the-algorithm / core /analytics.py
github-actions[bot]
deploy: HF sync (Run 194)
1ac9f32
import os
import re
from urllib.parse import urlparse
from itertools import chain
import pandas as pd
import numpy as np
from transformers import pipeline
import torch
import emoji
from collections import Counter
# Pre-compiled Regex Patterns for Analytics (V5.4 Optimization)
# Module-level compilation avoids redundant overhead in high-traffic request cycles.
STRESS_RE = re.compile(r'work|tired|sad|stressed|deadline|exhausted|unhappy|worry|anxious|sick|bad day|hard time')
AFFIRMATIVE_RE = re.compile(r'love|thanks|happy|we|miss|appreciate|glad|proud|beautiful|care')
DISMISSIVE_RE = re.compile(r'whatever|fine|okay|sure|k|ok|busy|tired|idk|anyway')
# Topic-specific regexes
LOGISTICS_RE = re.compile(r'dinner|lunch|bill|home|work|done|todo|buy|shop|cleaning')
EXTERNAL_RE = re.compile(r'friends|party|movie|news|gym|weather|job')
CONFLICT_RE = re.compile(r'sorry|why|fight|angry|stop|listen|mean|hurt|annoyed')
INTIMACY_RE = re.compile(r'love|miss|baby|darling|honey|kiss|hug|beautiful|forever')
BONDING_RE = re.compile(r'miss|love|haha|lol|fun|crazy|remember|bro|dude|bestie')
COLLABORATION_RE = re.compile(r'help|thanks|appreciate|great|good job|team|meeting|sync')
FALLBACK_BONDING_RE = re.compile(r'miss|care|fun')
sentiment_pipeline = None
def get_sentiment_pipeline():
"""Lazy load and quantize the Hinglish sentiment model on CPU."""
global sentiment_pipeline
if sentiment_pipeline is None:
print("Loading and quantizing Hinglish sentiment model...")
model_name = "pascalrai/hinglish-twitter-roberta-base-sentiment"
# Determine device-downloaded model dir (Docker) or fall back to HuggingFace cache (local dev)
model_dir = os.environ.get("MODEL_DIR")
model_kwargs = {"model": model_name, "device": -1}
if model_dir and os.path.isdir(model_dir):
print(f"Loading model from local directory: {model_dir}")
model_kwargs["model"] = model_dir
sentiment_pipeline = pipeline("sentiment-analysis", **model_kwargs)
# Apply dynamic quantization to Linear layers for 50% RAM reduction
sentiment_pipeline.model = torch.quantization.quantize_dynamic(
sentiment_pipeline.model,
{torch.nn.Linear},
dtype=torch.qint8
)
print("Model loaded successfully.")
return sentiment_pipeline
def validate_cloud_url(url: str) -> bool:
"""
Validates that the provided cloud GPU URL is secure and matches the allowed domain.
Prevents SSRF by enforcing HTTPS and restricting to *.lit.ai.
