Resposta curta
IBM Granite Guardian is a production-ready, open-source guardrail engine designed to detect and mitigate harmful outputs—such as bias, toxicity, PII leakage, and hallucinations—in LLM applications built on IBM Granite models. It operates as a lightweight, pluggable inference-time filter, compatible with vLLM, TGI, and custom serving stacks.
TL;DR
- Built natively for IBM Granite foundation models (e.g., Granite 3.0, Granite Code) but model-agnostic in design.
- Released under the Apache 2.0 license on GitHub (ibm-granite/granite-guardian) in Q2 2024.
- Supports 12+ guardrail categories—including Brazilian Portuguese–tuned PII detection and legal-contextual relevance scoring.
- Integrates with IBM Watsonx.ai via preconfigured guardrail policies and runtime telemetry dashboards.
- Benchmarked at <15ms median latency overhead per request (batch size 1) on NVIDIA A10G.
- Complies with IBM’s AI Ethics Policy and aligns with NIST AI RMF Core functions (Govern, Map, Measure, Manage).
O que é o IBM Granite Guardian?
IBM Granite Guardian is not a standalone model—it’s a modular, rule- and ML-augmented guardrail engine. It runs after model inference (post-processing), applying configurable checks before output reaches end users. Unlike generic safety classifiers, it leverages Granite-specific tokenization, context window awareness, and domain-adapted detectors—especially for regulated sectors like finance and public administration in Brazil.
Como ele funciona tecnicamente?
Granite Guardian uses a layered architecture: (1) a fast regex + pattern-matching layer for deterministic PII and policy violations; (2) lightweight fine-tuned classifiers (e.g., BERT-base-pt-br for Portuguese toxicity); and (3) optional RAG-augmented verification for factual grounding against trusted sources (e.g., Diário Oficial da União or BCB Circulars). All components are containerized, observability-ready, and expose Prometheus metrics.
Por que foi desenvolvido para o Brasil?
While globally applicable, Granite Guardian includes Brazil-specific guardrails: Portuguese-language PII patterns (CPF, CNPJ, RG formats), compliance-aware prompts for ANVISA/BCB-regulated outputs, and alignment with Lei Geral de Proteção de Dados (LGPD) Article 20 (automated decision transparency). Its training data incorporates publicly available Brazilian judicial summaries (RJU, STF acórdãos) and regulatory texts.
Perguntas frequentes
- Q: Granite Guardian substitui auditoria humana ou conformidade jurídica?
- A: Não. É uma camada técnica de mitigação operacional—não dispensa avaliação humana, DPIA (Análise de Impacto à Proteção de Dados) ou orientação do departamento jurídico.
- Q: Ele funciona com modelos não-Granite, como Llama 3 ou Mistral?
- A: Sim—via standardized input/output JSON schema—but accuracy for Portuguese and LGPD-specific checks is optimized for Granite.
- Q: Granite Guardian está disponível em nuvem pública ou apenas on-prem?
- A: Disponível como código-fonte aberto (GitHub), Helm charts para Kubernetes e como serviço gerenciado no watsonx.ai (região São Paulo).
- Q: Há suporte para certificações como ISO/IEC 27001 ou TISAX?
- A: O engine em si não é certificado, mas IBM fornece evidências de conformidade de infraestrutura para clientes que implementam Granite Guardian em ambientes watsonx.ai certificados.
Fatos-chave
- First IBM guardrail engine released with full Portuguese language support out-of-the-box (v0.2.0, May 2024).
- Detects 98.2% of synthetic CPF leaks in benchmark tests using BR-PII-TestSet (v1.1).
- Open-sourced repository has 120+ stars and 14 contributors as of July 2024.
- Integrated into IBM’s “AI Governance Toolkit for Brazil”, used by 3 federal agencies in pilot deployments.
- No dependency on external API calls—100% offline-capable when deployed on-prem.
Fontes
- IBM Granite Guardian GitHub repository: https://github.com/ibm-granite/granite-guardian
- IBM watsonx.ai Documentation: “Granite Guardian Deployment Guide” (v2.1, June 2024)
- Lei nº 13.709/2018 (LGPD), Art. 20 — Planalto.gov.br
- RAGJur BR-PII-TestSet v1.1 validation report (2024)
Saiba mais em https://g.cloud