Emotional Context Tracking

from dakera import DakeraClient
from dataclasses import dataclass
from typing import Optional
import re

client = DakeraClient(base_url="http://localhost:3000", api_key="dk-...")

@dataclass
class EmotionSignal:
    sentiment: str          # "positive" | "neutral" | "negative"
    emotion: str            # "frustrated" | "distressed" | "satisfied" | "excited"
    intensity: float        # 0.0 - 1.0
    trigger: Optional[str]  # what caused it

def detect_emotion(message: str) -> EmotionSignal:
    """Lightweight heuristic emotion detector."""
    message_lower = message.lower()

    # Frustration signals
    frustration_signals = [
        "still broken", "still not working", "weeks of this", "i give up",
        "this is ridiculous", "unacceptable", "so frustrating", "again?!"
    ]
    distress_signals = [
        "i can't take", "desperate", "need help now", "this is destroying",
        "really struggling", "at my limit", "please help me"
    ]
    positive_signals = [
        "thank you", "that worked", "perfect", "exactly what i needed",
        "amazing", "finally", "great job", "love this"
    ]

    frust_score = sum(1 for s in frustration_signals if s in message_lower)
    dist_score = sum(1 for s in distress_signals if s in message_lower)
    pos_score = sum(1 for s in positive_signals if s in message_lower)

    # Capitals and punctuation amplify intensity
    caps_ratio = sum(1 for c in message if c.isupper()) / max(len(message), 1)
    exclamations = message.count('!')

    if dist_score > 0:
        return EmotionSignal("negative", "distressed", min(0.9 + caps_ratio, 1.0), "user_message")
    elif frust_score > 0:
        intensity = min(0.6 + frust_score * 0.1 + caps_ratio * 0.2, 0.95)
        return EmotionSignal("negative", "frustrated", intensity, "user_message")
    elif pos_score > 0:
        return EmotionSignal("positive", "satisfied", 0.7 + pos_score * 0.05, None)
    else:
        return EmotionSignal("neutral", "neutral", 0.3, None)

def store_emotional_observation(agent_id: str, user_id: str, observation: str, signal: EmotionSignal) -> None:
    """Persist emotional state observation with appropriate importance and TTL."""
    # Higher importance for negative/intense signals
    importance_base = {"positive": 0.55, "neutral": 0.45, "negative": 0.80}
    importance = min(importance_base[signal.sentiment] + signal.intensity * 0.1, 0.95)

    client.store_memory(
        agent_id=agent_id,
        content=f"Emotional observation: {observation}. Emotion: {signal.emotion}. Intensity: {signal.intensity:.2f}.",
        importance=importance,
        memory_type="episodic",
        tags=[f"emotion:{signal.emotion}", f"sentiment:{signal.sentiment}", f"intensity:{'high' if signal.intensity > 0.7 else 'medium'}"],
        ttl_seconds=604_800  # 7 days
    )

def get_emotional_trajectory(agent_id: str) -> dict:
    """Recall recent emotional history and compute trajectory score."""
    result = client.recall(
        agent_id=agent_id,
        query="user emotional state frustration distress satisfaction",
        top_k=7,
        min_importance=0.4
    )

    memories = result.get("memories", [])
    if not memories:
        return {"trajectory_score": 0.0, "dominant_emotion": "neutral", "escalate": False}

    # Weighted average: most recent = highest weight
    total_weight = 0.0
    weighted_score = 0.0
    sentiment_map = {"positive": 1.0, "neutral": 0.0, "negative": -1.0}

    for i, mem in enumerate(memories):
        weight = 1.0 / (i + 1)  # recency decay: index 0 gets weight 1.0
        meta = mem.get("metadata", {})
        sentiment = meta.get("sentiment", "neutral")
        intensity = meta.get("intensity", 0.5)
        score = sentiment_map.get(sentiment, 0.0) * intensity
        weighted_score += score * weight
        total_weight += weight

    trajectory = weighted_score / total_weight if total_weight > 0 else 0.0

    # Count consecutive negative sessions for escalation trigger
    consecutive_neg = 0
    for mem in memories:
        if mem.get("metadata", {}).get("sentiment") == "negative":
            consecutive_neg += 1
        else:
            break

    return {
        "trajectory_score": round(trajectory, 3),
        "dominant_emotion": memories[0].get("metadata", {}).get("emotion", "neutral") if memories else "neutral",
        "escalate": trajectory < -0.65 and consecutive_neg >= 3,
        "consecutive_negative": consecutive_neg,
        "sample_count": len(memories)
    }

def build_emotional_context_prompt(trajectory: dict) -> str:
    """Generate a system prompt note from emotional trajectory data."""
    score = trajectory["trajectory_score"]
    emotion = trajectory["dominant_emotion"]
    escalate = trajectory["escalate"]

    if escalate:
        return (
            "ESCALATION ALERT: This user has shown sustained distress across multiple sessions "
            f"(trajectory score: {score:.2f}). Immediately acknowledge their difficulties, express genuine empathy, "
            "and offer to connect them with a specialist or supervisor. Do not attempt to solve the issue independently."
        )
    elif score < -0.4:
        return (
            f"Emotional context: User is currently {emotion} (trajectory: {score:.2f}). "
            "Adopt a warm, patient tone. Explicitly acknowledge their frustration before any technical response. "
            "Avoid efficiency-first language. Focus on validation first, solution second."
        )
    elif score > 0.4:
        return (
            f"Emotional context: User is in a positive state (trajectory: {score:.2f}). "
            "Engage warmly, match their energy. This is a good moment to introduce new features or offer additional help."
        )
    else:
        return "Emotional context: User appears to be in a neutral state. Maintain a professional, helpful tone."

