenable": True, # provisional, see _extract_and_correlate
|
||||
})
|
||||
await incident_correlator.maybe_resolve_parent(incident_id)
|
||||
logger.info(f"Auto-resolved incident {incident_id} (LLM closure detection)")
|
||||
@@ -466,6 +467,11 @@ async def _extract_and_correlate(
|
||||
"status": "resolved",
|
||||
"resolved_at": clock.now().isoformat(),
|
||||
"resolved_via": "llm_closure",
|
||||
# One transmission read as "it's over" ("transport complete")
|
||||
# closed the whole 09-22 bridge MVA at 14:44 and its next 56
|
||||
# calls opened a second incident (server-26#170). Inferred from
|
||||
# a single call, so provisional, like a timer close.
|
||||
"reopenable": True,
|
||||
})
|
||||
await incident_correlator.maybe_resolve_parent(incident_id)
|
||||
logger.info(f"Auto-resolved incident {incident_id} (LLM closure detection)")
|
||||
|
||||
@@ -0,0 +1,142 @@
|
||||
"""
|
||||
Dispatch→10-8 lifecycle, as measured by the first replay (server-26#170):
|
||||
0 of 19 incidents resolved on a clear although 25 transmissions said one.
|
||||
Three independent breaks, each pinned here.
|
||||
"""
|
||||
from app.internal import incident_correlator as ic
|
||||
from app.internal.intelligence import _clearance_scene, _short_clearance_unit
|
||||
|
||||
|
||||
def test_short_clearance_names_the_unit_that_cleared():
|
||||
assert _short_clearance_unit("45-9, I'm clear.") == "45-9"
|
||||
assert _short_clearance_unit("Vehicle 1, clear.") == "Vehicle 1"
|
||||
assert _short_clearance_unit("11 Adam, clear") == "11 Adam"
|
||||
assert _short_clearance_unit("Car 12 10-8") == "Car 12"
|
||||
assert _short_clearance_unit("45 9 clear") == "45-9"
|
||||
assert _short_clearance_unit("Warrant 4, clear from the jail") is None or True # >5 words: GPT's job
|
||||
assert _short_clearance_unit("45-9 clear, thank you") == "45-9"
|
||||
|
||||
|
||||
def test_short_clearance_never_guesses():
|
||||
for t in ("10-8, 10-8.", "CMT clear.", "10-8, I'm back now. Clear.",
|
||||
"10-8, thank you.", "Show us 10-8, post 4.", "7, Charlie Central.", "10-4.",
|
||||
# review findings: questions, negations, orders, places, times
|
||||
"45-9, are you clear?", "Is 45-9 clear", "45-9, not clear yet.",
|
||||
"Engine 5 not available.", "45-9, clear the scene.", "Medic 3, clear to transport.",
|
||||
"Room 2 clear.", "Route 9 is clear.", "1400 hours, clear.", "you clear 45-9"):
|
||||
assert _short_clearance_unit(t) is None, t
|
||||
|
||||
|
||||
def test_clearance_scene_cannot_open_an_incident():
|
||||
scene = _clearance_scene("45-9, I'm clear.", "45-9")
|
||||
ctx = {"call_vehicles": scene["vehicles"], "coords": scene["location_coords"], "tags": scene["tags"]}
|
||||
assert not ic.has_event_substance(ctx)
|
||||
assert scene["severity"] == "routine" and scene["incident_type"] is None
|
||||
|
||||
|
||||
def test_clearance_matches_a_differently_spoken_unit():
|
||||
inc = {"units_active": ["11 Adam", "45-9"], "units_cleared": []}
|
||||
active, cleared, resolved = ic._apply_unit_clearance(inc, ["11-Adam"])
|
||||
assert active == ["45-9"]
|
||||
active, cleared, resolved = ic._apply_unit_clearance(
|
||||
{"units_active": active, "units_cleared": cleared}, ["45 9"])
|
||||
assert active == [] and resolved
|
||||
|
||||
|
||||
def test_a_unit_never_on_the_incident_cannot_close_it():
|
||||
inc = {"units_active": ["45-9"], "units_cleared": []}
|
||||
active, cleared, resolved = ic._apply_unit_clearance(inc, ["22-1"])
|
||||
assert active == ["45-9"] and cleared == [] and not resolved
|
||||
# an incident with no numbered unit ever active never resolves on a clear
|
||||
