Gemini SRE Agent: Autonomous Monitoring and Remediation

Welcome to the Gemini SRE Agent, an autonomous system designed to enhance your Google Cloud operations by intelligently monitoring logs and automating incident response. This project leverages the power of Google's Gemini models to bring advanced AI capabilities directly into your Site Reliability Engineering (SRE) workflows.

At its core, the Gemini SRE Agent acts as a proactive digital assistant, continuously observing your Google Cloud environment. When anomalies or critical events are detected in your logs, it doesn't just alert you; it initiates a structured process of analysis, generates intelligent code fixes using its unified enhanced code generation system, and proposes concrete remediation steps, culminating in automated GitHub Pull Requests. This approach aims to reduce manual toil, accelerate incident resolution, and improve the overall reliability of your cloud services.

Whether you're looking to streamline your incident management, gain deeper insights from your operational data, or simply explore the practical applications of generative AI in SRE, the Gemini SRE Agent offers a robust and extensible foundation. It's built with an emphasis on configurability, resilience, and clear observability, ensuring it can adapt to diverse cloud environments and operational needs.

Dive in to discover how this agent can transform your cloud log monitoring into an intelligent, automated, and resilient operation.

System Architecture

The Gemini SRE Agent employs a sophisticated multi-model AI architecture with enhanced dynamic prompt generation, advanced pattern detection, a unified enhanced code generation system, and a comprehensive new log ingestion system for intelligent log monitoring and automated remediation:

graph TB
    subgraph "Google Cloud Platform"
        CL[Cloud Logging] --> PS[Pub/Sub Topics]
        PS --> |Log Messages| SUB[Pub/Sub Subscriptions]
    end

    subgraph "Gemini SRE Agent - New Log Ingestion System"
        SUB --> LM[Log Manager<br/>Orchestration Layer]
        LM --> |Multi-Source| GCP[GCP Pub/Sub Adapter]
        LM --> |Multi-Source| K8S[Kubernetes Adapter]
        LM --> |Multi-Source| FS[File System Adapter]
        LM --> |Multi-Source| AWS[AWS CloudWatch Adapter]

        GCP --> |LogEntry| LP[Log Processor]
        K8S --> |LogEntry| LP
        FS --> |LogEntry| LP
        AWS --> |LogEntry| LP

        LP --> |Structured Logs| MLPR[ML Pattern Refinement<br/>AI Enhancement Layer]
        MLPR --> |Enhanced Analysis| PD[Pattern Detection System<br/>Multi-Layer Analysis]
        PD --> |Pattern Match| TA[Triage Agent<br/>Gemini Flash]
        LP --> |Structured Logs| TA
        TA --> |TriagePacket| AA[Enhanced Analysis Agent<br/>Dynamic Prompt Generation]
        AA --> |ValidationRequest| QA[Quantitative Analyzer<br/>Code Execution]
        QA --> |EmpiricalData| AA
        AA --> |RemediationPlan| UECG[Unified Enhanced<br/>Code Generation System]
        UECG --> |Generated Code| RA[Remediation Agent]
    end

    subgraph "Legacy System (Backward Compatible)"
        SUB --> LS[Legacy Log Subscriber]
        LS --> |Raw Logs| TA
    end

    subgraph "Comprehensive Monitoring System"
        MM[Monitoring Manager] --> MC[Metrics Collector]
        MM --> HC[Health Checker]
        MM --> PM[Performance Monitor]
        MM --> AM[Alert Manager]

        LM --> |Health Status| MM
        LP --> |Processing Metrics| MM
        TA --> |Analysis Metrics| MM
        AA --> |Generation Metrics| MM
    end

    subgraph "ML Pattern Refinement Layer"
        MLPR --> LQV[Log Quality<br/>Validator]
        MLPR --> LSan[Log Sanitizer<br/>PII Removal]
        MLPR --> GEPD[Gemini Enhanced<br/>Pattern Detector]
        MLPR --> GRC[Response Cache<br/>Cost Optimization]
    end

    subgraph "Pattern Detection Layers"
        PD --> TW[Time Window<br/>Management]
        PD --> TE[Smart Threshold<br/>Evaluation]
        PD --> PC[Pattern<br/>Classification]
        PD --> CS[Confidence<br/>Scoring]
    end

    subgraph "Unified Enhanced Code Generation System"
        UECG --> UWO[Unified Workflow<br/>Orchestrator]
        UWO --> WCM[Workflow Context<br/>Manager]
        UWO --> WAE[Workflow Analysis<br/>Engine]
        UWO --> WCG[Workflow Code<br/>Generator]
        UWO --> WVE[Workflow Validation<br/>Engine]
        UWO --> WMC[Workflow Metrics<br/>Collector]

