Spaces:
Sleeping
Sleeping
File size: 35,893 Bytes
2744f33 0417c59 2744f33 21fbec4 2744f33 0417c59 2744f33 0417c59 2744f33 0417c59 2744f33 0417c59 2744f33 21fbec4 0417c59 2744f33 21fbec4 2744f33 | 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 | """
api/main.py
===========
Phase 3: Secure FastAPI Backend
Endpoints
---------
- ``POST /register`` β Hash password via bcrypt, create user profile.
- ``POST /login`` β Verify password, return JWT token.
- ``POST /analyze`` β (JWT required) Run CNN inference, temporal analysis,
and recommendation engine. Returns scores, interventions, attention weights.
- ``GET /history`` β (JWT required) Retrieve past analysis sessions.
Security
--------
- Passwords are NEVER stored in plaintext (bcrypt).
- JWT tokens authenticate all ``/analyze`` and ``/history`` requests.
- Stress history is encrypted at rest via Fernet (AES-256).
Persistence
-----------
- User accounts and analysis sessions are stored in a SQLite database.
- Sessions survive server restarts and are available upon re-login.
"""
from __future__ import annotations
import hashlib
import logging
import os
import statistics
import threading
import time
from typing import Any, Optional
import torch
from fastapi import Depends, FastAPI, HTTPException, Query, status
from fastapi.middleware.cors import CORSMiddleware
from fastapi.security import HTTPAuthorizationCredentials, HTTPBearer
from pydantic import BaseModel, Field
from database.db import DatabaseManager
from database.feedback import FeedbackStore
from intervention.engine import RecommendationEngine
from intervention.temporal_model import SecureTemporalModel
from models.architecture import (
DeBERTaStressClassifier,
MiniLMStressClassifier,
OptimizedMultichannelCNN,
)
from security.auth import (
create_jwt_token,
decode_jwt_token,
hash_password,
verify_password,
)
from utils.llm_reward import get_llm_reward
from utils.reward import compute_combined_reward
from utils.sentiment import compute_sentiment_dampening, get_sentiment_score
from utils.text_preprocessing import clean_text
from datetime import timedelta
# ---------------------------------------------------------------------------
# App & global state
# ---------------------------------------------------------------------------
_APP_START_TIME = time.time()
app = FastAPI(
title="Stress Detection API",
description="Secure, intervention-oriented stress detection system",
version="2.0.0",
)
# ---------------------------------------------------------------------------
# CORS β allow all origins in development / single-server deployments.
# For production, restrict ``allow_origins`` to your frontend domain(s).
# ---------------------------------------------------------------------------
app.add_middleware(
CORSMiddleware,
allow_origins=["*"],
allow_credentials=True,
allow_methods=["*"],
allow_headers=["*"],
)
# SQLite-backed user + session store
_db = DatabaseManager()
# Feedback / experience-replay store (same DB file)
_feedback_store = FeedbackStore()
# Singletons
_recommendation_engine = RecommendationEngine()
_temporal_model = SecureTemporalModel()
# Security
_bearer_scheme = HTTPBearer()
# Model (lazy-loaded on first request)
_model: Optional[torch.nn.Module] = None
_vocab: Optional[Any] = None
_model_type: str = "cnn"
_decision_threshold: float = 0.5
_tokenizer: Optional[Any] = None
_tokenizer_max_length: int = 256
_feature_dim: int = 0
_DEFAULT_VOCAB_SIZE = 10000
_CHECKPOINT_PATH = os.environ.get(
"STRESS_MODEL_CHECKPOINT", "checkpoints/model.pt"
)
logger = logging.getLogger(__name__)