active, cleared, resolved = ic._apply_unit_clearance({"units_active": [], "units_cleared": []}, ["22-1"])
|
||||
assert not resolved
|
||||
|
||||
|
||||
def test_clearance_only_call_skips_the_llm():
|
||||
from app.internal import llm_correlator
|
||||
ctx = {"call_units": ["45-9"], "call_cleared": ["45-9"], "tags": [], "location": None,
|
||||
"call_vehicles": [], "incident_type": None}
|
||||
assert llm_correlator._is_clearance_only(ctx)
|
||||
assert not llm_correlator._is_clearance_only({**ctx, "tags": ["mva"]})
|
||||
assert not llm_correlator._is_clearance_only({**ctx, "call_cleared": []})
|
||||
|
||||
|
||||
def test_only_numbered_units_hold_an_incident_open():
|
||||
for junk in ("Desk", "Central", "Division", "sergeant", "unknown", "John", "Zebra", "10-8", "10 4"):
|
||||
assert not ic._is_trackable_unit(junk), junk
|
||||
for real in ("45-9", "11-Adam", "Whitestone 1", "E-14", "Highway 3-4", "7"):
|
||||
assert ic._is_trackable_unit(real), real
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# Provisional timer close + reopen, substantive cap (server-26#170 replay)
|
||||
# ---------------------------------------------------------------------------
|
||||
from datetime import datetime, timedelta, timezone # noqa: E402
|
||||
|
||||
from app.config import settings # noqa: E402
|
||||
from app.internal import summarizer # noqa: E402
|
||||
|
||||
|
||||
def test_quiet_timer_scales_with_severity():
|
||||
assert summarizer._auto_resolve_minutes({"severity": "routine"}) == settings.incident_auto_resolve_minutes_routine
|
||||
assert summarizer._auto_resolve_minutes({"severity": "minor"}) == settings.incident_auto_resolve_minutes_routine
|
||||
assert summarizer._auto_resolve_minutes({"severity": "moderate"}) == settings.incident_auto_resolve_minutes_moderate
|
||||
assert summarizer._auto_resolve_minutes({"severity": "major"}) == settings.incident_auto_resolve_minutes
|
||||
assert summarizer._auto_resolve_minutes({}) == settings.incident_auto_resolve_minutes
|
||||
|
||||
|
||||
def test_thin_calls_do_not_fill_the_call_cap():
|
||||
now = datetime(2026, 9, 22, 15, 0, tzinfo=timezone.utc)
|
||||
inc = {"call_ids": [f"c{i}" for i in range(60)], "substantive_call_count": 12,
|
||||
"started_at": (now - timedelta(minutes=40)).isoformat(),
|
||||
"updated_at": now.isoformat()}
|
||||
assert ic._incident_at_capacity(inc, now) is None
|
||||
legacy = {k: v for k, v in inc.items() if k != "substantive_call_count"}
|
||||
assert ic._incident_at_capacity(legacy, now).startswith("call_cap")
|
||||
|
||||
|
||||
def test_reopen_only_for_a_call_after_the_close():
|
||||
closed = {"resolved_at": "2026-09-22T15:00:00+00:00"}
|
||||
assert ic._after_close(closed, datetime(2026, 9, 22, 15, 5, tzinfo=timezone.utc))
|
||||
assert not ic._after_close(closed, datetime(2026, 9, 22, 14, 55, tzinfo=timezone.utc))
|
||||
|
||||
|
||||
def test_traffic_stops_become_events():
|
||||
from app.internal.intelligence import _self_initiated_backstop as b
|
||||
for t in ("45 Adam on a stop, Eastbound Central Express.",
|
||||
"11-0. CM2 on the stop, southbound, KFLA on the right.",
|
||||
"Car 7 on a traffic stop, Route 9 at Main"):
|
||||
tags, typ, sev = b(t, [], None, "routine")
|
||||
assert "traffic-stop" in tags and typ == "police" and sev == "minor", t
|
||||
tags, typ, sev = b("Charlie 1. You put me out with a pedestrian on a parkway", [], None, "routine")
|
||||
assert "self-initiated" in tags
|
||||
# negation and unrelated chatter stay untouched
|
||||
assert b("Do you want me to not pull the car over", [], None, "routine") == ([], None, "routine")
|
||||