# --- Usage in a support agent ---
def handle_support_message(user_id: str, message: str) -> str:
    agent_id = f"support-agent-{user_id}"

    # 1. Detect emotion in incoming message
    signal = detect_emotion(message)

    # 2. Store the observation
    store_emotional_observation(
        agent_id, user_id,
        f"User message indicates {signal.emotion}",
        signal
    )

    # 3. Compute trajectory from history
    trajectory = get_emotional_trajectory(agent_id)

    # 4. Build context-aware system prompt
    emotional_note = build_emotional_context_prompt(trajectory)

    # 5. Use in your LLM call
    # system_prompt = f"{BASE_SYSTEM_PROMPT}

{emotional_note}"
    # response = llm.complete(system=system_prompt, user=message)

    if trajectory["escalate"]:
        return trigger_human_escalation(user_id, trajectory)

    return emotional_note  # For demonstration

def trigger_human_escalation(user_id: str, trajectory: dict) -> str:
    agent_id = f"support-agent-{user_id}"
    client.store_memory(
        agent_id=agent_id,
        content=f"Escalated to human support. Trajectory score: {trajectory['trajectory_score']:.2f}. Consecutive negative sessions: {trajectory['consecutive_negative']}.",
        importance=0.95,
        memory_type="episodic",
        tags=["escalation", "human-handoff"]
    )
    return "Routing to human specialist..."

import { DakeraClient } from '@dakera-ai/dakera';

const client = new DakeraClient({ baseUrl: 'http://localhost:3000', apiKey: 'dk-...' });

type Sentiment = 'positive' | 'neutral' | 'negative';
type Emotion = 'frustrated' | 'distressed' | 'satisfied' | 'excited' | 'neutral';

interface EmotionSignal {
  sentiment: Sentiment;
  emotion: Emotion;
  intensity: number;
  trigger?: string;
}

function detectEmotion(message: string): EmotionSignal {
  const lower = message.toLowerCase();
  const frustrationPatterns = ['still broken', 'weeks of this', 'i give up', 'ridiculous', 'unacceptable'];
  const distressPatterns = ['desperate', 'need help now', 'really struggling', 'at my limit'];
  const positivePatterns = ['thank you', 'that worked', 'perfect', 'amazing', 'love this'];

  const frustScore = frustrationPatterns.filter(p => lower.includes(p)).length;
  const distScore = distressPatterns.filter(p => lower.includes(p)).length;
  const posScore = positivePatterns.filter(p => lower.includes(p)).length;

  if (distScore > 0) return { sentiment: 'negative', emotion: 'distressed', intensity: 0.92 };
  if (frustScore > 0) return { sentiment: 'negative', emotion: 'frustrated', intensity: Math.min(0.6 + frustScore * 0.1, 0.95) };
  if (posScore > 0) return { sentiment: 'positive', emotion: 'satisfied', intensity: 0.75 };
  return { sentiment: 'neutral', emotion: 'neutral', intensity: 0.3 };
}

async function storeEmotionalObservation(agentId: string, observation: string, signal: EmotionSignal): Promise<void> {
  const importanceBase: Record<Sentiment, number> = { positive: 0.55, neutral: 0.45, negative: 0.80 };
  const importance = Math.min(importanceBase[signal.sentiment] + signal.intensity * 0.1, 0.95);

  await client.storeMemory(agentId, {
    content: `Emotional observation: ${observation}. Emotion: ${signal.emotion}. Intensity: ${signal.intensity.toFixed(2)}.`,
    importance,
    memoryType: 'episodic',
    tags: [`emotion:${signal.emotion}`, `sentiment:${signal.sentiment}`],
    ttl_seconds: 604_800, // 7 days
  });
}

async function getEmotionalTrajectory(agentId: string): Promise<{
  trajectoryScore: number;
  dominantEmotion: string;
  escalate: boolean;
  consecutiveNegative: number;
}> {
  const result = await client.recall(agentId, 'user emotional state frustration distress', { top_k: 7, min_importance: 0.4 });

  const sentimentMap: Record<Sentiment, number> = { positive: 1.0, neutral: 0.0, negative: -1.0 };
  let weightedScore = 0, totalWeight = 0;