# ---------------------------------------------------------------------------
# Inference guardrails
# ---------------------------------------------------------------------------
# The decision threshold is NEVER allowed to fall below this value, even if
# the training checkpoint was produced with a pathologically low threshold
# (e.g. 0.15) caused by unconstrained F1-only threshold calibration.
_MIN_DECISION_THRESHOLD: float = 0.50
# Clip raw model probabilities to this range before applying the threshold.
# Extreme values (< 0.10 or > 0.90) can amplify noise and indicate a
# poorly calibrated model; restricting the range keeps decisions sane.
_PROB_CLIP_MIN: float = 0.10
_PROB_CLIP_MAX: float = 0.90
# Dead-zone gating: predictions within this distance of the adaptive
# threshold are relabelled "uncertain" to avoid committing to a potentially
# false high/low near the decision boundary.
_CONFIDENCE_DEAD_ZONE: float = 0.07
# Ensemble Monte-Carlo Dropout: number of stochastic forward passes and the
# standard-deviation threshold above which the prediction is flagged as
# uncertain by the ensemble.
_ENSEMBLE_PASSES: int = 3
_ENSEMBLE_UNCERTAINTY_STD: float = 0.08
# Serialises MC-Dropout ensemble passes so concurrent requests do not
# interfere with each other's Dropout layer state.
_inference_lock: threading.Lock = threading.Lock()
def _enable_dropout(m: torch.nn.Module) -> None:
"""Set only Dropout layers to train mode for MC-Dropout ensemble passes."""
if isinstance(m, torch.nn.Dropout):
m.train()
def _classify_stress_level(stress_prob: float, decision_threshold: float) -> str:
"""Map a stress probability to a 4-way human-readable level.
Bands (all relative to ``decision_threshold``, default 0.50):
============ ================ ================================
Level Probability range Meaning
============ ================ ================================
low < thresholdβ0.10 Confidently not stressed
uncertain Β±0.10 of thresh. Near the boundary; unclear
moderate threshold+0.10 β¦ Clearly stressed but manageable
threshold+0.25
high β₯ threshold+0.25 High-stress; escalate
============ ================ ================================
"""
low_bound = decision_threshold - 0.10
uncertain_upper = decision_threshold + 0.10
high_lower = decision_threshold + 0.25
if stress_prob >= high_lower:
return "high"
if stress_prob >= uncertain_upper:
return "moderate"
if stress_prob >= low_bound:
return "uncertain"
return "low"
def _compute_confidence(stress_prob: float, decision_threshold: float) -> float:
"""Return a confidence score in [0, 1] derived from distance to the threshold.
A score of 1.0 means the prediction is maximally far from the boundary
(e.g. stress_prob = 0.0 or 1.0 with threshold = 0.5). Values close to
0.0 indicate the prediction is right on the decision boundary.
"""
dist = abs(stress_prob - decision_threshold)
return float(min(dist / _MIN_DECISION_THRESHOLD, 1.0))
# ---------------------------------------------------------------------------
# Short-input handler for common single-word inputs
# ---------------------------------------------------------------------------
_STRESS_WORDS = frozenset([
"tired", "exhausted", "overwhelmed", "burnt", "done",
"stressed", "anxious", "depressed", "hopeless", "miserable",
])
_NEUTRAL_WORDS = frozenset([
"fine", "ok", "okay", "good", "alright",
])
# ---------------------------------------------------------------------------
# Inference post-processing: signal strength and contrast filtering
# ---------------------------------------------------------------------------
# High-frequency function words that carry no stress signal on their own.
# Used to measure how much *meaningful* content is in the input.
_LOW_SIGNAL_WORDS: frozenset[str] = frozenset({
"i", "me", "my", "we", "you", "he", "she", "it", "they",
"am", "is", "are", "was", "were", "be", "been", "being",
"a", "an", "the", "and", "or", "of", "to", "in", "on",
"at", "by", "for", "with", "as", "this", "that", "do",
"did", "does", "have", "has", "had", "will", "would", "can",
"could", "should", "may", "might", "shall",
})
# Contrast conjunctions that signal a positive override following a
# stress-trigger phrase (e.g. "I am stressed but happy").
_CONTRAST_CONJUNCTIONS: frozenset[str] = frozenset({
"but", "however", "although", "though", "yet", "despite",
"nevertheless", "nonetheless", "whereas",
})
def _handle_short_input(text: str) -> float | None:
"""Return a preset stress probability for very short inputs.