        WCG --> CGF[Code Generator<br/>Factory]
        CGF --> APIG[API Code<br/>Generator]
        CGF --> DBG[Database Code<br/>Generator]
        CGF --> SECG[Security Code<br/>Generator]

        WVE --> CVP[Code Validation<br/>Pipeline]
        CVP --> SYN[Syntax<br/>Validation]
        CVP --> PAT[Pattern<br/>Validation]
        CVP --> SEC[Security<br/>Review]
        CVP --> PERF[Performance<br/>Assessment]
        CVP --> BP[Best Practices<br/>Check]
    end

    subgraph "External Services"
        RA --> |Create PR| GH[GitHub Repository]
        RA --> |Notifications| SLACK[Slack/PagerDuty]
        MLPR --> |Code Context| GH
        UECG --> |Repository Analysis| GH
    end

    subgraph "Configuration & Resilience"
        CONFIG[config.yaml<br/>Multi-Service Setup] --> LS
        RESILIENCE[Hyx Resilience<br/>Circuit Breakers] --> TA
        RESILIENCE --> AA
        RESILIENCE --> RA
        RESILIENCE --> MLPR
        RESILIENCE --> UECG
    end

    classDef aiComponent fill:#e1f5fe,stroke:#01579b,stroke-width:2px
    classDef gcpService fill:#fff3e0,stroke:#f57c00,stroke-width:2px
    classDef external fill:#f3e5f5,stroke:#7b1fa2,stroke-width:2px
    classDef config fill:#e8f5e8,stroke:#388e3c,stroke-width:2px
    classDef mlComponent fill:#fce4ec,stroke:#c2185b,stroke-width:2px
    classDef codeGenComponent fill:#e8f5e8,stroke:#2e7d32,stroke-width:3px
    classDef ingestionComponent fill:#e3f2fd,stroke:#1976d2,stroke-width:2px
    classDef monitoringComponent fill:#f1f8e9,stroke:#689f38,stroke-width:2px
    classDef legacyComponent fill:#fafafa,stroke:#616161,stroke-width:2px

    class TA,AA,QA aiComponent
    class CL,PS,SUB gcpService
    class GH,SLACK external
    class CONFIG,RESILIENCE config
    class MLPR,LQV,LSan,GEPD,GRC mlComponent
    class UECG,UWO,WCM,WAE,WCG,WVE,WMC,CGF,APIG,DBG,SECG,CVP,SYN,PAT,SEC,PERF,BP codeGenComponent
    class LM,GCP,K8S,FS,AWS,LP ingestionComponent
    class MM,MC,HC,PM,AM monitoringComponent
    class LS legacyComponent

Loading

Multi-Model AI Strategy

The system leverages different Gemini models optimized for specific tasks:

• Gemini Flash: High-speed log triage and classification (cost-optimized)
• Gemini Pro: Deep analysis and code generation (accuracy-optimized)
• Code Execution: Empirical validation and quantitative analysis

Enhanced Dynamic Prompt Generation System

The Gemini SRE Agent now features a revolutionary dynamic prompt generation system that automatically creates context-aware, specialized prompts for optimal code generation:

graph TB
    subgraph "Enhanced Prompt Generation Pipeline"
        TC[Task Context] --> APS[Adaptive Prompt Strategy]
        IC[Issue Context] --> APS
        RC[Repository Context] --> APS

        APS --> |Complex Issues| MPG[Meta-Prompt Generator<br/>Gemini Flash]
        APS --> |Specialized Issues| SPT[Specialized Templates<br/>Database/API/Security]
        APS --> |Simple Issues| GPT[Generic Prompt Template]

        MPG --> |Optimized Prompt| MPG2[Meta-Prompt Generator<br/>Gemini Pro]
        MPG2 --> |Final Prompt| EAA[Enhanced Analysis Agent]
        SPT --> |Specialized Prompt| EAA
        GPT --> |Generic Prompt| EAA