"""
if not url:
return False
try:
# 🛡️ Sentinel: Reject URLs with '@' to prevent credential-based SSRF bypasses
if '@' in url:
return False
parsed = urlparse(url)
# 🛡️ Sentinel: Use hostname instead of netloc to handle ports and auth safely
hostname = parsed.hostname
if not hostname:
return False
# Enforce HTTPS and restrict to Lightning AI domain (*.lit.ai)
if parsed.scheme == 'https' and hostname.endswith('.lit.ai'):
return True
return False
except Exception:
return False
def calculate_latency(df: pd.DataFrame) -> pd.DataFrame:
# DF is already sorted and index reset by run_analytics_pipeline
df['prev_sender'] = df['sender'].shift(1)
df['prev_timestamp'] = df['timestamp'].shift(1)
df['gap_mins'] = (df['timestamp'] - df['prev_timestamp']).dt.total_seconds() / 60.0
# Valid reply: Different sender, gap <= 24 hours (1440 mins)
valid_reply_mask = (df['sender'] != df['prev_sender']) & (df['gap_mins'] <= 1440)
df['latency_mins'] = np.nan
df.loc[valid_reply_mask, 'latency_mins'] = df.loc[valid_reply_mask, 'gap_mins']
# We preserve gap_mins for downstream functions like initiator_ratio and reengagement
df.drop(columns=['prev_sender', 'prev_timestamp'], inplace=True)
return df
def apply_sentiment(df: pd.DataFrame, hf_url: str = "", text_str: pd.Series = None) -> pd.DataFrame:
# We only score PARTNER messages for the risk algorithm
partner_mask = df['sender'] == 'PARTNER'
# ⚡ Bolt Optimization: Use pre-calculated text_str if provided to avoid redundant astype(str)
t_series = text_str if text_str is not None else df.loc[partner_mask, 'text'].astype(str)
partner_msgs_series = t_series[partner_mask].str[:512] if text_str is not None else t_series.str[:512]
partner_msgs = partner_msgs_series.tolist()
sentiment_scores = []
if hf_url:
# 🛡️ Sentinel: Validate URL to prevent SSRF
if not validate_cloud_url(hf_url):
raise ValueError("Security Error: Invalid cloud GPU URL. Must be a secure https://*.lit.ai endpoint.")
print(f"Offloading sentiment analysis of {len(partner_msgs)} messages to Cloud GPU...")
import requests
import concurrent.futures
import time as _time
# Ensure URL has /analyze endpoint precisely once
base_url = hf_url.rstrip('/').replace('/analyze', '')
api_endpoint = base_url + "/analyze"
chunk_size = 1500 # Send in batches of 1500 to prevent payload too large/timeouts
total_chunks = (len(partner_msgs) + chunk_size - 1) // chunk_size
sentiment_scores = [0] * len(partner_msgs)
MAX_RETRIES = 3
BASE_TIMEOUT = 120 # seconds; increased from 90 to handle cold starts
def fetch_chunk(chunk, chunk_index, start_idx):
"""Send a chunk to the Cloud GPU. Retries up to MAX_RETRIES on failure."""
last_error = None
for attempt in range(1, MAX_RETRIES + 1):
timeout = BASE_TIMEOUT + (attempt - 1) * 60 # 120s, 180s, 240s
try:
print(f" Chunk {chunk_index}/{total_chunks} ({len(chunk)} msgs) → Cloud GPU (attempt {attempt}/{MAX_RETRIES}, timeout={timeout}s)...")
response = requests.post(
api_endpoint,
json={"texts": chunk},
headers={"Content-Type": "application/json"},
timeout=timeout,
allow_redirects=False # 🛡️ Sentinel: Prevent SSRF redirect bypass
)
response.raise_for_status()
result = response.json()
if "scores" in result:
print(f" ✓ Chunk {chunk_index}/{total_chunks} completed.")
return start_idx, result["scores"]
else:
raise ValueError(f"Invalid API response format for chunk {chunk_index}: missing 'scores' key")
except Exception as e:
last_error = e
if attempt < MAX_RETRIES:
wait = 5 * attempt
print(f" ✗ Chunk {chunk_index} attempt {attempt} failed ({e}). Retrying in {wait}s...")
_time.sleep(wait)
# All retries exhausted — propagate the error (NO local fallback)
raise RuntimeError(f"Chunk {chunk_index} failed after {MAX_RETRIES} attempts: {last_error}")
chunks_data = []
for i in range(0, len(partner_msgs), chunk_size):
chunk = partner_msgs[i:i + chunk_size]
chunk_index = (i // chunk_size) + 1
chunks_data.append((chunk, chunk_index, i))
try:
with concurrent.futures.ThreadPoolExecutor(max_workers=5) as executor:
futures = [executor.submit(fetch_chunk, c, ci, si) for c, ci, si in chunks_data]
for future in concurrent.futures.as_completed(futures):
start_idx, scores = future.result()
for idx, score in enumerate(scores):
if start_idx + idx < len(sentiment_scores):
sentiment_scores[start_idx + idx] = score
except Exception as e:
print(f"CRITICAL: Cloud GPU offload failed: {e}.")
