File size: 49,273 Bytes
b2f5c42 d50d1e1 b2f5c42 d14dce3 4f20fa7 d14dce3 97ee6bf d14dce3 97ee6bf d14dce3 655c300 b2f5c42 d50d1e1 b2f5c42 d14dce3 97ee6bf b2f5c42 655c300 b2f5c42 847587d b2f5c42 f046303 d50d1e1 847587d b2f5c42 847587d b2f5c42 d50d1e1 97ee6bf d50d1e1 b2f5c42 d14dce3 d808a62 d14dce3 fea6929 b2f5c42 d14dce3 d808a62 d14dce3 fea6929 b2f5c42 d14dce3 d808a62 d14dce3 fea6929 b2f5c42 847587d b2f5c42 d14dce3 b2f5c42 655c300 fea6929 847587d 655c300 fea6929 655c300 fea6929 847587d 655c300 fea6929 847587d 655c300 fea6929 655c300 d14dce3 b2f5c42 847587d b2f5c42 847587d d50d1e1 b2f5c42 d50d1e1 b2f5c42 847587d b2f5c42 d14dce3 d808a62 d14dce3 d808a62 d14dce3 d808a62 d14dce3 d808a62 d14dce3 97ee6bf b2f5c42 aedf4f8 b2f5c42 d14dce3 b2f5c42 655c300 d14dce3 d808a62 d14dce3 d808a62 aedf4f8 847587d aedf4f8 847587d 4f20fa7 847587d aedf4f8 847587d 9ad0a4c 4f20fa7 9ad0a4c d808a62 4f20fa7 d808a62 9ad0a4c 97ee6bf 9ad0a4c 97ee6bf 655c300 9ad0a4c 655c300 97ee6bf 655c300 847587d 655c300 9ad0a4c 655c300 97ee6bf 655c300 b2f5c42 fea6929 b2f5c42 655c300 b2f5c42 655c300 b2f5c42 655c300 b2f5c42 655c300 d50d1e1 655c300 d50d1e1 655c300 b2f5c42 655c300 97ee6bf b2f5c42 d50d1e1 b2f5c42 f046303 b2f5c42 f046303 b2f5c42 ea1618f b2f5c42 ea1618f b2f5c42 f046303 511af68 f046303 b2f5c42 ea1618f b2f5c42 ea1618f b2f5c42 fea6929 b2f5c42 655c300 aedf4f8 b2f5c42 e143b4a b2f5c42 d808a62 97e19ad d808a62 97e19ad d808a62 9ad0a4c aedf4f8 d808a62 9ad0a4c b2f5c42 d50d1e1 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 | """EmpathRAG Core runtime.
One guarded conversational RAG interface used by the demo and evaluation.
"""
from __future__ import annotations
from dataclasses import asdict, dataclass
import os
from pathlib import Path
from typing import Literal
import sqlite3
import time
from .ml_router import MLRouter
from .output_guard import validate_output
from .response_planner import (
CLARIFY,
INTERNATIONAL_SOURCE_HINT,
build_response_plan,
classify_intl_topic,
decide_stage,
has_international_concern,
render_crisis_response,
render_intl_factual_offer,
)
from .rephraser import RephraseResult, ResponseRephraser
from .safety_policy import SafetyLevel, SafetyTriagePolicy
from .service_graph import match_services
from .v2_schema import SafetyTier, SupportRoute, classify_route, map_safety_level
AudienceMode = Literal["student", "helping_friend"]
BackendMode = Literal["demo_fast", "hybrid_ml", "real_llm"]
@dataclass
class EmpathRAGResult:
response: str
route_label: str
safety_tier: str
should_intercept: bool
retrieved_sources: list[dict]
recommended_action: str
output_guard: dict
trajectory_state: str
latency_ms: dict
classifier_confidence: dict
retrieval_mode: str
safety_precheck: dict
safety_explanation: dict
safety_reason: str
escalation_reason: str
retrieval_corpus: str
emotion_name: str = "neutral"
crisis: bool = False
crisis_confidence: float = 0.0
retrieved_chunks: list[str] | None = None
international_concern: bool = False
conversation_stage: str = "offer"
turn_index: int = 1
intl_topic: str = ""
rephraser_provider: str = "deterministic"
rephraser_used_llm: bool = False
rephraser_latency_ms: float = 0.0
rephraser_last_error: str = ""
def to_dict(self) -> dict:
row = asdict(self)
row["retrieved_chunks"] = row["retrieved_chunks"] or []
row["safety_level"] = row["safety_tier"]
row["route"] = row["route_label"]
row["latency_ms"] = self.latency_ms
return row
@dataclass
class _TurnPlan:
"""Everything ``_plan_turn`` decides before the rephrase step.
Shared between sync ``run_turn`` and ``run_turn_streaming`` so the
pre-rephrase planning logic isn't duplicated across the two paths.
