# Technical System Architecture Diagram: Automated CI Failure Resolution This document provides a comprehensive technical extraction of the provided system architecture diagram, which outlines a two-phase (Offline/Online) process for automated Root Cause Analysis (RCA) and solution execution for Continuous Integration (CI) pipeline failures. --- ## 1. High-Level Segmentation The diagram is divided into four primary quadrants based on two axes: * **Vertical Axis:** Separates **Offline** (top) preparation tasks from **Online** (bottom) real-time execution. * **Horizontal Axis:** Separates **Stage 1: Root Cause Analysis** (left) from **Stage 2: Solution Generation & Execution** (right). --- ## 2. Offline Preparation Phase ### Prep. 1: Success Log Template Deduplication (Top Left) * **Input:** "Old success logs" (represented by grey icons) and "Latest $x$ success logs" (represented by green icons). * **Process:** The logs are processed via a **Drain Algorithm**. * **Output:** A "Success log template" (green icon), which is fed into the Online Stage 1. ### Prep. 2: Offline Knowledge Base Construction (Top Right) * **Inputs:** * "Enterprise Document" (T-icon) * "On-call Q&A Records" (QA-icon) * **Process Flow:** 1. **Chunking:** Documents are broken down into smaller segments. 2. **Embedding Model:** Chunks are converted into vector representations. 3. **Vector Database:** The final storage for retrieved knowledge, used in Online Stage 2. --- ## 3. Online Execution Phase ### Stage 1: Root Cause Analysis (Bottom Left) This stage processes an **Original Failed CI Log** through five numbered steps: 1. **Key Log Filtering:** Uses a filter icon to isolate: * Keywords * Log Tail * Log Diff (The difference between the failed log and the success log template). 2. **Key Log Expansion:** Expands context around identified issues: * Previous Operation Extension * Later Consequence Expansion 3. **Token Overflow Pruning:** A "scissors" icon indicates optimization of data for LLM context windows: * Initial Weight Assignment * Block Weight Enhancement * Context Window Expansion * Density Based Ranking 4. **Input Assembly:** Combines **Critical Log Blocks** (red icon) with an **RCA Prompt Template** (green circle). 5. **LLM Processing:** An LLM generates a **Root Cause Analysis Report** (document icon with an 'i' info badge). ### Stage 2: Solution Generation & Execution (Bottom Right) This stage takes the RCA report and attempts to fix the pipeline through three numbered steps: 6. **Knowledge Retrieval:** Interacts with the Vector Database using: * Query Generation * Multi-Route Retrieval * Reranker * URL Mapping * **Output:** "Retrieved Knowledge" (T-icon). 7. **Executable Solution Generation:** * **Inputs:** Root Cause Analysis Report + Solution Generation Prompt Template + Retrieved Knowledge. * **Process:** LLM generates an **Executable Solution**. 8. **Tool Execution & Pipeline Rerun:** * **Inputs:** Executable Solution + Customized Tools + Tool Calling Prompt Template. * **Process:** LLM executes tools and reruns the pipeline. * **Final State:** **CI Pipeline Fixed** (blue checkmark icon). --- ## 4. Component and Flow Summary | Component | Description | | :--- | :--- | | **Data Sources** | Success logs, Failed CI logs, Enterprise docs, Q&A records. | | **Algorithms/Models** | Drain Algorithm, Embedding Model, LLM (used in 3 distinct steps). | | **Data Structures** | Success log templates, Critical log blocks, Vector Database, RCA Reports. | | **Logic Flow** | Log Filtering $\rightarrow$ Context Expansion $\rightarrow$ Pruning $\rightarrow$ RCA $\rightarrow$ Knowledge Retrieval $\rightarrow$ Solution Generation $\rightarrow$ Execution. | ## 5. Textual Transcriptions ### Labels and Headers * **Offline / Online:** Primary horizontal dividers. * **Prep. 1:** Success Log Template Deduplication. * **Prep. 2:** Offline Knowledge Base Construction. * **Stage 1:** Root Cause Analysis. * **Stage 2:** Solution Generation & Execution. ### Numbered Process Steps 1. **Key Log Filtering** 2. **Key Log Expansion** 3. **Token Overflow Pruning** 4. **Critical Log Blocks** 5. **Root Cause Analysis Report** 6. **Knowledge Retrieval** 7. **Executable Solution** 8. **Tool Execution & Pipeline Rerun** ### Embedded Technical Terms * **Drain Algorithm:** Used for log templating. * **Log Diff:** Comparison between success and failure. * **Token Overflow Pruning Sub-tasks:** Initial Weight Assignment, Block Weight Enhancement, Context Window Expansion, Density Based Ranking. * **Knowledge Retrieval Sub-tasks:** Query Generation, Multi-Route Retrieval, Reranker, URL Mapping. * **Final Outcome:** CI Pipeline Fixed.