# No local fallback when a cloud URL is provided — this prevents 70k+ messages from locking up the CPU.
raise RuntimeError(f"Cloud Sentiment Analysis Failed: {e}. Check your Lightning Studio instance.")
else:
# ONLY run local scoring if NO cloud URL was provided at all
if partner_msgs:
pipe = get_sentiment_pipeline()
batch_size = 32
print(f"Scoring {len(partner_msgs)} messages locally (pipeline batch_size={batch_size})...")
try:
# Performance Optimization (V5.2): Leverage Transformers native batching.
results = pipe(partner_msgs, batch_size=batch_size)
# Performance Optimization (V5.3): Replaced multiple str.contains scans
# and np.select with a direct O(1) dictionary map for label-to-score conversion.
labels = pd.Series([r['label'].lower() for r in results])
label_map = {
'label_0': -1, 'negative': -1,
'label_1': 0, 'neutral': 0,
'label_2': 1, 'positive': 1
}
sentiment_scores = labels.map(label_map).fillna(0).astype(int).tolist()
except Exception as e:
print(f"Local sentiment analysis failed: {e}")
sentiment_scores = [0] * len(partner_msgs)
df['sentiment'] = 0
if sentiment_scores:
df.loc[partner_mask, 'sentiment'] = sentiment_scores
return df
def aggregate_weekly(df: pd.DataFrame) -> pd.DataFrame:
# Anchor to Monday - Vectorized
df['week_start'] = df['timestamp'].dt.to_period('W').dt.start_time
# Pre-calculate filtered sentiment for vectorized aggregation
df['_partner_sent'] = df['sentiment'].where(df['sender'] == 'PARTNER')
# Aggregate using vectorized .agg() instead of slow .apply()
weekly = df.groupby('week_start').agg(
volume=('sentiment', 'size'),
median_latency=('latency_mins', 'median'),
mean_sentiment=('_partner_sent', 'mean')
).reset_index()
# Clean up temporary column
df.drop(columns=['_partner_sent'], inplace=True)
weekly.fillna({'median_latency': 0, 'mean_sentiment': 0}, inplace=True)
return weekly
def calculate_emoji_frequency(df: pd.DataFrame, text_str: pd.Series = None) -> dict:
"""Extract top-10 emoji usage per sender. Must be called BEFORE privacy firewall drops text."""
result = {}
# Use pre-calculated string series if provided to avoid redundant astype(str)
t_series = text_str if text_str is not None else df['text'].astype(str)
for sender in ['ME', 'PARTNER']:
mask = df['sender'] == sender
# Performance Optimization (V5.1): Use itertools.chain with Counter to eliminate
# the manual Python loop. This delegates character-level iteration to C-level
# routines while maintaining O(N_unique_chars) calls to emoji.is_emoji().
counts_all = Counter(chain.from_iterable(t_series[mask]))
emoji_counts = {char: count for char, count in counts_all.items() if emoji.is_emoji(char)}
counts = Counter(emoji_counts).most_common(10)
result[sender] = [{'emoji': e, 'count': c} for e, c in counts]
return result
def calculate_initiator_ratio(df: pd.DataFrame) -> dict:
"""Count conversation initiations. An initiation = message after a >=4 hour gap."""