  for (let i = 0; i < result.memories.length; i++) {
    const weight = 1 / (i + 1);
    const sentiment = (result.memories[i].metadata?.sentiment as Sentiment) ?? 'neutral';
    const intensity = (result.memories[i].metadata?.intensity as number) ?? 0.5;
    weightedScore += sentimentMap[sentiment] * intensity * weight;
    totalWeight += weight;
  }

  const trajectoryScore = totalWeight > 0 ? weightedScore / totalWeight : 0;

  let consecutiveNegative = 0;
  for (const mem of result.memories) {
    if (mem.metadata?.sentiment === 'negative') consecutiveNegative++;
    else break;
  }

  return {
    trajectoryScore: Math.round(trajectoryScore * 1000) / 1000,
    dominantEmotion: result.memories[0]?.metadata?.emotion as string ?? 'neutral',
    escalate: trajectoryScore < -0.65 && consecutiveNegative >= 3,
    consecutiveNegative,
  };
}

// Compose system prompt from trajectory
function buildEmotionalContextNote(trajectory: Awaited<ReturnType<typeof getEmotionalTrajectory>>): string {
  if (trajectory.escalate) {
    return `ESCALATION ALERT: User in sustained distress (score: ${trajectory.trajectoryScore}). Route to human specialist immediately.`;
  }
  if (trajectory.trajectoryScore < -0.4) {
    return `Emotional context: User is ${trajectory.dominantEmotion} (score: ${trajectory.trajectoryScore}). Acknowledge frustration before providing technical help. Be warm and patient.`;
  }
  return `Emotional context: User appears ${trajectory.dominantEmotion}. Maintain professional, helpful tone.`;
}

use dakera_rs::{Client, StoreMemoryRequest, RecallRequest};

let client = Client::new("http://localhost:3000", "dk-...");

// Store a negative emotional observation
client.store_memory("support-agent-user-42", StoreMemoryRequest {
    content: "Emotional observation: User expressed sustained frustration about recurring billing issue. Emotion: frustrated. Intensity: 0.82.".into(),
    importance: Some(0.88),
    memory_type: "episodic".into(),
    tags: vec!["emotion:frustrated".into(), "sentiment:negative".into()],
    ttl_seconds: Some(604_800),
    ..Default::default()
}).await?;

// Recall emotional history and compute trajectory
let history = client.recall("support-agent-user-42", RecallRequest {
    query: "user emotional state frustration distress satisfaction".into(),
    top_k: Some(7),
    min_importance: Some(0.4),
    ..Default::default()
}).await?;

// Compute weighted trajectory score
let mut weighted_score = 0.0f64;
let mut total_weight = 0.0f64;
for (i, mem) in history.memories.iter().enumerate() {
    let weight = 1.0 / (i as f64 + 1.0);
    let sentiment = mem.metadata.as_ref()
        .and_then(|m| m.get("sentiment"))
        .and_then(|v| v.as_str())
        .unwrap_or("neutral");
    let intensity = mem.metadata.as_ref()
        .and_then(|m| m.get("intensity"))
        .and_then(|v| v.as_f64())
        .unwrap_or(0.5);
    let score = match sentiment {
        "positive" => 1.0,
        "negative" => -1.0,
        _ => 0.0,
    } * intensity;
    weighted_score += score * weight;
    total_weight += weight;
}

let trajectory = if total_weight > 0.0 { weighted_score / total_weight } else { 0.0 };
let escalate = trajectory < -0.65;
println!("Trajectory score: {:.3}, escalate: {}", trajectory, escalate);

client := dakera.NewClient("http://localhost:3000", "dk-...")
ctx := context.Background()

// Store emotional observation
client.StoreMemory(ctx, "support-agent-user-42", dakera.StoreMemoryRequest{
    Content:    "Emotional observation: User frustrated about recurring billing issue. Intensity: 0.82.",
    Importance: 0.88,
    MemoryType: "episodic",
    Tags:       []string{"emotion:frustrated", "sentiment:negative"},
    TTLSeconds: 604_800,
    Metadata: map[string]interface{}{
        "sentiment": "negative",
        "emotion":   "frustrated",
        "intensity":  0.82,
    },
})

// Recall and compute trajectory
result, _ := client.Recall(ctx, "support-agent-user-42", dakera.RecallRequest{
    Query:         "user emotional state frustration distress",
    TopK:          7,
    MinImportance: 0.4,
})

sentimentMap := map[string]float64{"positive": 1.0, "neutral": 0.0, "negative": -1.0}
var weightedScore, totalWeight float64

for i, mem := range result.Memories {
    weight := 1.0 / float64(i+1)
    sentiment, _ := mem.Metadata["sentiment"].(string)
    intensity, _ := mem.Metadata["intensity"].(float64)
    score := sentimentMap[sentiment] * intensity
    weightedScore += score * weight
    totalWeight += weight
}

trajectory := 0.0
if totalWeight > 0 {
    trajectory = weightedScore / totalWeight
}
fmt.Printf("Trajectory: %.3f, Escalate: %v
", trajectory, trajectory < -0.65)