Returns ``None`` when the input is not recognised as a common
single-word pattern and should be passed to the model instead.
"""
cleaned = text.lower().strip()
if cleaned in _STRESS_WORDS:
return 0.8
if cleaned in _NEUTRAL_WORDS:
return 0.2
return None
def _apply_signal_filter(text: str, stress_prob: float) -> float:
"""Dampen stress probability when the input lacks meaningful content words.
Inputs consisting almost entirely of low-signal function words (e.g.
"I am the") carry no semantic content and should not trigger a high
stress prediction. Applies a 0.70 dampening factor when fewer than
three content words are detected.
Parameters
----------
text : str
Raw input text.
stress_prob : float
Current stress probability (after model + sentiment correction).
Returns
-------
float
Dampened stress probability.
"""
content_words = [
w for w in text.lower().split() if w not in _LOW_SIGNAL_WORDS
]
if len(content_words) < 3:
stress_prob *= 0.7
return stress_prob
def _apply_contrast_filter(text: str, stress_prob: float) -> float:
"""Dampen stress probability when a contrast conjunction is present.
Phrases like "I am stressed *but* happy" or "exhausted *however* grateful"
carry a positive override that should suppress the stress score. A 0.80
dampening factor is applied whenever any contrast conjunction is found,
regardless of position.
Parameters
----------
text : str
Raw input text.
stress_prob : float
Current stress probability.
Returns
-------
float
Dampened stress probability.
"""
tokens = set(text.lower().split())
if tokens & _CONTRAST_CONJUNCTIONS:
stress_prob *= 0.8
return stress_prob
def _get_model() -> torch.nn.Module:
"""Lazy-load or create the CNN model.
If a checkpoint file exists at ``_CHECKPOINT_PATH``, the function
attempts to load the saved ``model_state_dict``. When the checkpoint
was produced by an *older* architecture (e.g. one that used a single
``fc`` layer instead of the current ``attention`` + ``classifier``
head), loading with ``strict=True`` would raise a ``RuntimeError``.
To stay backward-compatible the loader:
1. Tries ``strict=True`` first.
2. On key-mismatch ``RuntimeError``, retries with ``strict=False``
so that all *compatible* weights (embedding, conv layers) are
restored while new layers keep their random initialisation.
3. Logs every missing / unexpected key for transparency.
If no checkpoint exists the model is created with random weights.
"""
global _model, _decision_threshold, _model_type, _tokenizer, _tokenizer_max_length, _feature_dim
if _model is None:
checkpoint = None
if os.path.isfile(_CHECKPOINT_PATH):
try:
checkpoint = torch.load(
_CHECKPOINT_PATH, map_location="cpu", weights_only=True,
)
except Exception as exc:
logger.warning(
"Failed to read checkpoint %s (%s); using randomly "
"initialised weights.",
_CHECKPOINT_PATH,
exc,
)
if isinstance(checkpoint, dict):
_model_type = checkpoint.get("model_type", "cnn")
threshold = checkpoint.get("decision_threshold")
# Backward-compatible: threshold may be serialized as a tensor.
if isinstance(threshold, torch.Tensor):
threshold = float(threshold.item())
if isinstance(threshold, (float, int)):
_decision_threshold = float(threshold)