# Optimization: DF is already sorted by calculate_latency at the start of the pipeline
if len(df) < 2:
return {'me_initiations': 0, 'partner_initiations': 0, 'me_ratio': 0.0}
gap_threshold_mins = 240 # 4 hours
# First message is always an initiation
initiation_mask = (df['gap_mins'] >= gap_threshold_mins) | (df.index == 0)
initiations = df.loc[initiation_mask]
me_count = int((initiations['sender'] == 'ME').sum())
partner_count = int((initiations['sender'] == 'PARTNER').sum())
total = me_count + partner_count
return {
'me_initiations': me_count,
'partner_initiations': partner_count,
'me_ratio': round(me_count / total, 4) if total > 0 else 0.0
}
def calculate_risk_score(weekly_df: pd.DataFrame) -> pd.DataFrame:
if weekly_df.empty: return weekly_df
# Sentiment: -1 (bad) to 1 (good). Inverted: 1 (high risk) to 0 (low risk)
weekly_df['sentiment_inv'] = (1 - weekly_df['mean_sentiment']) / 2.0
# Latency: Normalize 0 to 1
max_lat = weekly_df['median_latency'].max()
min_lat = weekly_df['median_latency'].min()
if max_lat > min_lat:
weekly_df['latency_norm'] = (weekly_df['median_latency'] - min_lat) / (max_lat - min_lat)
else:
weekly_df['latency_norm'] = 0
# Volume: Normalize and Invert
max_vol = weekly_df['volume'].max()
min_vol = weekly_df['volume'].min()
if max_vol > min_vol:
vol_norm = (weekly_df['volume'] - min_vol) / (max_vol - min_vol)
weekly_df['volume_inv'] = 1.0 - vol_norm
else:
weekly_df['volume_inv'] = 0
# Formula from PRD 2.0
weekly_df['risk_score'] = (0.5 * weekly_df['sentiment_inv']) + (0.3 * weekly_df['latency_norm']) + (0.2 * weekly_df['volume_inv'])
# Round metrics for clean UI
weekly_df['risk_score'] = weekly_df['risk_score'].round(4)
weekly_df['mean_sentiment'] = weekly_df['mean_sentiment'].round(4)
weekly_df['median_latency'] = weekly_df['median_latency'].round(2)
return weekly_df
def detect_risk_phases(weekly_df: pd.DataFrame) -> pd.DataFrame:
"""Label each week with a relationship phase based on risk score."""
def _phase(score):
if score < 0.3: return 'Honeymoon'
elif score < 0.6: return 'Stable'
elif score < 0.85: return 'Tension'
else: return 'Danger'
if not weekly_df.empty:
weekly_df['phase'] = weekly_df['risk_score'].apply(_phase)
return weekly_df
def calculate_power_dynamics(df: pd.DataFrame, text_str: pd.Series = None) -> dict:
"""Calculate the Word Count ratio to establish Power Dynamics (V3.0)."""
if 'text' not in df.columns: return {}
# Optimization: Use str.count for faster vectorized word counting
# Use pre-calculated string series if provided
t_series = text_str if text_str is not None else df['text'].astype(str)
df['word_count'] = t_series.str.count(r'\S+')
counts = df.groupby('sender')['word_count'].sum().to_dict()
me_words = int(counts.get('ME', 0))
partner_words = int(counts.get('PARTNER', 0))
# Ratio: ME / PARTNER. If > 1, ME is dominating the conversation volume.
ratio = float(round(me_words / partner_words, 2)) if partner_words > 0 else 0.0
return {
'me_word_count': me_words,
'partner_word_count': partner_words,
'power_ratio': ratio
}
def calculate_affection_friction(df: pd.DataFrame, text_lower: pd.Series = None) -> dict:
"""Detect 'Burnout' via affirmative vs dismissive language trends (V3.0)."""
if 'text' not in df.columns: return {}
# Use pre-calculated lowercased series if provided
text_lower = text_lower if text_lower is not None else df['text'].astype(str).str.lower()
# Performance Optimization (V5.4): Use pre-compiled module-level regexes.
aff_count = text_lower.str.contains(AFFIRMATIVE_RE).sum()
dis_count = text_lower.str.contains(DISMISSIVE_RE).sum()
return {
'affirmative_count': int(aff_count),
'dismissive_count': int(dis_count)
}
def calculate_support_gap(df: pd.DataFrame, text_lower: pd.Series = None, text_str: pd.Series = None) -> dict:
"""Identify stress messages and measure partner's response quality (V4.0)."""
