🤖 AI

PulseCX: Breaking the Closed-World Assumption in Real-Time CX

PulseCX is a novel framework that overcomes the closed-world limitations of real-time conversational AI by decoupling knowledge acquisition from consumption through an asynchronous agent that builds a decay-aware temporal knowledge graph, thereby eliminating search latency and significantly improving intent resolution and customer satisfaction in dynamic environments.

Original authors: Rajat Agarwal, Suvidha Tripathi, Shubham Sharma

Published 2026-06-23

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Rajat Agarwal, Suvidha Tripathi, Shubham Sharma

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are working as a customer service representative for a huge company. You have a thick, heavy manual on your desk that lists every product, policy, and answer you know. This manual is perfect for normal days. But, the world outside is chaotic and fast-moving. A viral meme just made a specific product famous, a server just crashed, or a competitor just launched a flash sale.

The problem with most AI customer service bots today is that they are like you, but they only have the manual. They don't know what's happening outside their office walls. If a customer asks, "Why is everyone talking about the 'Glazed Donut Stick'?" the bot looks at its manual, sees nothing about donuts, and says, "I don't know what that is." The customer gets frustrated and leaves. This is what the paper calls the "Closed-World Constraint."

The authors of this paper, Rajat Agarwal and his team at Sprinklr, built a new system called PulseCX to fix this. Here is how it works, explained simply:

1. The Problem: The "Live Search" Trap

You might think, "Why doesn't the bot just Google the answer right when the customer asks?"
The paper says this is a bad idea for two reasons:

It's too slow: If the bot has to stop, go search the internet, read the results, and then answer, the customer has to wait too long (like waiting 1.5 seconds in a real-time chat).
It's dangerous: The internet is full of noise, rumors, and fake news. If the bot reads a fake rumor and tells the customer, it creates "context poisoning." It might give wrong advice or look stupid.

2. The Solution: PulseCX (The "Newsroom" Approach)

Instead of searching the internet when a customer asks a question, PulseCX does the searching before the customer even shows up.

Think of PulseCX as a 24-hour newsroom working in the background while the customer service agents are on break.

The Asynchronous Agent (The Scout): A special AI agent (called the Social Search Agent) constantly scans social media and news sites every 15 minutes. It doesn't wait for a question; it just watches the world.
The "Decay-Aware" Memory (The Smart Filing Cabinet): When the Scout finds something important (like a real outage or a viral trend), it doesn't just dump the raw text into a database. It turns it into a clean, structured "Context Object" (like a press release).
- The "Decay" Magic: This is the clever part. The system knows that some news is fleeting (like a flash sale that ends in an hour) and some is permanent (like a server outage).
- If a trend is viral, the system gives it a high "weight" (importance).
- As time passes, if the news isn't confirmed again, the system automatically lowers its weight (it "decays").
- Analogy: Imagine a whiteboard in the office. If a rumor is just a whisper, it fades away quickly. If it's a confirmed fact, someone keeps writing it in bold marker. If the event is over, someone erases it. This prevents the bot from talking about a sale that ended yesterday.

3. How It Answers You (The "Cheat Sheet")

When a customer finally chats with the bot:

The Gatekeeper: The bot first checks: "Is this a boring question (like 'What's my password?') or a dynamic one (like 'Why is the site slow?')"
The Quick Grab: If it's a dynamic question, the bot doesn't search the web. Instead, it instantly grabs the pre-made "Cheat Sheet" (Context Card) from its smart filing cabinet.
The Answer: The bot reads the Cheat Sheet, which says: "Hey, there is a viral trend about 'Glazed Donut Sticks'—that's actually our Lip Balm. Also, there is an outage in Texas."
The Result: The bot answers instantly (in milliseconds) with the correct, up-to-date info, sounding like it knows exactly what's happening.

Why This Matters (The Results)

The authors tested this in a simulation with 1,000 customer questions. Here is what happened:

Old Way (Static Manual): The bot was fast but often wrong or unhelpful for trending topics. Customers got frustrated (Low Satisfaction).
Naive Way (Live Search): The bot knew the answers but took too long to find them. Customers got annoyed waiting.
PulseCX: The bot was fast (almost as fast as the static one) but smart (it knew the viral trends and outages).
- It solved 89% of the tricky questions (compared to 64% for the old way).
- Customers were much happier because the bot validated their feelings (e.g., "Yes, your internet is down, we see the reports too") instead of gaslighting them ("Our system says everything is fine").

In a Nutshell

PulseCX stops the AI from being stuck in the past. Instead of panicking and searching the internet every time a customer asks a question, it has a dedicated team constantly updating its "brain" in the background, filtering out the noise, and keeping only the fresh, relevant facts ready to go. This makes the AI feel alive, responsive, and trustworthy without slowing down the conversation.

Technical Summary: PulseCX – Breaking the Closed-World Assumption in Real-Time CX

Problem Statement: The Closed-World Constraint

Current Conversational AI agents in Customer Experience (CX) operate under a fundamental Closed-World Constraint. These agents are grounded in static internal artifacts (manuals, FAQs, dashboards) while operating in a reality shaped by high-velocity external events (viral trends, outages, PR crises). This misalignment leads to "context blindness" regarding Dynamic intents, where successful resolution depends on time-sensitive external knowledge.