# Technical Document Extraction: CI Pipeline Root Cause Analysis & Solution Generation System ## Overview The diagram illustrates a two-stage CI/CD pipeline analysis system combining offline preprocessing with online LLM-driven root cause analysis and solution generation. The system is divided into **Offline** (pink) and **Online** (blue) components, with bidirectional data flow between stages. --- ## Offline Processing (Preprocessing) ### Prep. 1: Success Log Template Deduplication 1. **Input**: Original Failed CI Log 2. **Processing Steps**: - **Keyword Filtering**: Extracts keywords and log tail from CI logs - **Log Diff Analysis**: Compares with latest success logs - **Token Overflow Pruning**: Applies context window expansion and density-based ranking 3. **Output**: Critical Log Blocks + RCA Prompt Template ### Prep. 2: Offline Knowledge Base Construction 1. **Input**: Enterprise Document 2. **Processing Steps**: - **Chunking**: Splits documents into embeddable units - **Embedding Model**: Converts chunks to vector representations - **Vector Database**: Stores embeddings for retrieval 3. **Output**: Vector Database with enhanced knowledge blocks --- ## Online Processing (Execution) ### Stage 1: Root Cause Analysis 1. **Input**: Critical Log Blocks + RCA Prompt Template 2. **Processing Steps**: - **LLM Initial Analysis**: Generates preliminary root cause hypotheses - **Knowledge Retrieval**: Cross-references with vector database 3. **Output**: Root Cause Analysis Report ### Stage 2: Solution Generation & Execution 1. **Input**: Root Cause Analysis Report 2. **Processing Steps**: - **Solution Generation**: Creates executable solutions using LLM - **Customized Tools**: Maps solutions to available CI tools - **Tool Execution**: Implements solutions via pipeline rerun 3. **Output**: CI Pipeline Fixed status --- ## Key Components & Flow 1. **Offline-Online Integration**: - Offline components prepare templates and knowledge bases - Online components use LLMs for dynamic analysis and solution generation 2. **Data Flow**: - Failed CI logs → Offline preprocessing → Online analysis → Solution execution - Knowledge base updates feed into both stages --- ## Technical Architecture - **Offline System**: - Success Log Template Deduplication Engine - Knowledge Base Construction Pipeline - **Online System**: - LLM-Driven Root Cause Analysis Module - Automated Solution Generation Framework - CI Pipeline Execution Controller --- ## Critical Data Points 1. **Log Processing**: - Keyword extraction from failed CI logs - Temporal comparison with latest success logs 2. **Knowledge Base**: - Embedding model converts enterprise documents to vector representations - Vector database enables semantic search for solution retrieval 3. **LLM Workflow**: - Initial root cause hypothesis generation - Solution proposal and tool mapping - Pipeline rerun execution --- ## Process Flow Diagram ``` [Original Failed CI Log] → [Keyword Filtering] → [Token Overflow Pruning] → [Critical Log Blocks] → [RCA Prompt Template] → [LLM Root Cause Analysis] → [Knowledge Retrieval] → [Solution Generation] → [Customized Tools] → [Tool Execution] → [CI Pipeline Fixed] ``` --- ## Implementation Notes 1. **Offline Preprocessing**: - Success log template deduplication reduces noise - Knowledge base construction enables context-aware analysis 2. **Online Analysis**: - LLM combines log analysis with knowledge base retrieval - Solution generation uses tool-specific prompt templates 3. **Execution**: - Customized tools interface with CI/CD pipelines - Pipeline rerun validates solution effectiveness This system architecture enables automated root cause analysis and solution implementation for CI/CD failures through integrated offline preprocessing and online LLM-driven processing.