        EAA --> |Analysis Result| VAL[Validation & Refinement]
        VAL --> |Validated Code| OUT[Code Fix Output]
    end

    classDef metaPrompt fill:#e1f5fe,stroke:#01579b,stroke-width:2px
    classDef specialized fill:#f3e5f5,stroke:#7b1fa2,stroke-width:2px
    classDef generic fill:#e8f5e8,stroke:#388e3c,stroke-width:2px

    class MPG,MPG2 metaPrompt
    class SPT specialized
    class GPT generic

Loading

Key Capabilities:

• Meta-Prompt Generation: Uses Gemini Flash to generate optimized prompts for Gemini Pro, creating a "AI teaching AI" approach
• Context-Aware Templates: Automatically selects specialized prompt templates based on issue type (database errors, API failures, security issues, etc.)
• Adaptive Strategy Selection: Intelligently chooses between meta-prompt, specialized, or generic approaches based on issue complexity
• Multi-Stage Validation: Implements iterative refinement with validation feedback loops
• Fallback Mechanisms: Gracefully degrades to simpler approaches if advanced features fail
• Performance Optimization: Caching and similarity matching for cost-effective prompt generation

Issue Type Specialization:

• Database Errors: Specialized prompts for connection issues, query optimization, deadlocks
• API Errors: Authentication, rate limiting, endpoint failures, response validation
• Security Issues: Vulnerability assessment, access control, encryption problems
• Service Errors: Microservice communication, dependency failures, resource exhaustion
• Infrastructure Issues: Deployment problems, configuration errors, scaling issues

Unified Enhanced Code Generation System

The Gemini SRE Agent now features a comprehensive unified enhanced code generation system that provides end-to-end automated remediation capabilities. This system represents a significant advancement in AI-powered incident response, combining specialized code generators, multi-level validation, performance optimization, and adaptive learning.

Unified Code Generation Architecture

graph TB
    subgraph "Unified Enhanced Code Generation Pipeline"
        UWO[Unified Workflow<br/>Orchestrator] --> WCM[Workflow Context<br/>Manager]
        UWO --> WAE[Workflow Analysis<br/>Engine]
        UWO --> WCG[Workflow Code<br/>Generator]
        UWO --> WVE[Workflow Validation<br/>Engine]
        UWO --> WMC[Workflow Metrics<br/>Collector]

        WCM --> |Context Building| IC[Issue Context<br/>Cache]
        WCM --> |Repository Analysis| RC[Repository Context<br/>Cache]

        WAE --> |Enhanced Analysis| EAA[Enhanced Analysis<br/>Agent]
        EAA --> |Meta-Prompt Generation| MPG[Meta-Prompt<br/>Generator]
        EAA --> |Adaptive Strategy| APS[Adaptive Prompt<br/>Strategy]

        WCG --> |Specialized Generation| CGF[Code Generator<br/>Factory]
        CGF --> |API Issues| APIG[API Code<br/>Generator]
        CGF --> |Database Issues| DBG[Database Code<br/>Generator]
        CGF --> |Security Issues| SECG[Security Code<br/>Generator]

        WVE --> |Multi-Level Validation| CVP[Code Validation<br/>Pipeline]
        CVP --> |Syntax Check| SYN[Syntax<br/>Validation]
        CVP --> |Pattern Check| PAT[Pattern<br/>Validation]
        CVP --> |Security Review| SEC[Security<br/>Review]
        CVP --> |Performance Check| PERF[Performance<br/>Assessment]
        CVP --> |Best Practices| BP[Best Practices<br/>Check]

        WMC --> |Performance Monitoring| PM[Performance<br/>Monitor]
        WMC --> |Cost Tracking| CT[Cost<br/>Tracker]
        WMC --> |Learning System| ECL[Enhanced Code<br/>Generation Learning]
    end

    classDef orchestrator fill:#e1f5fe,stroke:#01579b,stroke-width:3px
    classDef workflow fill:#f3e5f5,stroke:#7b1fa2,stroke-width:2px
    classDef generator fill:#e8f5e8,stroke:#388e3c,stroke-width:2px
    classDef validation fill:#fff3e0,stroke:#f57c00,stroke-width:2px
    classDef monitoring fill:#fce4ec,stroke:#c2185b,stroke-width:2px

    class UWO orchestrator
    class WCM,WAE,WCG,WVE,WMC workflow
    class CGF,APIG,DBG,SECG generator
    class CVP,SYN,PAT,SEC,PERF,BP validation
    class PM,CT,ECL monitoring