# Safety guard: never use a threshold below the minimum, regardless
# of what the checkpoint reports. Unconstrained F1-only calibration
# during training can produce pathological values like 0.15.
_decision_threshold = max(_decision_threshold, _MIN_DECISION_THRESHOLD)
_tokenizer_max_length = int(
checkpoint.get("tokenizer_max_length", _tokenizer_max_length)
)
_feature_dim = int(checkpoint.get("feature_dim", 0))
feature_columns = checkpoint.get("feature_columns")
if _feature_dim == 0 and isinstance(feature_columns, list):
_feature_dim = len(feature_columns)
dropout = float(
checkpoint.get(
"dropout",
0.3 if _model_type == "cnn" else 0.1,
)
)
else:
_model_type = "cnn"
dropout = 0.3
_feature_dim = 0
if _model_type == "deberta":
_model = DeBERTaStressClassifier(dropout=dropout)
elif _model_type == "minilm":
_model = MiniLMStressClassifier(dropout=dropout)
else:
_model = OptimizedMultichannelCNN(
vocab_size=_DEFAULT_VOCAB_SIZE,
embed_dim=128,
num_filters=64,
kernel_sizes=(2, 3, 5),
num_classes=2,
dropout=dropout,
aux_dim=_feature_dim,
)
if _feature_dim > 0:
logger.info(
"Checkpoint expects %d auxiliary features; inference "
"will use zero-filled features unless provided.",
_feature_dim,
)
if _model_type in {"deberta", "minilm"}:
from transformers import AutoTokenizer
model_name = None
if isinstance(checkpoint, dict):
model_name = checkpoint.get("model_name")
if model_name is None:
model_name = _model.MODEL_NAME
_tokenizer = AutoTokenizer.from_pretrained(model_name)
if checkpoint is not None:
state_dict = (
checkpoint.get("model_state_dict", checkpoint)
if isinstance(checkpoint, dict)
else checkpoint
)
try:
_model.load_state_dict(state_dict, strict=True)
logger.info("Loaded checkpoint from %s", _CHECKPOINT_PATH)
except RuntimeError as exc:
logger.warning(
"Strict checkpoint load failed (%s); retrying with "
"strict=False to restore compatible weights.",
exc,
)
result = _model.load_state_dict(state_dict, strict=False)
if result.missing_keys:
logger.warning(
"Missing keys (randomly initialised): %s",
result.missing_keys,
)
if result.unexpected_keys:
logger.warning(
"Unexpected keys (ignored): %s",
result.unexpected_keys,
)
else:
logger.info(
"No checkpoint found at %s; using randomly initialised "
"weights.",
_CHECKPOINT_PATH,
)
_model.eval()
return _model
# ---------------------------------------------------------------------------
# Pydantic schemas
# ---------------------------------------------------------------------------
class RegisterRequest(BaseModel):
username: str = Field(..., min_length=3, max_length=50)
password: str = Field(..., min_length=8)
class LoginRequest(BaseModel):
username: str
password: str
remember_me: bool = True
class TokenResponse(BaseModel):
access_token: str
token_type: str = "bearer"
class AnalyzeRequest(BaseModel):
text: str = Field(..., min_length=1)
class InterventionResponse(BaseModel):
title: str
description: str
category: str
priority: int
class AnalyzeResponse(BaseModel):
stress_score: float
stress_label: str
stress_level: str # "low" | "moderate" | "high" | "uncertain"
confidence: float # how far the prediction is from the decision boundary [0, 1]
temporal: dict
interventions: list[InterventionResponse]
is_crisis: bool
crisis_message: Optional[str] = None
matched_triggers: list[str]
attention_weights: list[float]
requires_escalation: bool = False # True when 3+ consecutive above-threshold sessions
is_uncertain: bool = False # True when ensemble std is high or near-boundary
class SessionResponse(BaseModel):
"""A single past analysis session."""
id: int
stress_score: float
stress_label: str
temporal_data: dict
interventions: list[dict]
is_crisis: bool
crisis_message: Optional[str] = None
matched_triggers: list[str]
attention_weights: list[float]
created_at: float
class HistoryResponse(BaseModel):
"""Paginated list of past analysis sessions."""
sessions: list[SessionResponse]
total: int
class FeedbackRequest(BaseModel):
"""User-submitted feedback on a single prediction."""
text: str = Field(..., min_length=1)
prediction: float = Field(..., ge=0.0, le=1.0)
user_feedback: int = Field(..., ge=0, le=1,
description="1 = prediction was correct, 0 = wrong")
class FeedbackResponse(BaseModel):
"""Acknowledgement returned after storing feedback."""
status: str
reward: float
llm_reward: Optional[int] = None
feedback_id: int
class FeedbackStatsResponse(BaseModel):
"""Aggregated feedback statistics for the authenticated user."""