"""
message: str
session_id: str
audience_mode: str
backend_mode: str
turn_index: int
template_response: str
retrieved: list[dict]
recommended_action: str
route_label: str
safety_tier: SafetyTier
safety_reason: str
stage: str
intl_concern: bool
intl_topic: str
should_intercept: bool
retrieval_mode: str
latency: dict
t_total: float
stage1_level: str
stage1_reason: str
stage1_should_intercept: bool
ml_prediction: object
guardrail_info: dict
escalation_reason: str
class EmpathRAGCore:
# Safety layers that can be selectively disabled for ablation evaluation.
# See eval/run_ablation_eval.py for the harness that uses this.
VALID_DISABLED_LAYERS = frozenset({
"stage1_precheck", # Stage-1 lexical safety policy precheck
"output_guard", # validate_output at OFFER stage
"rephrase_safety", # verify_rephrased_safety on LLM output
"registry_filter", # resource registry + retrieval filter
})
def __init__(
self,
curated_db_path: Path | str = Path("data/curated/indexes/metadata_curated.db"),
retrieval_corpus: str = "curated_support",
top_k: int = 5,
router_model_dir: Path | str = Path("models/router"),
ml_confidence_threshold: float = 0.35,
use_model_guardrail: bool | None = None,
compute_ig_on_intercept: bool | None = None,
guardrail_threshold: float = 0.5,
disable_layers: set[str] | None = None,
):
self.disable_layers = set(disable_layers or set())
invalid = self.disable_layers - self.VALID_DISABLED_LAYERS
if invalid:
raise ValueError(
f"Unknown disable_layers: {invalid}. "
f"Valid: {sorted(self.VALID_DISABLED_LAYERS)}"
)
self.curated_db_path = Path(curated_db_path)
self.retrieval_corpus = "curated_support" if self.curated_db_path.exists() else retrieval_corpus
self.top_k = top_k
self.safety_policy = SafetyTriagePolicy()
self.ml_router = MLRouter(Path(router_model_dir), min_confidence=ml_confidence_threshold)
self.use_model_guardrail = _env_flag("EMPATHRAG_CORE_USE_GUARDRAIL") if use_model_guardrail is None else use_model_guardrail
self.compute_ig_on_intercept = _env_flag("EMPATHRAG_CORE_COMPUTE_IG") if compute_ig_on_intercept is None else compute_ig_on_intercept
self.guardrail_threshold = guardrail_threshold
self.rephraser = ResponseRephraser()
self._guardrail = None
self._guardrail_error = ""
self.tier_history: dict[str, list[str]] = {}
self.locked_sessions: dict[str, str] = {}
# Session-scoped sticky flags so context detected on one turn carries
# forward (international concern, last specific route) instead of being
# re-derived from each message in isolation.
self.session_intl_flag: dict[str, bool] = {}
# Turns since the user last mentioned F-1 / visa / international
# status keywords. ISSS auto-surfacing decays after 2 silent turns
# so the conversation can fully shift topic without ISSS hijacking
# every subsequent response.
self.session_turns_since_intl: dict[str, int] = {}
self.session_last_specific_route: dict[str, str] = {}
# Rolling per-session message log so the rephraser has continuity.
# Only the last few user-assistant pairs are passed to the LLM; full
# history is kept here for diagnostics. Each entry: {"role", "content"}.
self.session_message_history: dict[str, list[dict]] = {}
def reset_session(self, session_id: str | None = None) -> None:
if session_id:
self.tier_history.pop(session_id, None)
self.locked_sessions.pop(session_id, None)
self.session_intl_flag.pop(session_id, None)
self.session_turns_since_intl.pop(session_id, None)
self.session_last_specific_route.pop(session_id, None)
self.session_message_history.pop(session_id, None)
else:
self.tier_history.clear()
self.locked_sessions.clear()
self.session_intl_flag.clear()
self.session_turns_since_intl.clear()
self.session_last_specific_route.clear()
self.session_message_history.clear()
# Input length cap: defends against (1) accidental wall-of-text pastes
# ballooning LLM cost and planner regex matching, (2) deliberate DoS
# via runaway-length payloads. Configurable via env so larger contexts
# can be unlocked for evaluation runs that legitimately need them.
MAX_USER_MESSAGE_CHARS = int(os.getenv("EMPATHRAG_MAX_USER_MESSAGE_CHARS", "2000"))
def run_turn(
self,
message: str,
session_id: str,
audience_mode: AudienceMode = "student",
resource_profile: str = "umd",
backend_mode: BackendMode = "hybrid_ml",
turn_index: int | None = None,
) -> EmpathRAGResult:
plan = self._plan_turn(message, session_id, audience_mode, backend_mode, turn_index)
if plan.should_intercept:
return self._finalize_turn(plan, plan.template_response, None)
rephrase_result = self.rephraser.rephrase(
user_message=message,
template_response=plan.template_response,
retrieved_sources=plan.retrieved,
recommended_action=plan.recommended_action,
history=self._recent_history(session_id),
skip_safety_check="rephrase_safety" in self.disable_layers,
)
return self._finalize_turn(plan, rephrase_result.response, rephrase_result)
def run_turn_streaming(
self,
message: str,
session_id: str,
audience_mode: AudienceMode = "student",
resource_profile: str = "umd",
backend_mode: BackendMode = "hybrid_ml",
turn_index: int | None = None,
):
"""Generator variant of :meth:`run_turn`.
Yields:
* ``("token", accumulated_text)`` for each streamed chunk
* ``("done", EmpathRAGResult)`` exactly once at the end
Crisis interception, deterministic mode, and safety-rejected rephrases
all collapse to a single deterministic ``("token", final_text)`` plus
a ``("done", ...)``. Real Groq/Anthropic streaming only kicks in when
``EMPATHRAG_REPHRASER_ENABLED=1`` and the route is non-crisis.