if 'text' not in df.columns or len(df) < 5: return {}
# Use input df directly as it is already sorted
df_temp = df
# Use pre-calculated series if provided
t_lower = text_lower if text_lower is not None else df_temp['text'].astype(str).str.lower()
# Performance Optimization (V5.4): Use pre-compiled STRESS_RE.
# Vectorized stress detection outside the loop is much faster
is_stress = t_lower.str.contains(STRESS_RE).values
# Performance Optimization (V5.3): Refactored the Python loop to use integer indexing
# and NumPy-native state tracking. This eliminates multiple dictionary lookups
# and string key overhead in every iteration of the hot loop (O(N)).
senders = df_temp['sender'].values
# ME -> 0, PARTNER -> 1
s_idx = (senders == 'PARTNER').astype(np.int8)
timestamps = df_temp['timestamp'].values
# Pre-calculate message lengths to avoid calling len() in the loop
text_lens = text_str.str.len().values if text_str is not None else df_temp['text'].astype(str).str.len().values
# State tracking using arrays (index 0: ME, index 1: PARTNER)
stress_counts = np.zeros(2, dtype=np.int32)
support_received = np.zeros(2, dtype=np.int32)
active_stress_ts = np.full(2, np.datetime64('NaT'), dtype=timestamps.dtype)
# Comparison threshold for response time
threshold = np.timedelta64(60, 'm')
for i in range(len(s_idx)):
s = s_idx[i]
ts = timestamps[i]
# Did this person just send a stress message?
if is_stress[i]:
stress_counts[s] += 1
active_stress_ts[s] = ts
# Did this person just respond to the OTHER person's stress message?
other_s = 1 - s # Flip 0 to 1, 1 to 0
ast = active_stress_ts[other_s]
if not np.isnat(ast):
# Direct comparison of timedeltas avoids division overhead
if (ts - ast) <= threshold and text_lens[i] > 10:
support_received[other_s] += 1
# Clear their stress state so we don't double count
active_stress_ts[other_s] = np.datetime64('NaT')
return {
'ME': {'stress_count': int(stress_counts[0]), 'support_received': int(support_received[0])},
'PARTNER': {'stress_count': int(stress_counts[1]), 'support_received': int(support_received[1])}
}
def calculate_reengagement(df: pd.DataFrame) -> dict:
"""Detect who reaches out first after a long silence (> 24h) (V4.0)."""
# Optimization: DF is already sorted
if len(df) < 10: return {}
# We already have gap_mins from calculate_latency
# Long silence = gap > 24 hours (1440 mins)
reengagements = df[df['gap_mins'] > 1440]
counts = reengagements['sender'].value_counts().to_dict()
return {
'me_reengagements': int(counts.get('ME', 0)),
'partner_reengagements': int(counts.get('PARTNER', 0))
}
def calculate_linguistic_mirroring(df: pd.DataFrame, text_lower: pd.Series = None) -> dict:
"""Measure how frequently partners adopt each others vocabulary (V4.0)."""
if 'text' not in df.columns or len(df) < 100:
return {}
# Simplified approach: Look for rare punctuation/emoji habits or unique high-frequency words
punctuation_habits = ['!!!', '...', '??', 'haha', 'lol', 'lmao']
results = {}
# Optimization: Use vectorized .str.contains().any() to avoid massive string joins
# Joining 100k messages into one string causes major memory spikes and slow search.
# Use pre-calculated lowercased series if provided
text_lower = text_lower if text_lower is not None else df['text'].astype(str).str.lower()
# Pre-calculate habit presence for each sender using vectorized operations
habit_presence = {}
for sender in ['ME', 'PARTNER']:
sender_mask = df['sender'] == sender
sender_msgs = text_lower[sender_mask]
habit_presence[sender] = {
habit: sender_msgs.str.contains(habit, regex=False).any()
for habit in punctuation_habits
}
for sender in ['ME', 'PARTNER']:
other = 'PARTNER' if sender == 'ME' else 'ME'
mirror_score = sum(
1 for habit in punctuation_habits
if habit_presence[sender][habit] and habit_presence[other][habit]
)
results[f"{sender}_mirroring"] = mirror_score
return results
def calculate_topic_mix(df: pd.DataFrame, connection_type: str, text_lower: pd.Series = None) -> dict:
"""Categorize conversation dynamically based on connection type (V4.0)."""
