# Conversational AI Agent

**Component Classification:** LLM-powered natural language processing and orchestration interface within the aarnâ Multi-Agent Architecture.

#### 1. Functional Purpose

The Conversational Agent acts as the semantic interface layer between user intent and system-level configuration logic. It converts natural language inputs into structured, machine-interpretable command objects and governs the decision pipeline before execution. The Conversational Agent ensures that all user-driven actions (configuration, insight queries, transaction intent, or reasoning requests) are processed deterministically and aligned with protocol rules, safeguards, and vault risk constraints.

#### 2. Core Capabilities

**Intent Classification & Semantic Parsing**

* Utilizes transformer-based LLM inference for:
  * Intent detection
  * Slot/entity extraction
  * Constraint identification (risk thresholds, asset boundaries, timeline constraints).
* Normalizes ambiguous expressions (e.g., "maximum safe yield" or "higher returns but low volatility") into system-defined feature vectors.

**Policy-Aligned Strategy Mapping**

* Converts parsed intent into validated vault configuration instructions.
* Applies:
  * Risk framework lookups
  * Eligibility matrices
  * TVL ceilings
  * Chain availability constraints
  * Tactical vs strategic sleeve logic.
* Generates configuration instructions for downstream agents without violating governance or execution rules.

**Execution Mediation**

* The Conversational Agent does not directly trigger blockchain transactions.
* Instead, it:
  * Generates execution signatures,
  * Confirms user approval,
  * Passes final instruction packets to the Execution Agent.

#### 3. Memory + Personalization Layer

* Persistent, wallet-scoped memory (no identity leakage).
* Uses vector embeddings + semantic recall rather than rule-based storage.
* Supports session memory (short-term reasoning) and long-tail preference modeling.

Memory examples stored as feature vectors:

* Risk preferences
* Typical asset exposure
* Interaction patterns (active vs passive user)
* Prior allocations and rebalance approvals

#### 4. Communication Model

The Conversational Agent supports:

* User-initiated flow (queries, instructions)
* System-initiated flow (alerts, deviation signals, yield opportunity push)

Communication signals include:

* Risk shift notifications
* Rebalancing triggers
* Yield deltas > threshold
* PT opportunity upgrades (from Yield Curation Agent)

All messaging is generated in natural language but anchored in a quantifiable system state.

#### 5. Guarantees & Guardrails

| Domain                  | Enforcement                                          |
| ----------------------- | ---------------------------------------------------- |
| Safety                  | Risk policy enforcement, transaction sanity checks   |
| Consistency             | Deterministic mapping of preferences → configuration |
| Explainability          | All actions include natural language reasoning       |
| Non-Custodial Integrity | No execution without explicit approval               |

<figure><img src="https://677490501-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FO3jlJDGYTE4CuS69YpTx%2Fuploads%2FNdPRrjyXNgjc3Vsvrm9P%2Fhow%20aTARS%20works.png?alt=media&#x26;token=4b8ac909-dde5-4018-8f6a-8cbd989675ee" alt=""><figcaption></figcaption></figure>