Attempts to bridge this gap via synchronous web search or news APIs introduce prohibitive latency and severe context poisoning. Flooding the inference window with noisy, contradictory, or dangerous signals can compromise security and degrade performance. The core challenge is to inject real-time situational context without sacrificing the low latency and reliability required for real-time interaction.

Methodology: The PulseCX Framework

PulseCX addresses this challenge by decoupling knowledge acquisition from consumption, adopting a structure-first paradigm. The framework consists of two distinct stages:

1. Offline Asynchronous Stage: World Modeling

This stage operates independently of user queries to structure world knowledge into a Decay-Aware Temporal Knowledge Graph (DA-TKG).

Social Search Agent (SSA): A periodic agent (e.g., every 15 minutes) scans external sources (social media, news) for high-velocity signals.
Linearization & Context Object Creation: Raw, unstructured signals are immediately linearized into a structured Context Object ( $O = \{Type, Scope, Confidence, State\}$ ). This ensures only high-quality, provenance-aware signals enter the system.
Memory Evolution (Reinforcement-Decay): The DA-TKG is self-evolving. Node weights ( $W_t$ ) decay over time ( $W_t = W_{last} \cdot e^{-\lambda(t-t_{last})}$ ) to prune stale context (e.g., expired sales). Conversely, signals re-confirmed by the SSA are reinforced ( $\beta$ ), amplifying persistent events (e.g., infrastructure outages) while allowing transient noise to fade.

2. Online Synchronous Stage: Grounding

When a user queries the system, PulseCX executes a high-speed retrieval pipeline to inject relevant context.

Hierarchical Intent Gating: The system classifies user intent. Static intents (e.g., "Reset Password") bypass the graph to prevent context poisoning. Dynamic intents (e.g., "Why is the site slow?") trigger retrieval.
Salience-Weighted Semantic Reranker: Candidate nodes are scored by combining semantic similarity to the query with their current decay-adjusted weight ( $R(N) = Sim(Q, N) \cdot W_t$ ). This prioritizes nodes that are both relevant and temporally active.
Context Card Synthesis: Top-ranked nodes are synthesized into a deterministic Context Card (a structured briefing) injected into the agent's system prompt. This card includes specific "Playbook" logic based on event type (e.g., "Priority: External Signal > Internal Dashboard" for outages).
Feedback Loop: A Query Refiner analyzes chat logs for information gaps, generating targeted search queries for the next offline SSA cycle.

Key Contributions

Decoupled Architecture: PulseCX separates high-latency intelligence gathering (Offline) from real-time context injection (Online), eliminating the latency penalties of synchronous web search.
Decay-Aware Temporal Knowledge Graph (DA-TKG): A novel memory substrate that explicitly models the lifecycle of external events via reinforcement-decay dynamics, ensuring the agent's "world state" reflects current reality rather than historical noise.
Structured Context Injection: By enforcing a "Structure-First" paradigm, the framework transforms raw noise into bounded, low-latency context objects, mitigating the risk of hallucination and context poisoning.
Intent-Aware Gating: The system selectively injects external context only when necessary (Dynamic intents), preserving the efficiency of static interactions.

Experimental Results

Evaluated in a "World-Lab Simulation" with 1,000 queries across static and dynamic intents, PulseCX was compared against Static RAG (industry standard) and a Naive Online Agent (synchronous web search).

Latency: PulseCX achieved a Time-to-First-Token (TTFT) of 430 ms, comparable to Static RAG (415 ms) and significantly faster than the Naive Online Agent (1500 ms).
Intent Resolution Rate (IRR): PulseCX achieved 89.2%, outperforming both Static RAG (64.1%) and the Naive Online Agent (78.5%).
Customer Satisfaction (s-CSAT): PulseCX scored 4.2/5, compared to 2.4/5 for Static RAG and 3.1/5 for the Naive Online Agent.
Escalation Rate (ER): PulseCX reduced escalations to 13.5%, a significant improvement over Static RAG (35.2%).

Business Impact Analysis:

Vocabulary Bridge: PulseCX mapped viral slang (e.g., "Glazed Donut Stick") to internal SKUs, converting search intent into sales opportunities where static agents failed.
Situational Validation: In outage scenarios, PulseCX acknowledged external reality rather than relying on lagging internal logs ("Green" status), reducing user frustration and unnecessary human handoffs by 21.7%.

Ablation Study: Disabling the decay function ( $\lambda=0$ ) resulted in the agent recommending expired deals (e.g., flash sales) hours after conclusion, confirming that temporal lifecycle management is essential for maintaining trust.

Significance and Claims

The paper claims that PulseCX demonstrates the Closed-World Constraint of modern agents is an architectural choice, not a fundamental limitation. By formalizing unstructured signals into a DA-TKG and adopting a structure-first approach to real-time memory evolution, the framework enables agentic systems to be:

Faster: Maintaining sub-450ms latency despite dynamic context.
Safer: Reducing hallucination risks and context poisoning through structured injection and gating.
Aligned: Truly reflecting the rapidly shifting reality of human experience and external events.

The authors acknowledge limitations, including potential temporal blindness between fixed update cycles, vulnerability to adversarial poisoning via coordinated inauthentic activity, and reliance on heuristic decay parameters rather than learned coefficients. However, the framework establishes a new paradigm for grounding agents in the "wild" without the overhead of synchronous retrieval.