if 'text' not in df.columns: return {}
# Performance Optimization (V5.4): Use pre-compiled module-level regexes.
if connection_type == 'romantic':
categories = {'Logistics': LOGISTICS_RE, 'Intimacy': INTIMACY_RE, 'Conflict': CONFLICT_RE, 'External': EXTERNAL_RE}
elif connection_type in ['friendship', 'casual', 'family']:
categories = {'Logistics': LOGISTICS_RE, 'Bonding': BONDING_RE, 'Disagreement': CONFLICT_RE, 'External': EXTERNAL_RE}
elif connection_type == 'professional':
categories = {'Operations': LOGISTICS_RE, 'Collaboration': COLLABORATION_RE, 'Blockers': CONFLICT_RE, 'External': EXTERNAL_RE}
else:
categories = {'Logistics': LOGISTICS_RE, 'Bonding': FALLBACK_BONDING_RE, 'Conflict': CONFLICT_RE, 'External': EXTERNAL_RE}
# Use pre-calculated lowercased series if provided
text_lower = text_lower if text_lower is not None else df['text'].astype(str).str.lower()
results = {}
for cat, regex in categories.items():
results[cat] = int(text_lower.str.contains(regex).sum())
return results
def run_analytics_pipeline(df: pd.DataFrame, hf_url: str = "", connection_type: str = "romantic") -> dict:
"""Runs the full analytics pipeline and returns a dict with weekly stats, emoji freq, and initiator ratio."""
# ⚡ Bolt Optimization (V5.4): Reset index at entry to ensure alignment for pre-calculated
# series and remove redundant O(N) reset_index calls from downstream functions.
# DF is already sorted by timestamp in app.py.
df = df.reset_index(drop=True)
# ⚡ Bolt Optimization: Pre-calculate common series once at the pipeline entry
# to avoid redundant O(N) operations across multiple analytics functions.
text_str = df['text'].astype(str)
text_lower = text_str.str.lower()
df = calculate_latency(df)
df = apply_sentiment(df, hf_url=hf_url, text_str=text_str)
# Phase 6: Extract enhanced features BEFORE privacy firewall
emoji_freq = calculate_emoji_frequency(df, text_str=text_str)
initiator_ratio = calculate_initiator_ratio(df)
# Phase 8 (V3.0): Power Dynamics & Burnout NLP
power_dynamics = calculate_power_dynamics(df, text_str=text_str)
affection_friction = calculate_affection_friction(df, text_lower=text_lower)
# Phase 11 (V4.0): Advanced Personalization
support_gap = calculate_support_gap(df, text_lower=text_lower, text_str=text_str)
reengagement = calculate_reengagement(df)
mirroring = calculate_linguistic_mirroring(df, text_lower=text_lower)
topic_mix = calculate_topic_mix(df, connection_type, text_lower=text_lower)
# Privacy handling: text is needed for flashbacks in app.py, so we don't drop it here anymore.
# The app.py will handle the session storage and eventual purging.
weekly_df = aggregate_weekly(df)
weekly_df = calculate_risk_score(weekly_df)
weekly_df = detect_risk_phases(weekly_df)
# Format date for JSON
weekly_df['week_start'] = weekly_df['week_start'].dt.strftime('%Y-%m-%d')
return {
'weekly': weekly_df.to_dict(orient='records'),
'emoji_freq': emoji_freq,
'initiator_ratio': initiator_ratio,
'power_dynamics': power_dynamics,
'affection_friction': affection_friction,
'support_gap': support_gap,
'reengagement': reengagement,
'mirroring': mirroring,
'topic_mix': topic_mix
}