assert b("45-8, go ahead.", [], None, "routine") == ([], None, "routine")
|
||||
# an existing type/severity is never downgraded
|
||||
assert b("on a stop", ["dwi"], "police", "moderate") == (["dwi", "traffic-stop"], "police", "moderate")
|
||||
|
||||
|
||||
def test_stop_backstop_stays_off_rail_bridge_and_ems_channels():
|
||||
from app.internal.intelligence import _self_initiated_backstop as b
|
||||
none = ([], None, "routine")
|
||||
assert b("Train 4 holding on the stop at Grand Central", [], None, "routine",
|
||||
"MTA PD Districts 6/7/11 - Police Dispatch") == none
|
||||
assert b("out with a disabled on the bridge, toll plaza", [], None, "routine",
|
||||
"MTA Bridges and Tunnels - Whitestone/Throgs Neck Bridge Patrols") == none
|
||||
assert b("Medic 2 pull over to the side and wait", [], None, "routine") == none
|
||||
assert b("ran a plate for a car stop", [], None, "routine") == none
|
||||
assert b("I dont think he is on a stop", [], None, "routine") == none
|
||||
assert b("45 Adam on a stop", [], None, "routine", "Ch 1 (Patched with 155.310)")[0] == ["traffic-stop"]
|
||||
|
||||
|
||||
def test_short_stop_report_opens_a_scene():
|
||||
import asyncio
|
||||
from unittest.mock import patch
|
||||
from app.internal import firestore as fstore, intelligence
|
||||
|
||||
async def run():
|
||||
with patch.object(fstore, "doc_set"), patch.object(fstore, "doc_get_cached", return_value=None):
|
||||
return await intelligence.extract_scenes("c1", "Adam 3 on a stop.", "Ch 1 (Patched with 155.310)")
|
||||
scenes = asyncio.run(run())
|
||||
assert len(scenes) == 1 and scenes[0]["tags"] == ["traffic-stop"] and scenes[0]["incident_type"] == "police"
|
||||
@@ -67,7 +67,10 @@ def test_clearing_a_unit_not_tracked_as_active_is_a_noop_for_active_list():
|
||||
inc = _incident(units_active=["6-7"], units_cleared=[])
|
||||
active, cleared, resolved = _apply_unit_clearance(inc, ["ghost-unit"])
|
||||
assert active == ["6-7"]
|
||||
assert cleared == ["ghost-unit"]
|
||||
# A unit never active on this incident is not recorded as cleared here
|
||||
# either (server-26#170 replay review): its 10-8 says nothing about this
|
||||
# incident, and recording it let the all-clear gate pass on a stray clear.
|
||||
assert cleared == []
|
||||
assert resolved is False
|
||||
|
||||
|
||||
|
||||
@@ -269,12 +269,14 @@ async def test_run_writes_only_to_its_sandbox_and_pins_the_clock(store):
|
||||
# The two Car 12 calls are one job; the Car 40 call five hours later is another.
|
||||
groups = sorted(sorted(i["call_ids"]) for i in sb_incidents.values())
|
||||
assert groups == [["call-1", "call-2"], ["call-3"]]
|
||||
# Each aged out on the replayed clock the way it would have live —
|
||||
# incident_auto_resolve_minutes after its last activity, not "now".
|
||||
# Each aged out on the replayed clock the way it would have live — its
|
||||
# severity's quiet timer after its last activity, not "now".
|
||||
assert run["metrics"]["resolved_via"] == {"idle_timeout": 2}
|
||||
first = next(i for i in sb_incidents.values() if "call-1" in i["call_ids"])
|
||||
idle = datetime.fromisoformat(first["resolved_at"]) - datetime.fromisoformat(first["updated_at"])
|
||||
assert timedelta(minutes=90) < idle <= timedelta(minutes=95)
|
||||
from app.internal.summarizer import _auto_resolve_minutes
|
||||
limit = timedelta(minutes=_auto_resolve_minutes(first))
|
||||
assert limit < idle <= limit + timedelta(minutes=5)
|
||||
assert run["metrics"]["calls"] == 3
|
||||
assert set(store.data[f"{root}/scenes"]) == {"call-1", "call-2", "call-3"}
|
||||
|
||||
|
||||
Reference in New Issue
Block a user