total: int
mean_reward: float
n_correct: int
n_wrong: int
accuracy_rate: float
class PersonalizationResponse(BaseModel):
"""Per-user score adjustment derived from their feedback history."""
user_bias: float
feedback_count: int
description: str
# ---------------------------------------------------------------------------
# Auth dependency
# ---------------------------------------------------------------------------
def _get_current_user(
credentials: HTTPAuthorizationCredentials = Depends(_bearer_scheme),
) -> str:
"""Decode the JWT and return the username (``sub`` claim)."""
try:
payload = decode_jwt_token(credentials.credentials)
username: str | None = payload.get("sub")
if username is None:
raise HTTPException(
status_code=status.HTTP_401_UNAUTHORIZED,
detail="Invalid token: missing subject",
)
if not _db.user_exists(username):
raise HTTPException(
status_code=status.HTTP_401_UNAUTHORIZED,
detail="User not found",
)
return username
except Exception as exc:
raise HTTPException(
status_code=status.HTTP_401_UNAUTHORIZED,
detail=f"Could not validate credentials: {exc}",
)
# ---------------------------------------------------------------------------
# Endpoints
# ---------------------------------------------------------------------------
@app.get("/health")
def health() -> dict:
"""Liveness / readiness probe.
Returns the service status, uptime in seconds, and whether the
prediction model has been loaded into memory.
"""
return {
"status": "ok",
"uptime_seconds": round(time.time() - _APP_START_TIME, 1),
"model_loaded": _model is not None,
"model_type": _model_type,
}
@app.get("/model/info")
def model_info() -> dict:
"""Return metadata about the currently loaded prediction model.
Useful for the UI settings panel and for debugging.
"""
return {
"model_type": _model_type,
"decision_threshold": _decision_threshold,
"vocab_size": _DEFAULT_VOCAB_SIZE,
"checkpoint_path": _CHECKPOINT_PATH,
"checkpoint_exists": os.path.isfile(_CHECKPOINT_PATH),
"prob_clip_min": _PROB_CLIP_MIN,
"prob_clip_max": _PROB_CLIP_MAX,
"feature_dim": _feature_dim,
}
@app.post("/login", response_model=TokenResponse)
def login(req: LoginRequest) -> TokenResponse:
"""Verify credentials and return a JWT token."""
username_normalized = req.username.strip().lower()
logger.info("Login attempt for username: %s", username_normalized)
user = _db.get_user(username_normalized)
if user is None:
logger.warning("Login failed: user '%s' not found in database.", username_normalized)
raise HTTPException(
status_code=status.HTTP_401_UNAUTHORIZED,
detail="Invalid username or password",
)
password_match = verify_password(req.password, user["password_hash"])
logger.info("Password match result for '%s': %s", username_normalized, password_match)
if not password_match:
logger.warning("Login failed: wrong password for user '%s'.", username_normalized)
raise HTTPException(
status_code=status.HTTP_401_UNAUTHORIZED,
detail="Invalid username or password",
)
expiry = timedelta(days=7) if req.remember_me else timedelta(hours=1)
token = create_jwt_token({"sub": username_normalized}, expires_delta=expiry)
logger.info("Token generated successfully for user '%s'.", username_normalized)
return TokenResponse(access_token=token)
@app.post("/register", response_model=TokenResponse, status_code=201)
def register(req: RegisterRequest) -> TokenResponse:
"""Register a new user with bcrypt-hashed password."""
username_normalized = req.username.strip().lower()
if _db.user_exists(username_normalized):
raise HTTPException(
status_code=status.HTTP_400_BAD_REQUEST,
detail="Username already exists",
)
hashed = hash_password(req.password)
logger.info("Registering user '%s' with hashed password.", username_normalized)
_db.create_user(username_normalized, hashed)
token = create_jwt_token({"sub": username_normalized})
return TokenResponse(access_token=token)
@app.post("/token/refresh", response_model=TokenResponse)
def refresh_token(username: str = Depends(_get_current_user)) -> TokenResponse:
"""Issue a fresh JWT for the authenticated user."""
token = create_jwt_token({"sub": username})
return TokenResponse(access_token=token)
@app.post("/analyze", response_model=AnalyzeResponse)
def analyze(
req: AnalyzeRequest,
username: str = Depends(_get_current_user),
) -> AnalyzeResponse:
"""Run full stress analysis pipeline (JWT required).