Loading

Core Components

Unified Workflow Orchestrator:

• Central Coordination: Manages the entire code generation workflow from analysis to validation
• Modular Architecture: Broken down into specialized workflow components for maintainability
• Async Optimization: Concurrent task execution with priority queuing and retry mechanisms
• Performance Monitoring: Real-time metrics collection and trend analysis

Enhanced Analysis Engine:

• Context-Aware Analysis: Incorporates repository structure, recent commits, and issue patterns
• Meta-Prompt Generation: Uses AI to generate optimized prompts for better code generation
• Adaptive Strategy Selection: Chooses optimal analysis approach based on issue complexity
• Fallback Mechanisms: Graceful degradation to simpler approaches when needed

Specialized Code Generators:

• API Code Generator: Handles authentication, rate limiting, endpoint failures, response validation
• Database Code Generator: Manages connection issues, query optimization, deadlocks, migrations
• Security Code Generator: Addresses vulnerability assessment, access control, encryption problems
• Factory Pattern: Dynamic generator selection based on issue type and context

Multi-Level Validation Pipeline:

• Syntax Validation: Ensures generated code is syntactically correct
• Pattern Validation: Verifies code follows established patterns and conventions
• Security Review: Identifies potential security vulnerabilities and compliance issues
• Performance Assessment: Evaluates performance impact and optimization opportunities
• Best Practices Check: Ensures adherence to coding standards and architectural principles

Performance & Cost Management:

• Adaptive Rate Limiting: Smart API request management with circuit breaker patterns
• Cost Tracking: Real-time budget monitoring with configurable limits and alerts
• Performance Monitoring: Comprehensive metrics collection and trend analysis
• Async Optimization: Concurrent processing with intelligent task prioritization

Learning & Improvement System:

• Generation History Tracking: Records successful and failed code generation attempts
• Pattern Learning: Identifies successful patterns for future use
• Feedback Integration: Incorporates validation results to improve future generations
• Performance Insights: Provides actionable recommendations for system optimization

Key Capabilities

Intelligent Code Generation:

• Context-Aware: Incorporates repository structure, recent changes, and issue patterns
• Specialized Approaches: Domain-specific generators for different types of issues
• Multi-Model Strategy: Leverages different Gemini models for optimal results
• Iterative Refinement: Continuous improvement based on validation feedback

Quality Assurance:

• Multi-Level Validation: Comprehensive checking from syntax to security
• Automated Testing: Integration with existing test suites and validation frameworks
• Code Review Integration: Generates pull requests with detailed explanations
• Compliance Checking: Ensures adherence to security and coding standards

Performance Optimization:

• Concurrent Processing: Async task execution for improved throughput
• Intelligent Caching: Context and response caching to reduce API costs
• Resource Management: Adaptive rate limiting and circuit breaker patterns
• Cost Control: Budget monitoring with automatic throttling and alerts

Learning & Adaptation:

• Success Pattern Recognition: Identifies and reuses successful generation patterns
• Failure Analysis: Learns from failed attempts to improve future generations
• Performance Tuning: Continuous optimization based on metrics and feedback
• Adaptive Strategies: Dynamic adjustment of generation approaches based on success rates

This unified system represents a significant advancement in AI-powered incident response, providing a complete end-to-end solution for automated code generation, validation, and deployment in SRE workflows.

New Log Ingestion System

The Gemini SRE Agent now features a comprehensive new log ingestion system that provides enterprise-grade log processing capabilities with full backward compatibility. This system represents a major architectural advancement, offering pluggable adapters, comprehensive monitoring, and production-ready resilience patterns.

