qarai-agent-guard 0.1.0

A lightweight guardrails framework for AI agents that supports multiple languages, primarily Arabic and its dialects, for safety and PII governance.

Qarai Agent Guard

Qarai Agent Guard is a lightweight, composable guardrails framework for AI agents. It inspects agent memory and tool I/O at runtime, detects unsafe or sensitive content, and enforces policy-driven actions — block, warn, redact, or allow — before data is persisted or returned to users.

The project is built with multilingual safety in mind. Arabic is a first-class language today, with pattern libraries and detectors tuned for Modern Standard Arabic. Arabic dialects (Maghrebi, Egyptian, Levantine, Gulf, and others) are on the roadmap, reflecting the linguistic diversity of real-world deployments across the MENA region and the global Arabic-speaking diaspora.

Whether you run agents in English, French, Arabic, or a mix of languages, Qarai Agent Guard gives you a single, policy-driven layer to govern PII, secrets, prompt injection, and model-reasoning leakage — without rewriting your agent stack.

---

Why Qarai Agent Guard?

AI agents accumulate context in memory, call external tools, and exchange messages across sessions. That context is a high-value attack surface: prompt injection can poison long-term memory, PII can leak into logs, and chain-of-thought content can surface where it should not.

Qarai Agent Guard addresses this by sitting between your agent and its memory layer, running detectors on every create, read, update, and delete operation. Policies map detection severity to concrete actions so security teams can tune behavior without touching application code.

Capability	Description
Multilingual detection	YAML-driven regex patterns for English, French, and Arabic; extensible to dialects
Policy engine	Severity-based rules: block, redact, warn, allow
Memory governance	Hooks on all CRUD paths for agent memory
Framework integrations	LangChain buffer memory support out of the box
Composable detectors	PII, secrets, model-reasoning / injection patterns
Custom policies	Built-in presets or load your own YAML policy files

---

Quick start

Install the package (see install.md for full instructions):

pip install qarai-agent-guard
# LangChain integration (optional)
pip install langchain-community

Configure guards and attach them to LangChain buffer memory:

from qarai_agent_guard import (
    BufferMemoryMiddleware,
    GuardedBufferChatMessageHistory,
    ModelReasoningDetector,
    PIIDetector,
    strict_policy,
)

middleware = BufferMemoryMiddleware(
    detectors=[
        ModelReasoningDetector(lang="ar"),
        PIIDetector(),
    ],
    policy=strict_policy(),
)

history = GuardedBufferChatMessageHistory(
    middleware=middleware,
    session_id="user-42",
)

history.add_user_message("مرحبا، كيف يمكنني المساعدة؟")

You pass detectors and policy to BufferMemoryMiddleware. The library wires AgentGuard, MemoryGuard, and the internal middleware for you.

---

Documentation

Document	Description
install.md	Installation, dependencies, and environment setup
contributing.md	How to contribute code, patterns, and documentation
docs/index.md	Component reference and architecture guide

---

Project status

Qarai Agent Guard is under active development. The core detector, policy, and guard pipeline is stable; integrations and dialect coverage are expanding. We welcome contributions — especially Arabic dialect patterns, new integrations, and policy presets.

---

License

Apache License 2.0. See LICENSE for details.