Pipeline:
0. Clean and normalise the input text (HTML, URLs, emojis, etc.).
1. Tokenize text and run OptimizedMultichannelCNN inference.
2. Decrypt user's temporal history, update profile, re-encrypt.
3. Run RecommendationEngine.
4. Persist session to database.
5. Return scores, interventions, and attention weights.
"""
model = _get_model()
# ββ 0. Text preprocessing ββ
# Normalise input before any downstream processing so that the text
# seen by the model exactly matches what was seen during training.
text = clean_text(req.text)
if not text:
raise HTTPException(
status_code=status.HTTP_422_UNPROCESSABLE_ENTITY,
detail="Input text is empty after preprocessing.",
)
# ββ 0b. Short-input shortcut ββ
short_result = _handle_short_input(text)
# ββ 1. Model Inference ββ
attn_weights: list[float] = []
is_uncertain_ensemble: bool = False
if short_result is not None:
stress_prob = short_result
elif _model_type == "cnn":
tokens = _simple_tokenize(text)
input_tensor = torch.tensor([tokens], dtype=torch.long)
# Single eval-mode pass β captures attention weights.
with torch.no_grad():
if _feature_dim > 0:
# No auxiliary features at inference time; use zero-filled inputs.
aux_features = torch.zeros(
(1, _feature_dim), dtype=torch.float
)
output = model(input_tensor, aux_features=aux_features)
else:
output = model(input_tensor)
logits = output["logits"]
attn_weights = output["attention_weights"][0].tolist()
p_eval = float(torch.softmax(logits, dim=-1)[0, 1])
# Ensemble MC-Dropout: additional stochastic passes with dropout only.
# We selectively set Dropout layers to train mode instead of the whole
# model, so the BatchNorm / LayerNorm statistics stay in eval mode.
# The lock serialises model-state mutations so concurrent requests
# do not interfere with each other's Dropout state.
ensemble_probs: list[float] = [p_eval]
with _inference_lock:
model.apply(_enable_dropout)
try:
for _ in range(_ENSEMBLE_PASSES - 1):
with torch.no_grad():
if _feature_dim > 0:
out = model(input_tensor, aux_features=aux_features)
else:
out = model(input_tensor)
ensemble_probs.append(
float(torch.softmax(out["logits"], dim=-1)[0, 1])
)
finally:
model.eval()
stress_prob = statistics.mean(ensemble_probs)
ensemble_std = statistics.pstdev(ensemble_probs)
is_uncertain_ensemble = ensemble_std > _ENSEMBLE_UNCERTAINTY_STD
else:
if _tokenizer is None:
raise HTTPException(
status_code=500,
detail="Tokenizer not initialized for transformer model.",
)
encoded = _tokenizer(
text,
return_tensors="pt",
truncation=True,
max_length=_tokenizer_max_length,
)
sentiment_val = get_sentiment_score(text)
sentiment_tensor = torch.tensor([sentiment_val], dtype=torch.float)
with torch.no_grad():
output = model(
input_ids=encoded["input_ids"],
attention_mask=encoded.get("attention_mask"),
sentiment=sentiment_tensor,
)
logits = output["logits"]
attn_weights = []
if short_result is None and _model_type != "cnn":