"""
plan = self._plan_turn(message, session_id, audience_mode, backend_mode, turn_index)
if plan.should_intercept:
result = self._finalize_turn(plan, plan.template_response, None)
yield ("token", result.response)
yield ("done", result)
return
history = self._recent_history(session_id)
if not self.rephraser.enabled:
# Deterministic mode: no LLM, no real streaming. Hand the template
# through and let the demo's word-chunk reveal layer fake-stream it
# the way it always has.
rephrase_result = self.rephraser.rephrase(
user_message=message,
template_response=plan.template_response,
retrieved_sources=plan.retrieved,
recommended_action=plan.recommended_action,
history=history,
skip_safety_check="rephrase_safety" in self.disable_layers,
)
result = self._finalize_turn(plan, rephrase_result.response, rephrase_result)
yield ("token", result.response)
yield ("done", result)
return
accumulated = ""
rephrase_result: RephraseResult | None = None
for event in self.rephraser.rephrase_streaming(
user_message=message,
template_response=plan.template_response,
retrieved_sources=plan.retrieved,
recommended_action=plan.recommended_action,
history=history,
skip_safety_check="rephrase_safety" in self.disable_layers,
):
if event[0] == "chunk":
_, accumulated, _provider_name = event
yield ("token", accumulated)
elif event[0] == "final":
_, rephrase_result = event
# rephrase_streaming guarantees exactly one ("final", ...).
assert rephrase_result is not None
result = self._finalize_turn(plan, rephrase_result.response, rephrase_result)
# If the output guard corrected the response or the streamed text was
# safety-rejected and replaced with the deterministic fallback, the
# bubble currently shows stale text. Yield one more correction.
if result.response != accumulated:
yield ("token", result.response)
yield ("done", result)
def _recent_history(self, session_id: str, max_messages: int = 4) -> list[dict]:
"""Return the last ``max_messages`` entries from this session's
message history. Used to give the rephraser continuity across turns
without leaking older context the planner has already moved past."""
hist = self.session_message_history.get(session_id, [])
return hist[-max_messages:] if hist else []
# ------------------------------------------------------------------
# Internal: plan + finalize
# ------------------------------------------------------------------
def _plan_turn(
self,
message: str,
session_id: str,
audience_mode: AudienceMode,
backend_mode: BackendMode,
turn_index: int | None,
) -> _TurnPlan:
if turn_index is None:
turn_index = len(self.tier_history.get(session_id, [])) + 1
t_total = time.perf_counter()
latency: dict[str, float] = {}
# Length cap: a very long input is either an accidental wall-of-text
# paste or a DoS attempt. We truncate before the planner sees it (so
# regex matching + downstream LLM cost stay bounded) and surface a
# short clarify-style response if the truncation actually mattered.
message_was_truncated = False
if message and len(message) > self.MAX_USER_MESSAGE_CHARS:
message = message[: self.MAX_USER_MESSAGE_CHARS]
message_was_truncated = True
t0 = time.perf_counter()
if "stage1_precheck" in self.disable_layers:
# Ablation: pretend the message passes safety. Used to measure
# what fraction of escalation catches depend on the Stage-1
# lexical check vs. downstream layers (ML router, contextual
# overrides, output guard). Never set in production.
from .safety_policy import SafetyDecision, SafetyLevel as _SL
stage1_decision = SafetyDecision(
level=_SL.PASS, confidence=0.0,
reason="stage1_disabled_ablation", should_intercept=False,
)
else:
stage1_decision = self.safety_policy.classify(message, confidence=0.0, model_flag=False)
latency["stage1_precheck_ms"] = _elapsed_ms(t0)
t0 = time.perf_counter()
guardrail_info = self._run_optional_guardrail(message, skip_ig=True) if not stage1_decision.should_intercept else _guardrail_skipped("stage1_intercept")
latency["model_guardrail_ms"] = _elapsed_ms(t0)
if guardrail_info["available"] and guardrail_info["model_flag"]:
safety_decision = self.safety_policy.classify(
message,
confidence=float(guardrail_info["confidence"]),
model_flag=True,
)
else:
safety_decision = stage1_decision
latency["hard_safety_ms"] = latency["stage1_precheck_ms"]
wellbeing_request = _wellbeing_request(message)
safety_tier = map_safety_level(safety_decision.level, wellbeing_request=wellbeing_request)
safety_reason = safety_decision.reason
safety_tier, safety_reason = _apply_contextual_safety_overrides(
message, safety_tier, safety_reason, audience_mode
)
rule_decision = classify_route(message, safety_tier, audience_mode=audience_mode)
t0 = time.perf_counter()
ml_prediction = self.ml_router.predict(message, rule_decision.route, safety_tier)
latency["classifier_ms"] = _elapsed_ms(t0)
route_label = ml_prediction.route_label if backend_mode in {"hybrid_ml", "real_llm"} else rule_decision.route.value
if safety_tier == SafetyTier.IMMINENT_SAFETY:
route_label = SupportRoute.PEER_HELPER.value if rule_decision.route == SupportRoute.PEER_HELPER else SupportRoute.CRISIS_IMMEDIATE.value
else:
safety_tier = SafetyTier(ml_prediction.safety_tier) if ml_prediction.used_ml else safety_tier
escalation_reason = self._update_trajectory(session_id, safety_tier.value, message)
if session_id in self.locked_sessions:
safety_tier = SafetyTier.IMMINENT_SAFETY
safety_reason = self.locked_sessions[session_id]
should_intercept = safety_decision.should_intercept or safety_tier == SafetyTier.IMMINENT_SAFETY
retrieval_mode = _retrieval_mode(backend_mode, should_intercept)
if should_intercept and self.compute_ig_on_intercept and self.use_model_guardrail:
t0 = time.perf_counter()
guardrail_info = self._run_optional_guardrail(message, skip_ig=False)
latency["integrated_gradients_ms"] = _elapsed_ms(t0)
t0 = time.perf_counter()
if "registry_filter" in self.disable_layers:
