> ## Documentation Index
> Fetch the complete documentation index at: https://docs.lyzr.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Lyzr ADK (Agent Development Kit)

The **Lyzr ADK (Agent Development Kit)** is a high-performance ADK designed to embed **Safe-and-Responsible-AI guardrails** into every autonomous agent. It serves as the core compute layer for building, deploying, and managing enterprise-grade agents, comparable to frameworks like Langgraph, CrewAI, or Autogen.

The ADK provides a governed workflow that empowers both engineers and business users to ship production-grade agents in weeks while satisfying the strictest security, compliance, and model-risk standards.

***

## Core Value Proposition

Lyzr.ai delivers an enterprise-ready agent platform that fuses autonomous intelligence with governed workflow automation. The ADK is the engine behind this, offering:

* **Native Safe & Responsible Module**: An AI framework that provides built-in security and ethical decision-making.
* **Knowledge-Graph Core**: A semantic layer that empowers agents to move beyond simple keyword matching to relationship-based reasoning.
* **Low-Code Integration**: Seamlessly bridges the gap between a visual design studio and full code-level control for developers.

***

## 3-Layer Architecture

The ADK is part of a comprehensive three-layer ecosystem:

1. **Organizational General Intelligence (OGI)**: A shared vector-based knowledge graph that captures prompts, outputs, and lineage to enable enterprise-wide intelligence reuse.
2. **Lyzr Agent Studio**: A visual environment for building single agents, multi-agent orchestrations (managerial, DAG, sequential), and knowledge bases.
3. **Lyzr Agent Framework (ADK)**: The SDK layer that embeds guardrails and exposes connectors for MCP, OpenAI functions, Google A2A, and classic REST/gRPC services.

***

## Key Features & Building Blocks

### 1. Memory Management

The ADK offers three specialized memory varieties to ensure agents deliver coherent and personalized responses:

* **Lyzr Cognis**: A state-of-the-art module with temporal awareness and Matryoshka embeddings for production-grade context retention.
* **Lyzr Memory**: Intelligently combines short-term (word-for-word) and long-term (summarized) memory to prevent latency issues and hallucinations.
* **Amazon Bedrock AgentCore**: A managed service for maintaining both immediate and long-term knowledge using AWS credentials.

### 2. RAG & Knowledge Systems

The framework allows for the creation of no-code RAG pipelines for fast, searchable document understanding:

* **Unstructured Data Support**: Processes PDF, DOCX, TXT, and Website URLs.
* **Advanced Retrieval**: Supports Basic, MMR (Maximal Marginal Relevance), and HyDE (Hypothetical Document Embeddings) retrieval types.
* **Graph RAG**: Uses Neo4J to answer multi-hop and context-heavy queries that traditional RAG pipelines cannot handle.

### 3. Modular Agent Skills

Agent skills are modular building blocks that define what an agent can do. They contain executable logic, tool integrations, or multi-step prompts. This allows developers to write logic once and reuse it across multiple agents, creating a composable system rather than one-off flows.

***

## Safe & Responsible AI

Lyzr is one of the few platforms with Responsible AI built into its foundation. The ADK enforces guardrails that include:

* **Toxicity & Bias Control**: Automatically detects and prevents harmful content or skewed outputs.
* **PII Redaction**: Controls how agents handle sensitive personal data by blocking or redacting information like SSNs, credit card numbers, and emails.
* **Prompt Injection Protection**: Blocks malicious manipulation attempts before they reach the Large Language Model.
* **Reflection Mechanism**: Allows agents to critically assess their own responses against instructions to reduce hallucinations.

***

## Deployment & Interoperability

Every agent created with the ADK is instantly deployable.

* **Structured Output**: Supports machine-readable formats like JSON, enabling agents to act as modular, programmable components in larger automation systems.

Would you like me to generate a specific **Python code snippet** to initialize an agent with these guardrails using the ADK?