# CNN stress_prob is already averaged across ensemble passes above.
probs = torch.softmax(logits, dim=-1)
stress_prob = float(probs[0, 1])
if short_result is None:
# ββ Sentiment correction ββ
dampening = compute_sentiment_dampening(text)
stress_prob = stress_prob * dampening
# ββ Signal-strength filter ββ
stress_prob = _apply_signal_filter(text, stress_prob)
# ββ Contrast-phrase filter ββ
stress_prob = _apply_contrast_filter(text, stress_prob)
# ββ Probability calibration ββ
stress_prob = float(min(max(stress_prob, _PROB_CLIP_MIN), _PROB_CLIP_MAX))
stress_label = (
"stress" if stress_prob >= _decision_threshold else "no_stress"
)
# ββ Multi-level classification + confidence ββ
stress_level = _classify_stress_level(stress_prob, _decision_threshold)
confidence = _compute_confidence(stress_prob, _decision_threshold)
# ββ 2. Temporal Analysis (decrypt β compute β re-encrypt) ββ
user_data = _db.get_user(username)
analysis, new_encrypted = _temporal_model.process(
score=stress_prob,
encrypted_history=user_data["encrypted_history"] if user_data else None,
)
_db.update_encrypted_history(username, new_encrypted)
# ββ Dead-zone gating ββ
# If the probability is within _CONFIDENCE_DEAD_ZONE of the adaptive
# threshold, the prediction is too close to the boundary to be reliable.
# Override the label to "uncertain" to avoid a false high/low call.
if abs(stress_prob - analysis.adaptive_threshold) < _CONFIDENCE_DEAD_ZONE:
stress_level = "uncertain"
# ββ Layer 4: Escalation detection ββ
# Query the 2 most recent saved sessions (before saving the current one).
# If the last 3 sessions (including the current) all exceed the adaptive
# threshold, flag the user for escalation to a professional.
recent_sessions = _db.get_sessions(username, limit=2, offset=0)
past_scores = [s["stress_score"] for s in recent_sessions]
all_recent_scores = [stress_prob] + past_scores
requires_escalation = (
analysis.score_count >= 3
and len(all_recent_scores) >= 3
and all(s >= analysis.adaptive_threshold for s in all_recent_scores[:3])
)
# ββ 3. Recommendation Engine ββ
recommendation = _recommendation_engine.recommend(
text=text,
stress_score=stress_prob,
is_volatile=analysis.is_volatile,
requires_escalation=requires_escalation,
)
# ββ 4. Build response ββ
temporal_dict = {
"stress_velocity": analysis.stress_velocity,
"adaptive_threshold": analysis.adaptive_threshold,
"exceeds_threshold": analysis.exceeds_threshold,
"is_volatile": analysis.is_volatile,
"volatility": analysis.volatility,
"score_count": analysis.score_count,
}
interventions_list = [
{
"title": iv.title,
"description": iv.description,
"category": iv.category,
"priority": iv.priority,
}
for iv in recommendation.interventions
]
# ββ 5. Persist session to database ββ
_db.save_session(
username=username,
stress_score=stress_prob,
stress_label=stress_label,
temporal_data=temporal_dict,
interventions=interventions_list,
is_crisis=recommendation.is_crisis,
crisis_message=recommendation.crisis_message,
matched_triggers=recommendation.matched_triggers,
attention_weights=attn_weights,
)
return AnalyzeResponse(
stress_score=stress_prob,
stress_label=stress_label,
stress_level=stress_level,
confidence=confidence,
temporal=temporal_dict,
interventions=[
InterventionResponse(**iv) for iv in interventions_list
],
is_crisis=recommendation.is_crisis,
crisis_message=recommendation.crisis_message,
matched_triggers=recommendation.matched_triggers,
attention_weights=attn_weights,
requires_escalation=requires_escalation,
is_uncertain=is_uncertain_ensemble or stress_level == "uncertain",
)
@app.get("/history", response_model=HistoryResponse)
def history(
limit: int = Query(default=50, ge=1, le=200),
offset: int = Query(default=0, ge=0),
username: str = Depends(_get_current_user),
) -> HistoryResponse:
"""Retrieve past analysis sessions for the authenticated user.
Sessions are returned newest-first and support pagination via
``limit`` and ``offset`` query parameters.