# Ablation: no resource-registry filtering, no curated retrieval.
# Planner falls back to generic templates with no named UMD
# resources. Used to measure how much of the system's
# resource-grounding depends on the registry layer.
retrieved = []
else:
retrieved = self._retrieve(message, route_label, safety_tier.value, audience_mode, should_intercept)
latency["retrieval_ms"] = _elapsed_ms(t0)
# Cross-cutting: when an F-1 / visa / international-status worry shows up
# at any point in the session, keep the ISSS surface and the
# international-aware framing for subsequent turns β but decay it so
# ISSS doesn't hijack every later turn after the topic has shifted.
intl_now = has_international_concern(message)
if intl_now:
self.session_intl_flag[session_id] = True
self.session_turns_since_intl[session_id] = 0
else:
self.session_turns_since_intl[session_id] = (
self.session_turns_since_intl.get(session_id, 0) + 1
)
intl_session = self.session_intl_flag.get(session_id, False)
# "Active" means F-1 framing should drive retrieval + sub-topic logic
# this turn. After 2 turns with no F-1 keyword the flag remains True
# for diagnostics but stops dominating responses.
intl_active = intl_session and self.session_turns_since_intl.get(session_id, 0) <= 2
if not should_intercept and intl_active:
already_has_isss = any(
(s.get("source_name") or "").lower().startswith("umd international")
or (s.get("service_id") or "").startswith("umd_isss")
for s in retrieved
)
if not already_has_isss:
retrieved = [INTERNATIONAL_SOURCE_HINT] + retrieved
intl_topic = ""
if should_intercept:
template_response = render_crisis_response(
route_label, audience_mode=audience_mode, user_message=message,
)
recommended_action = _recommended_action(route_label, safety_tier.value)
stage = "offer"
else:
response_plan = build_response_plan(
message,
route_label,
safety_tier.value,
retrieved,
audience_mode,
international_concern_override=intl_active,
)
stage = decide_stage(message, route_label, safety_tier.value, turn_index)
# Only classify F-1 sub-topic when the framing is actively in play.
# After 2 silent turns, decayed; don't pin the planner to OPT/RCL/etc.
intl_topic = classify_intl_topic(message) if intl_active else ""
# Conversational utterances (greeting / goodbye / meta) route to
# short purposeful templates BEFORE the regular minimal /
# incomplete checks. These aren't support content β they're
# conversation openers and closers. Treating them as OFFER
# responses produces jarring "let me tell you about UMD CC"
# replies to a simple "hi".
conv_intent = (
"" if message_was_truncated else _conversational_intent(message)
)
# Length-cap / minimal / incomplete short-circuits, in priority
# order. Each picks a clarify-style template instead of trying
# to OFFER a full route response on insufficient or excessive
# input. Output guard skipped for the clarify stage.
minimal_kind = "" if (message_was_truncated or conv_intent) else _minimal_response_kind(message)
if message_was_truncated:
template_response = (
"That's a lot, and I want to make sure I focus on what matters most to you. "
"Your message was long enough that I only have the first part. Could you say "
"which piece feels most pressing right now, in a sentence or two?"
)
stage = CLARIFY
recommended_action = response_plan.recommended_action
elif conv_intent == "greeting":
template_response = _render_greeting()
stage = CLARIFY
recommended_action = response_plan.recommended_action
elif conv_intent == "goodbye":
template_response = _render_goodbye()
stage = CLARIFY
recommended_action = response_plan.recommended_action
elif conv_intent == "meta":
template_response = _render_meta()
stage = CLARIFY
recommended_action = response_plan.recommended_action
elif minimal_kind:
template_response = _render_minimal_followup(minimal_kind, intl_session)
stage = CLARIFY
recommended_action = response_plan.recommended_action
elif _is_incomplete_message(message):
# The student's message trails off ("what should i do to",
# "honestly", "kind of"). Guessing the intent and producing
# a full OFFER response is worse than asking them to finish.
template_response = _render_incomplete_followup()
stage = CLARIFY
recommended_action = response_plan.recommended_action
elif intl_topic and stage == "offer":
template_response = render_intl_factual_offer(intl_topic, message)
recommended_action = response_plan.recommended_action
else:
template_response = response_plan.render(stage)
recommended_action = response_plan.recommended_action
return _TurnPlan(
message=message,
session_id=session_id,
audience_mode=audience_mode,
backend_mode=backend_mode,
turn_index=turn_index,
template_response=template_response,
retrieved=retrieved,
recommended_action=recommended_action,
route_label=route_label,
safety_tier=safety_tier,
safety_reason=safety_reason,
stage=stage,
intl_concern=intl_session,
intl_topic=intl_topic,
should_intercept=should_intercept,
retrieval_mode=retrieval_mode,
latency=latency,
t_total=t_total,
stage1_level=stage1_decision.level.value,
stage1_reason=stage1_decision.reason,
stage1_should_intercept=stage1_decision.should_intercept,
ml_prediction=ml_prediction,
guardrail_info=guardrail_info,
escalation_reason=escalation_reason,
)
def _finalize_turn(
self,
plan: _TurnPlan,
response: str,
rephrase_result: RephraseResult | None,
) -> EmpathRAGResult:
if "output_guard" in self.disable_layers:
output_guard = {"allowed": True, "reason": "output_guard_disabled_ablation", "flags": []}
elif plan.should_intercept:
output_guard = {"allowed": True, "reason": "crisis_template", "flags": []}
elif plan.stage == "clarify":
# Clarifying replies after minimal user input (yes/no/maybe) are
# intentionally short open-ended invitations. The OFFER-stage
# missing-action / pure-validation checks would punish them.
output_guard = {"allowed": True, "reason": "minimal_response_clarify", "flags": []}
elif plan.stage == "offer":