Enhanced Log Ingestion Architecture

graph TB
    subgraph "Log Ingestion System"
        LM[Log Manager<br/>Orchestration] --> |Multi-Source| GCP[GCP Pub/Sub<br/>Adapter]
        LM --> |Multi-Source| K8S[Kubernetes<br/>Adapter]
        LM --> |Multi-Source| FS[File System<br/>Adapter]
        LM --> |Multi-Source| AWS[AWS CloudWatch<br/>Adapter]

        GCP --> |LogEntry| LP[Log Processor<br/>Standardization]
        K8S --> |LogEntry| LP
        FS --> |LogEntry| LP
        AWS --> |LogEntry| LP

        LP --> |Structured Logs| AI[AI Analysis<br/>Pipeline]
    end

    subgraph "Monitoring & Observability"
        MM[Monitoring Manager] --> MC[Metrics Collector<br/>Real-time Metrics]
        MM --> HC[Health Checker<br/>Component Health]
        MM --> PM[Performance Monitor<br/>Processing Analytics]
        MM --> AM[Alert Manager<br/>Intelligent Alerts]

        LM --> |Health Status| MM
        LP --> |Processing Metrics| MM
        GCP --> |Adapter Metrics| MM
        K8S --> |Adapter Metrics| MM
    end

    subgraph "Resilience & Reliability"
        CB[Circuit Breakers] --> GCP
        CB --> K8S
        CB --> FS
        CB --> AWS

        BP[Backpressure Manager] --> LP
        RT[Retry Logic] --> GCP
        RT --> K8S
        RT --> FS
        RT --> AWS
    end

    classDef ingestion fill:#e3f2fd,stroke:#1976d2,stroke-width:2px
    classDef monitoring fill:#f1f8e9,stroke:#689f38,stroke-width:2px
    classDef resilience fill:#fff3e0,stroke:#f57c00,stroke-width:2px

    class LM,GCP,K8S,FS,AWS,LP ingestion
    class MM,MC,HC,PM,AM monitoring
    class CB,BP,RT resilience

Loading

Log Ingestion Capabilities

Multi-Source Log Ingestion:

• GCP Pub/Sub: Production-ready Google Cloud Pub/Sub integration with flow control and acknowledgment management
• Kubernetes: Container log collection with namespace filtering and pod-level granularity
• File System: Local and remote file monitoring with rotation support and pattern matching
• AWS CloudWatch: Amazon CloudWatch Logs integration with log group and stream management
• Extensible Architecture: Easy addition of new log sources through the adapter pattern

Enterprise-Grade Monitoring:

• Real-time Metrics: Processing rates, error counts, latency tracking, and throughput analysis
• Health Monitoring: Component-level health checks with automated status reporting
• Performance Analytics: Resource utilization tracking and bottleneck identification
• Intelligent Alerting: Configurable alerts with escalation policies and notification channels

Production-Ready Resilience:

• Circuit Breakers: Prevent cascade failures and protect downstream systems
• Backpressure Management: Handle high-volume log streams without overwhelming the system
• Automatic Retries: Resilient error handling with exponential backoff and jitter
• Graceful Degradation: Fallback mechanisms and graceful shutdown procedures

Unified Configuration:

• Single Configuration System: Centralized configuration for all log sources and adapters
• Runtime Updates: Dynamic configuration changes without system restarts
• Validation: Comprehensive configuration validation with detailed error reporting
• Environment Integration: Seamless integration with environment variables and secrets management

Migration Strategy

The new log ingestion system is designed for zero-downtime migration with full backward compatibility:

Feature Flag Control:

# Enable new ingestion system
export USE_NEW_INGESTION_SYSTEM=true

# Enable comprehensive monitoring
export ENABLE_MONITORING=true

# Keep legacy fallback enabled
export ENABLE_LEGACY_FALLBACK=true

Migration Phases:

1. Phase 1: Deploy with new system disabled, enable monitoring
2. Phase 2: Enable new system for non-critical services
3. Phase 3: Gradual expansion based on confidence and metrics
4. Phase 4: Full migration with legacy system deprecation

Safety Features:

• Automatic Fallback: Seamless fallback to legacy system if new system fails
• Health Monitoring: Real-time health checks and status reporting
• Performance Tracking: Comprehensive metrics for migration validation
• Rollback Capability: Instant rollback via environment variable changes

Benefits

Operational Excellence:

• Enhanced Observability: Real-time monitoring and alerting capabilities
• Improved Reliability: Circuit breakers, retries, and graceful error handling
• Better Performance: Optimized processing with backpressure management
• Simplified Operations: Unified configuration and monitoring interface

Future-Proof Architecture:

• Multi-Cloud Ready: Support for GCP, AWS, and hybrid environments
• Container Native: Kubernetes integration with pod and namespace awareness
• Extensible Design: Easy addition of new log sources and processing capabilities
• Scalable Processing: Horizontal scaling with load balancing and distribution