"""
sessions = _db.get_sessions(username, limit=limit, offset=offset)
total = _db.get_session_count(username)
return HistoryResponse(
sessions=[SessionResponse(**s) for s in sessions],
total=total,
)
@app.post("/feedback", response_model=FeedbackResponse)
def submit_feedback(
req: FeedbackRequest,
username: str = Depends(_get_current_user),
) -> FeedbackResponse:
"""Store user feedback on a prediction and compute the RL reward signal.
Pipeline
--------
1. Compute a ``Β±1`` reward from the user's binary rating.
2. Optionally call an LLM judge (if ``OPENAI_API_KEY`` or
``GEMINI_API_KEY`` is set) and average with the user reward.
3. Persist the feedback event and a corrected training sample to the
``feedback`` / ``experience`` tables.
4. Return the reward so the UI can display it to the user.
"""
llm_r = get_llm_reward(req.text, req.prediction)
reward = compute_combined_reward(req.user_feedback, llm_r)
feedback_id = _feedback_store.save_feedback(
username=username,
text=req.text,
prediction=req.prediction,
user_feedback=req.user_feedback,
reward=reward,
llm_reward=llm_r,
)
return FeedbackResponse(
status="saved",
reward=reward,
llm_reward=llm_r,
feedback_id=feedback_id,
)
@app.get("/feedback/stats", response_model=FeedbackStatsResponse)
def feedback_stats(
username: str = Depends(_get_current_user),
) -> FeedbackStatsResponse:
"""Return aggregated feedback statistics for the authenticated user."""
stats = _feedback_store.get_user_stats(username)
return FeedbackStatsResponse(**stats)
@app.get("/personalization", response_model=PersonalizationResponse)
def personalization(
username: str = Depends(_get_current_user),
) -> PersonalizationResponse:
"""Return a per-user stress-score bias derived from feedback history.
The bias is a small additive correction (β0.1 to +0.1) that shifts the
model's raw prediction toward what past feedback suggests is accurate
for this specific user. A positive bias indicates the model has
historically under-predicted stress for this user; a negative bias
indicates over-prediction.
The correction can be applied at inference time by client code.
"""
stats = _feedback_store.get_user_stats(username)
total = stats["total"]
if total == 0:
return PersonalizationResponse(
user_bias=0.0,
feedback_count=0,
description="No feedback yet β bias is neutral.",
)
# Derive bias: mean_reward of +1 means the model is mostly right (no
# correction needed); mean_reward near -1 means it is mostly wrong.
# We map [-1, +1] β [+0.1, -0.1]: if the model is wrong more often,
# nudge the score up (positive bias) to force the threshold to be met.
mean_r = stats["mean_reward"]
user_bias = round(-mean_r * 0.1, 4)
accuracy_pct = int(stats["accuracy_rate"] * 100)
description = (
f"Based on {total} feedback event(s), model accuracy for you is "
f"~{accuracy_pct}%. Bias adjustment: {user_bias:+.4f}."
)
return PersonalizationResponse(
user_bias=user_bias,
feedback_count=total,
description=description,
)
# ---------------------------------------------------------------------------
# Helpers
# ---------------------------------------------------------------------------
_CHUNK_SIZE = 200
def _simple_tokenize(text: str) -> list[int]:
"""Hash-based tokenization for inference without a stored vocabulary.
Maps each whitespace-delimited token to an index in [1, VOCAB_SIZE-1]
via ``hashlib.md5`` β a fully deterministic hash that produces the
same token IDs on every platform and Python process regardless of
``PYTHONHASHSEED``. Index 0 is reserved for padding.
This must stay in sync with ``_tokenize`` in ``training/train.py``
so that a checkpoint trained on Colab loads and infers correctly on
Windows (or any other machine).
"""
tokens = text.lower().split()
ids = [
int(hashlib.md5(t.encode("utf-8"), usedforsecurity=False).hexdigest(), 16)
% (_DEFAULT_VOCAB_SIZE - 1) + 1
for t in tokens
]
# Pad or truncate to CHUNK_SIZE
if len(ids) < _CHUNK_SIZE:
ids = ids + [0] * (_CHUNK_SIZE - len(ids))
else:
ids = ids[:_CHUNK_SIZE]
return ids
|