# Output guard runs after rephrase (or fall-through to template)
# only at OFFER stage. LISTEN/PERMISSION are intentionally
# reflective and would fail missing-action / pure-validation checks.
guard = validate_output(
response, plan.retrieved, plan.safety_tier.value, plan.route_label, []
)
output_guard = {"allowed": guard.allowed, "reason": guard.reason, "flags": guard.flags}
if guard.fallback_required and guard.corrected_response:
response = guard.corrected_response
else:
output_guard = {"allowed": True, "reason": f"listening_stage_{plan.stage}", "flags": []}
latency = dict(plan.latency)
latency["total_ms"] = _elapsed_ms(plan.t_total)
# Append this turn to the per-session history so the next turn has
# context. Trimmed to last ~6 messages to bound memory + LLM payload.
hist = self.session_message_history.setdefault(plan.session_id, [])
hist.append({"role": "user", "content": plan.message})
hist.append({"role": "assistant", "content": response})
if len(hist) > 12: # 6 user-assistant pairs
del hist[: len(hist) - 12]
return EmpathRAGResult(
response=response,
route_label=plan.route_label,
safety_tier=plan.safety_tier.value,
should_intercept=plan.should_intercept,
retrieved_sources=_source_summaries(plan.retrieved),
recommended_action=plan.recommended_action,
output_guard=output_guard,
trajectory_state="locked" if plan.session_id in self.locked_sessions else "active",
latency_ms=latency,
classifier_confidence={
"route": plan.ml_prediction.route_confidence,
"tier": plan.ml_prediction.tier_confidence,
"model_available": plan.ml_prediction.model_available,
"used_ml": plan.ml_prediction.used_ml and plan.backend_mode in {"hybrid_ml", "real_llm"},
"reason": plan.ml_prediction.reason,
},
retrieval_mode=plan.retrieval_mode,
safety_precheck={
"stage": "hard_lexical_precheck",
"level": plan.stage1_level,
"reason": plan.stage1_reason,
"should_intercept": plan.stage1_should_intercept,
"ran_before_ml": True,
},
safety_explanation=plan.guardrail_info,
safety_reason=plan.safety_reason,
escalation_reason=plan.escalation_reason,
retrieval_corpus=self.retrieval_corpus,
emotion_name=_emotion_name(plan.message),
crisis=plan.should_intercept,
crisis_confidence=float(plan.guardrail_info.get("confidence") or (1.0 if plan.should_intercept else 0.0)),
retrieved_chunks=[row.get("text", "") for row in plan.retrieved],
international_concern=plan.intl_concern,
conversation_stage=plan.stage,
turn_index=plan.turn_index,
intl_topic=plan.intl_topic,
rephraser_provider=(rephrase_result.provider_name if rephrase_result else "deterministic"),
rephraser_used_llm=(rephrase_result.used_llm if rephrase_result else False),
rephraser_latency_ms=(rephrase_result.latency_ms if rephrase_result else 0.0),
rephraser_last_error=(rephrase_result.last_error if rephrase_result else ""),
)
def _run_optional_guardrail(self, message: str, skip_ig: bool) -> dict:
if not self.use_model_guardrail:
return _guardrail_skipped("disabled")
try:
guardrail = self._load_guardrail()
if guardrail is None:
return {
"available": False,
"model": "deberta_nli",
"reason": self._guardrail_error or "load_failed",
"confidence": 0.0,
"model_flag": False,
"ig_tokens": [],
}
model_flag, confidence, ig_tokens = guardrail.check(
message,
threshold=self.guardrail_threshold,
skip_ig=skip_ig,
)
return {
"available": True,
"model": "deberta_nli",
"reason": "model_guardrail_checked",
"confidence": float(confidence),
"model_flag": bool(model_flag),
"ig_tokens": ig_tokens,
"ig_computed": not skip_ig and bool(ig_tokens),
}
except Exception as exc:
self._guardrail_error = str(exc)
return {
"available": False,
"model": "deberta_nli",
"reason": f"guardrail_error: {exc}",
"confidence": 0.0,
"model_flag": False,
"ig_tokens": [],
}
def _load_guardrail(self):
if self._guardrail is not None:
return self._guardrail
try:
from src.models.guardrail_ig import SafetyGuardrail
except ImportError:
try:
from models.guardrail_ig import SafetyGuardrail
except ImportError as exc:
self._guardrail_error = str(exc)
return None
try:
self._guardrail = SafetyGuardrail()
except Exception as exc:
self._guardrail_error = str(exc)
return None
return self._guardrail
def _update_trajectory(self, session_id: str, safety_tier: str, message: str) -> str:
history = self.tier_history.setdefault(session_id, [])
history.append(safety_tier)
self.tier_history[session_id] = history[-3:]
text = message.lower()
if len(history[-3:]) == 3 and all(tier in {"imminent_safety", "high_distress"} for tier in history[-3:]):
self.locked_sessions[session_id] = "three_consecutive_high_risk_turns"
return "three_consecutive_high_risk_turns"
if safety_tier in {"imminent_safety", "high_distress"} and "goodbye" in text:
return "peer_goodbye_or_farewell_escalation"
if safety_tier in {"imminent_safety", "high_distress"} and any(
phrase in text
for phrase in (
"you are the only one",
"only one i can talk to",
"keep this secret",
"don't tell anyone",
"refuse external help",
"secrecy",
"never suggest counseling",
)
):
return "dependency_or_secrecy_redirect"
return ""
def _retrieve(
self,
message: str,
route: str,
safety_tier: str,
audience_mode: str,
should_intercept: bool,
) -> list[dict]:
if route == SupportRoute.OUT_OF_SCOPE.value:
return []
usage_modes = ("crisis_only",) if should_intercept else ("retrieval", "wellbeing_only") if safety_tier == "wellbeing" else ("retrieval",)