Cost Optimization:

• Efficient Processing: Optimized resource utilization and processing patterns
• Smart Monitoring: Cost-effective monitoring with configurable sampling
• Intelligent Caching: Response caching and similarity matching for API cost reduction
• Resource Management: Adaptive rate limiting and circuit breaker patterns

Key Features

• 🚀 New Log Ingestion System: Enterprise-grade log processing with pluggable adapters, comprehensive monitoring, and production-ready resilience patterns. Supports GCP Pub/Sub, Kubernetes, File System, and AWS CloudWatch with zero-downtime migration.
• 📊 Comprehensive Monitoring: Real-time metrics collection, health checks, performance monitoring, and intelligent alerting with configurable escalation policies and notification channels.
• 🛡️ Production-Ready Resilience: Circuit breakers, backpressure management, automatic retries, and graceful degradation with seamless fallback to legacy systems.
• 🔧 Feature Flag Migration: Safe, gradual rollout with environment variable controls and automatic fallback capabilities for risk-free deployment.
• Unified Enhanced Code Generation System: Complete AI-powered code generation pipeline with specialized generators, validation, and learning capabilities for automated remediation.
• AI-Enhanced Pattern Detection: Multi-layer analysis engine enhanced with Gemini AI models for intelligent pattern recognition, confidence scoring, and context-aware analysis.
• Gemini ML Pattern Refinement: Advanced AI pipeline with quality validation, PII sanitization, smart caching, and cost optimization for production-ready AI analysis.
• Enhanced Dynamic Prompt Generation: Revolutionary AI-powered prompt generation system with meta-prompt optimization, context-aware templates, and adaptive strategy selection for superior code generation.
• Configuration Management: Modern, type-safe configuration system with Pydantic validation, environment variable integration, hot reloading, and comprehensive CLI tools for validation and management.
• Intelligent Log Analysis: Leverages dual Gemini models (Flash for speed, Pro for accuracy) with enhanced dynamic prompting and few-shot learning capabilities.
• Proactive Incident Detection: Identifies 7+ distinct failure patterns with AI-powered classification and confidence scoring across 15+ quantitative factors.
• Code-Context Integration: Automated analysis incorporating recent commits, dependencies, and repository structure for more accurate remediation.
• Cost-Optimized AI Operations: Intelligent response caching with similarity matching, reducing API costs by up to 60% while maintaining accuracy.
• Quality Assurance Pipeline: Pre-processing validation, PII sanitization, and performance monitoring ensuring reliable AI analysis.
• Automated Remediation: Generates and submits GitHub Pull Requests with AI-enhanced context analysis and code repository integration.
• Multi-Service & Multi-Repository Monitoring: Designed to monitor logs from various services and manage remediation across different GitHub repositories.
• Dynamic Baseline Tracking: Maintains adaptive baselines for anomaly detection with AI-powered threshold evaluation.
• Built-in Resilience: Incorporates robust resilience patterns (circuit breakers, retries, bulkheads, rate limiting) with AI-aware error handling.
• Structured Observability: Employs structured logging with AI interaction tracking for comprehensive operational visibility.

4-Layer Pattern Detection System

The agent implements a sophisticated pattern detection engine that processes log streams through four distinct analytical layers. This system enables proactive incident detection and reduces response times by identifying emerging issues before they escalate.

Layer 1: Time Window Management

The system accumulates incoming log entries into configurable sliding time windows. Each window maintains temporal boundaries and automatically processes accumulated logs when the window expires. This approach enables the analysis of log patterns across time intervals rather than processing individual log entries in isolation.

Technical Implementation:

• Configurable window duration (default: 5 minutes)
• Automatic log aggregation by timestamp
• Service-based log grouping within windows
• Memory-efficient window rotation and cleanup

Layer 2: Smart Threshold Evaluation

Multiple threshold types evaluate each time window against dynamic baselines and absolute limits. The system maintains historical baselines and compares current metrics against these learned patterns to detect deviations.