selected: list[dict] = []
graph_rows = [
node.as_source("resource registry route match")
for node in match_services(route, safety_tier, audience_mode, limit=self.top_k)
if (node.usage_modes[0] if node.usage_modes else "retrieval") in usage_modes
]
selected.extend(graph_rows)
if self.curated_db_path.exists() and len(selected) < self.top_k:
selected.extend(self._retrieve_curated(message, route, usage_modes, limit=self.top_k - len(selected)))
return _dedupe_sources(selected)[: self.top_k]
def _retrieve_curated(self, message: str, route: str, usage_modes: tuple[str, ...], limit: int) -> list[dict]:
topics, source_names = _targets_for_route(route, message)
conn = sqlite3.connect(self.curated_db_path)
conn.row_factory = sqlite3.Row
rows = conn.execute(
"""
SELECT id, resource_id, text, source_id, source_name, source_type,
title, url, topic, audience, risk_level, usage_mode, summary,
last_checked, notes
FROM chunks
WHERE usage_mode IN ({})
""".format(",".join("?" * len(usage_modes))),
tuple(usage_modes),
).fetchall()
conn.close()
query = message.lower()
scored = []
for row in rows:
score = 0
reasons = []
if row["topic"] in topics:
score += 8
reasons.append(f"topic match: {row['topic']}")
if row["source_name"] in source_names:
score += 7
reasons.append(f"preferred source: {row['source_name']}")
haystack = f"{row['title']} {row['summary']} {row['text']}".lower()
for token in _keywords(query):
if token in haystack:
score += 1
row_dict = dict(row)
row_dict["why_retrieved"] = "; ".join(reasons[:2]) if reasons else "curated retrieval match"
scored.append((score, row_dict))
scored.sort(key=lambda item: item[0], reverse=True)
return [row for score, row in scored[: max(limit, 0)] if score > 0 or not scored[:1]]
def run_turn(
message: str,
session_id: str,
audience_mode: AudienceMode = "student",
resource_profile: str = "umd",
backend_mode: BackendMode = "hybrid_ml",
) -> EmpathRAGResult:
core = EmpathRAGCore()
return core.run_turn(message, session_id, audience_mode, resource_profile, backend_mode)
def _apply_contextual_safety_overrides(
message: str,
safety_tier: SafetyTier,
safety_reason: str,
audience_mode: str,
) -> tuple[SafetyTier, str]:
text = message.lower()
peer_context = audience_mode == "helping_friend" or any(
phrase in text for phrase in ("my friend", "my roommate", "my labmate", "my teammate", "someone i know")
)
if peer_context and "goodbye" in text and any(phrase in text for phrase in ("locked", "will not answer", "won't answer", "not answering")):
return SafetyTier.IMMINENT_SAFETY, "peer_goodbye_unreachable"
if safety_tier == SafetyTier.SUPPORT_NAVIGATION and any(
phrase in text for phrase in ("you are the only one", "only one i can talk to", "keep this secret", "don't tell anyone", "no one can help")
):
return SafetyTier.HIGH_DISTRESS, "dependency_or_secrecy_redirect"
if safety_tier == SafetyTier.SUPPORT_NAVIGATION and any(
phrase in text for phrase in ("crises if it gets worse", "dark moods", "pointless", "scared to escalate")
):
return SafetyTier.HIGH_DISTRESS, "high_distress_language"
if any(
phrase in text
for phrase in (
"panic attacks",
"devastated",
"terrified",
"scared about",
"hopeless",
"everything hollow",
"ideation creeping",
"goodbye texts",
"mentioned goodbye",
"refuse external help",
"secrecy",
)
):
return SafetyTier.HIGH_DISTRESS, "high_distress_language"
if safety_tier == SafetyTier.SUPPORT_NAVIGATION and any(
phrase in text
for phrase in (
"not in immediate danger",
"intrusive thoughts spike",
"canceling plans",
"five-minute mindfulness",
"sensory overwhelm",
"quick reset",
"journaling prompts",
"breathing technique",
)
):
return SafetyTier.WELLBEING, "wellbeing_support_language"
return safety_tier, safety_reason
def _retrieval_mode(backend_mode: str, should_intercept: bool) -> str:
if should_intercept:
return "registry_filtered_crisis_only"
if backend_mode == "demo_fast":
return "registry_filtered_faiss"
return "registry_filtered_faiss_plus_router"
def _targets_for_route(route: str, message: str) -> tuple[set[str], set[str]]:
mapping = {
SupportRoute.CRISIS_IMMEDIATE.value: ({"crisis_immediate_help", "emergency_services"}, {"988 Suicide & Crisis Lifeline", "UMD Counseling Center"}),
SupportRoute.PEER_HELPER.value: ({"crisis_immediate_help", "help_seeking_script", "counseling_services"}, {"988 Suicide & Crisis Lifeline", "UMD Counseling Center", "JED Foundation"}),
SupportRoute.ACCESSIBILITY_ADS.value: ({"accessibility_disability", "campus_navigation"}, {"UMD Accessibility & Disability Service"}),
SupportRoute.ADVISOR_CONFLICT.value: ({"advisor_conflict", "graduate_student_support"}, {"UMD Graduate School Ombuds", "UMD Graduate School"}),
SupportRoute.BASIC_NEEDS.value: ({"help_seeking_script", "campus_navigation", "graduate_student_support"}, {"UMD Dean of Students", "UMD Graduate School"}),
SupportRoute.ANXIETY_PANIC.value: ({"anxiety_stress", "grounding_exercise", "counseling_services"}, {"NIMH", "NAMI", "UMD Counseling Center"}),
SupportRoute.AUTHORITY_MISCONDUCT.value: (
{"authority_misconduct", "care_violence_confidential", "campus_navigation"},
{"UMD Office of Civil Rights & Sexual Misconduct (OCRSM)", "UMD Office of Student Conduct", "UMD Dean of Students", "UMD CARE to Stop Violence"},
),
SupportRoute.LOW_MOOD.value: ({"depression_support", "counseling_services"}, {"NIMH", "NAMI", "UMD Counseling Center"}),
SupportRoute.COUNSELING_NAVIGATION.value: ({"counseling_services", "campus_navigation", "therapy_expectations"}, {"UMD Counseling Center"}),
SupportRoute.ACADEMIC_SETBACK.value: ({"academic_burnout", "graduate_student_support", "counseling_services"}, {"UMD Counseling Center", "UMD Graduate School"}),
SupportRoute.EXAM_STRESS.value: ({"academic_burnout", "anxiety_stress", "grounding_exercise"}, {"UMD Counseling Center", "CDC", "NIMH"}),
SupportRoute.LONELINESS_ISOLATION.value: ({"isolation_loneliness", "counseling_services"}, {"NAMI", "UMD Counseling Center", "CDC"}),
}
return mapping.get(route, ({"counseling_services", "anxiety_stress", "academic_burnout"}, {"UMD Counseling Center", "NIMH"}))
def _source_summaries(rows: list[dict]) -> list[dict]:
return [
{
"title": row.get("title", ""),
"source_name": row.get("source_name", ""),
"url": row.get("url", ""),
"topic": row.get("topic", ""),
"risk_level": row.get("risk_level", ""),
"usage_mode": row.get("usage_mode", ""),
"source_type": row.get("source_type", ""),
"why_retrieved": row.get("why_retrieved", ""),
"documents": row.get("documents", []) or [],
# Trust signal: surface the verification date in the UI so users
# (especially clinicians) can see how fresh the link is.
"last_verified": row.get("last_verified") or row.get("last_checked", ""),
}
for row in rows
]
def _dedupe_sources(rows: list[dict]) -> list[dict]:
selected = []
seen = set()
for row in rows:
key = (row.get("source_name", ""), row.get("title", ""))
if key in seen:
continue
seen.add(key)
selected.append(row)
return selected
def _recommended_action(route: str, safety_tier: str) -> str:
if safety_tier == SafetyTier.IMMINENT_SAFETY.value:
if route == SupportRoute.PEER_HELPER.value:
return "Do not handle this alone; contact emergency or crisis support now and involve a trusted nearby person."
return "Contact 988 or emergency services now, and move near another person if you can."
plan = build_response_plan("", route, safety_tier, [], "student")
return plan.recommended_action
# Words that, when they're the last token of a short message, strongly
# suggest the user trailed off mid-sentence. Conservative list: only
# prepositions/conjunctions/articles/auxiliaries that almost never close
# a real sentence. We deliberately omit "is", "are", "was" β those can
# legitimately close a sentence ("It is what it is").
_INCOMPLETE_TERMINAL_WORDS = frozenset({
# Prepositions
"to", "for", "with", "about", "of", "on", "at", "by", "from", "into",
"over", "under", "between", "without", "through", "during",
# Conjunctions / subordinators
"and", "but", "or", "because", "if", "when", "while", "though",
"although", "since", "until", "whereas",
# Articles
"a", "an", "the",
# Auxiliary fragments
"should", "would", "could", "might", "may", "will", "can", "do",
"does", "did",
# Fragment determiners (only ones that never legitimately close a sentence)
"my", "your", "their", "our",
# Transitive verbs that almost always need an object
"want", "need", "wish",
})
# Deliberately NOT in the terminal set: "this", "that", "these", "those",
# "her", "his", "him". These DO legitimately close sentences ("I don't know
# what to do with that.", "I told her."), and putting them in the terminal
# set produces false positives on real emotional disclosure.
# Words that close a sentence legitimately in longer messages ("It is what
# it is") but signal incompleteness in very short ones ("the thing is").
_SHORT_ONLY_TERMINAL_WORDS = frozenset({
"is", "are", "was", "were", "i",
})
# Whole-message hedges that carry no content. Distinct from minimal
# affirmations (yes/no/ok) β those have a routing meaning; these don't.
_HEDGE_ONLY_MESSAGES = frozenset({
"honestly", "kind of", "kinda", "sort of", "sorta", "i mean",
"well", "uh", "um", "you know", "like", "i guess", "hmm",
"idk maybe", "idk i guess", "maybe idk",
})
def _is_incomplete_message(message: str) -> bool:
"""True if the message looks like it trailed off or is hedge-only."""
if not message:
return True
raw = message.strip()
if not raw:
return True
# Pure ellipsis / punctuation-only
if all(c in ".,;:!? \t-" for c in raw):
return True
# Strip trailing punctuation/whitespace before testing the terminal word
text = raw.rstrip(".,!?;:- \t").strip()
if not text:
return True
text_lower = " ".join(text.lower().split())
if text_lower in _HEDGE_ONLY_MESSAGES:
return True
# Last word ending the (short) message is a known incomplete-terminal
words = text_lower.split()
if not words:
return True
if words[-1] in _INCOMPLETE_TERMINAL_WORDS and len(text) < 80:
return True
# Short-only terminals: "is" / "are" close real sentences in longer
# messages but signal incompleteness in fragments like "the thing is".
if words[-1] in _SHORT_ONLY_TERMINAL_WORDS and len(words) <= 3:
return True
return False
_MINIMAL_AFFIRM = frozenset({
"yes", "yeah", "yep", "yup", "sure", "ok", "okay", "kk", "k",
"alright", "fine", "definitely", "absolutely",
})
_MINIMAL_NEGATE = frozenset({"no", "nope", "nah", "not really", "nuh"})
_MINIMAL_UNSURE = frozenset({"maybe", "idk", "dunno", "unsure", "not sure", "perhaps"})
def _minimal_response_kind(message: str) -> str:
"""Return 'affirm' / 'negate' / 'unsure' / '' for very short reply-only turns.