Threshold Types:

• Error Frequency: Absolute error count thresholds with service grouping
• Error Rate: Percentage increase from rolling baseline averages
• Service Impact: Multi-service failure detection across correlated services
• Severity Weighted: Weighted scoring system based on log severity levels
• Cascade Failure: Cross-service correlation analysis within time windows

Baseline Tracking:

• Rolling window baselines (configurable history depth)
• Service-specific baseline calculation
• Automatic baseline adaptation over time

Layer 3: Pattern Classification

When thresholds trigger, the system applies classification algorithms to identify specific failure patterns. Each pattern type has distinct characteristics and triggers different remediation approaches.

Detected Pattern Types:

• Cascade Failure: Sequential failures across dependent services
• Service Degradation: Performance degradation within a single service
• Traffic Spike: Volume-induced system stress and failures
• Configuration Issue: Deployment or configuration-related problems
• Dependency Failure: External service or dependency problems
• Resource Exhaustion: Memory, CPU, or storage capacity issues
• Sporadic Errors: Random distributed failures without clear correlation

Layer 4: Confidence Scoring

A quantitative confidence assessment system evaluates pattern matches using 15+ measurable factors. This scoring system provides numerical confidence levels and detailed explanations for each classification decision.

Confidence Factors:

• Temporal Analysis: Time concentration, correlation, onset patterns (rapid vs gradual)
• Service Impact: Service count, distribution uniformity, cross-service correlation
• Error Characteristics: Frequency, severity distribution, type consistency, message similarity
• Historical Context: Baseline deviation, trend analysis, seasonal patterns
• External Factors: Dependency health, resource utilization, deployment timing

Scoring Process:

• Weighted factor combination with configurable rules per pattern type
• Decay functions (linear, exponential, logarithmic) for factor processing
• Threshold-based factor filtering
• Normalized confidence scores (0.0 to 1.0) with categorical levels (VERY_LOW to VERY_HIGH)

Output:

• Overall confidence score with factor breakdown
• Human-readable explanations for classification decisions
• Raw factor values for debugging and tuning

This multi-layer approach reduces false positives while maintaining high sensitivity to genuine incidents. The confidence scoring system provides transparency into classification decisions, enabling operators to understand and tune detection behavior based on their specific environment characteristics.

Gemini ML Pattern Refinement System

The Gemini ML Pattern Refinement System enhances the 4-layer pattern detection with advanced AI capabilities, providing intelligent analysis of incident patterns and automated code-context integration for more accurate remediation.

ML Pattern Refinement Components

AI-Powered Analysis Pipeline:

• GeminiPromptEngine: Advanced prompt management with structured templates and few-shot learning capabilities
• GeminiEnhancedPatternDetector: Ensemble pattern detection combining rule-based and AI-driven classification
• GeminiResponseCache: Intelligent response caching with similarity-based matching to optimize API costs
• PatternContextExtractor: Context extraction from incident data with code repository integration

Quality Assurance & Performance:

• LogQualityValidator: Pre-processing validation ensuring high-quality input data for AI analysis
• LogSanitizer: Sensitive data sanitization removing PII, tokens, and credentials before AI processing
• ModelPerformanceMonitor: Real-time monitoring of AI model performance and drift detection
• CostTracker: API usage monitoring with budget controls and cost optimization

Resilience & Reliability:

• AdaptiveRateLimiter: Smart rate limiting with circuit breaker patterns for API stability
• Structured output validation using Pydantic schemas for consistent AI responses
• Comprehensive error handling and fallback mechanisms

ML Pattern Refinement Capabilities

graph TB
    subgraph "ML Pattern Refinement Pipeline"
        LS[Log Stream] --> LQV[Log Quality<br/>Validator]
        LQV --> LSan[Log Sanitizer<br/>PII Removal]
        LSan --> PCE[Pattern Context<br/>Extractor]
        PCE --> |Context Data| GEPD[Gemini Enhanced<br/>Pattern Detector]

        subgraph "AI Processing Layer"
            GEPD --> GPE[Gemini Prompt<br/>Engine]
            GPE --> |Structured Prompts| GF[Gemini Flash<br/>Fast Classification]
            GPE --> |Deep Analysis| GP[Gemini Pro<br/>Detailed Analysis]
        end

        subgraph "Performance & Cost Management"
            GF --> GRC[Response Cache<br/>Similarity Matching]
            GP --> GRC
            GRC --> CT[Cost Tracker<br/>Budget Control]
            CT --> MPM[Performance<br/>Monitor]
        end

        subgraph "Code Integration"
            PCE --> |Repository Data| CCE[Code Context<br/>Extractor]
            CCE --> |Static Analysis| GPE
        end
    end

    classDef aiComponent fill:#e1f5fe,stroke:#01579b,stroke-width:2px
    classDef qualityComponent fill:#f3e5f5,stroke:#7b1fa2,stroke-width:2px
    classDef performanceComponent fill:#e8f5e8,stroke:#388e3c,stroke-width:2px

    class GPE,GEPD,GF,GP,GRC aiComponent
    class LQV,LSan,PCE,CCE qualityComponent
    class CT,MPM performanceComponent