Single-word affirmations are ambiguous β without context, re-rendering the
same OFFER template is the worst possible response. We catch them here and
redirect to a clarifying-question response instead.
"""
text = (message or "").strip().lower()
if not text or len(text) > 24:
return ""
# Strip trailing punctuation; collapse internal whitespace.
clean = " ".join(text.replace(".", " ").replace("!", " ").replace("?", " ").replace(",", " ").split())
if clean in _MINIMAL_AFFIRM:
return "affirm"
if clean in _MINIMAL_NEGATE:
return "negate"
if clean in _MINIMAL_UNSURE:
return "unsure"
return ""
# Conversation openers / closers / meta β these are NOT student-support
# content. They route to short purposeful templates instead of a heavy
# OFFER response that would feel jarring for an opener or premature for a
# close.
_GREETING_PATTERNS = frozenset({
"hi", "hello", "hey", "yo", "hiya", "hey there", "hi there", "hello there",
"good morning", "good afternoon", "good evening",
"sup", "what's up", "whats up", "wassup",
"hey empathrag", "hi empathrag",
})
_GOODBYE_PATTERNS = frozenset({
"thanks", "thank you", "thx", "ty", "thank u",
"bye", "goodbye", "see you", "see ya", "later",
"i'll think about it", "ill think about it", "let me think",
"got it", "cool", "appreciate it", "that helps",
"alright then", "ok thanks", "thanks bye",
})
_META_PATTERNS = (
"are you a bot", "are you an ai", "are you a robot", "are you human",
"what are you", "who are you", "are you real", "are you a person",
"is this an ai", "is this a bot", "is this real",
"what can you do", "how do you work", "what is this",
)
def _conversational_intent(message: str) -> str:
"""Returns 'greeting' / 'goodbye' / 'meta' / '' for non-support
conversational utterances. These are routed to short purposeful
templates rather than full OFFER responses."""
text = (message or "").strip().lower()
if not text or len(text) > 60:
return ""
clean = " ".join(text.rstrip(".,!?;:").split())
if clean in _GREETING_PATTERNS:
return "greeting"
if clean in _GOODBYE_PATTERNS:
return "goodbye"
if any(pattern in clean for pattern in _META_PATTERNS):
return "meta"
return ""
def _render_greeting() -> str:
return (
"Hi. I'm here if there's something on your mind β academic stress, "
"visa or status worry, feeling low, helping a friend through "
"something, or anything else weighing on you. Say what feels "
"easiest to share first."
)
def _render_goodbye() -> str:
return (
"Take care of yourself. The resources we talked through will still "
"be here when you're ready to follow up. If anything shifts in the "
"meantime β especially anything that feels urgent β UMD Counseling "
"Center is at counseling.umd.edu and 988 is always there."
)
def _render_meta() -> str:
return (
"I'm EmpathRAG β a research prototype built at UMD to help students "
"navigate campus support. I'm not a counselor, therapist, or "
"emergency service. I listen, then point to verified UMD resources "
"when it feels like the right moment. Anything you'd like to start with?"
)
def _render_incomplete_followup() -> str:
"""Response for messages that trail off mid-sentence or are content-free
hedges. Invite completion without guessing at intent."""
return (
"Looks like the thought might have cut off. Take your time and say "
"the rest when you're ready, in whatever words feel right."
)
def _render_minimal_followup(kind: str, intl_session: bool) -> str:
"""Short clarifying response when the user's reply was just yes/no/maybe.
Deliberately doesn't re-name resources or push action β that would just be
the previous turn repeated. Keep the conversation open instead.
"""
if kind == "affirm":
body = (
"Okay. Which part of what we just talked about feels most useful to "
"start with? You can name it in your own words, or pick one of the "
"directions I mentioned a moment ago."
)
elif kind == "negate":
body = (
"Got it. Want to keep talking it through, or try a different angle? "
"We can slow down or pick a different starting point."
)
else:
body = (
"That's okay. Want me to keep listening for a bit, or sit with what's "
"making it hard to know? Either is fine."
)
if intl_session:
# Hold the international-aware frame open without re-naming ISSS.
body += " I'll keep what you said about your status in mind."
return body
def _wellbeing_request(message: str) -> bool:
text = message.lower()
return any(word in text for word in ("grounding", "ground", "panic", "breathing", "cope", "mindfulness"))
def _emotion_name(message: str) -> str:
text = message.lower()
if any(word in text for word in ("safe tonight", "hurt myself", "suicide", "goodbye")):
return "distress"
if any(word in text for word in ("panic", "anxiety", "stress", "exam", "deadline")):
return "anxiety"
if any(word in text for word in ("advisor", "retaliatory", "funding")):
return "frustration"
if any(word in text for word in ("better", "hopeful", "proud")):
return "hopeful"
return "neutral"
def _keywords(query: str) -> list[str]:
return [token for token in query.replace("?", " ").replace(".", " ").split() if len(token) > 4]
def _elapsed_ms(start: float) -> float:
return round((time.perf_counter() - start) * 1000, 2)
def _env_flag(name: str) -> bool:
return os.getenv(name, "").strip().lower() in {"1", "true", "yes", "on"}
def _guardrail_skipped(reason: str) -> dict:
return {
"available": False,
"model": "deberta_nli",
"reason": reason,
"confidence": 0.0,
"model_flag": False,
"ig_tokens": [],
"ig_computed": False,
}
|