Loading

Enhanced Pattern Analysis:

• Multi-model AI strategy leveraging both Gemini Flash (speed) and Gemini Pro (accuracy)
• Similarity-based response caching reducing API costs by up to 60% while maintaining accuracy
• Context-aware analysis incorporating recent code changes, dependencies, and repository structure
• Structured output schemas ensuring consistent, validated AI responses

Quality & Cost Controls:

• Pre-processing validation filtering low-quality logs before expensive AI analysis
• Intelligent caching system matching similar incident contexts to avoid redundant API calls
• Real-time cost monitoring with configurable budget limits and alerts
• Performance drift detection identifying model degradation over time

Security & Compliance:

• Comprehensive PII sanitization removing sensitive data (emails, IPs, tokens, API keys)
• Pattern-based credential detection and replacement before AI processing
• Audit logging for all AI interactions and data transformations

This ML enhancement layer transforms raw log analysis into intelligent, context-aware incident detection with automated code repository integration, significantly improving remediation accuracy while optimizing operational costs.

Documentation

For detailed information on the Gemini SRE Agent, please refer to the following documentation sections:

• Quick Start Guide: Get the agent up and running in 15 minutes.

• Architecture Overview: Understand the core components and data flow of the agent.

• 🚀 New Log Ingestion System Migration Guide: Complete guide to migrating to the new enterprise-grade log ingestion system with feature flags, monitoring, and zero-downtime deployment.

• Unified Enhanced Code Generation System: Comprehensive guide to the complete AI-powered code generation pipeline.

• ML Pattern Refinement System: Comprehensive guide to the Gemini AI enhancement layer.

• Enhanced Prompt Generation System: Deep dive into the revolutionary dynamic prompt generation capabilities.

• GCP Infrastructure Setup Guide: Instructions for setting up necessary Google Cloud infrastructure.

• Setup and Installation: A comprehensive guide to getting the project up and running.

• Configuration Guide: Learn how to customize the agent's behavior with the enhanced type-safe configuration system.

• Deployment Guide: Instructions for deploying the agent to Google Cloud Run and other environments.

• Multi-Environment Guide: Strategies for managing the agent across different environments.

• Security Guide: Best practices and considerations for securing the agent.

• Performance Tuning Guide: Recommendations for optimizing agent performance and cost.

• Operations Runbook: Guidelines for operating, monitoring, and maintaining the agent.

• Logging and Flow Tracking: Complete guide to the flow tracking system and structured logging format.

• Error Handling System: Comprehensive guide to the robust error handling, circuit breakers, retry mechanisms, and graceful degradation.

• Error Handling Quick Reference: Quick reference guide for developers using the error handling system.

• Troubleshooting Guide: Systematic approaches for debugging issues using flow tracking and execution path tracing.

• Development Guide: Information for contributors, including testing, code style, and contributing.

Getting Started (Quick Overview)

To quickly get started, ensure you have Python 3.12+ and uv installed. Clone the repository, install dependencies with uv sync, authenticate your gcloud CLI, and set your GITHUB_TOKEN environment variable. Then, explore config/config.yaml to define your monitoring services.

New Log Ingestion System: The agent now features a comprehensive new log ingestion system with enterprise-grade monitoring and resilience. Enable it with:

export USE_NEW_INGESTION_SYSTEM=true
export ENABLE_MONITORING=true
export ENABLE_LEGACY_FALLBACK=true

You can run the agent locally with python main.py or deploy it to Cloud Run using the provided deploy.sh script. The new system provides seamless migration with full backward compatibility.

Contributing

We welcome contributions! Please see the Development Guide for details on how to get involved.

License

This project is licensed under the MIT License - see the LICENSE file for details.