Skip to main content

​
Overview

The Incident Resolution Agent is an AI-powered assistant designed to help teams quickly resolve incidents by searching through a knowledge base of past incidents. Built with LangGraph and powered by large language models, the agent provides intelligent, context-aware responses to incident-related queries.

​
What is the Agent?

The agent is a conversational AI system that:
  • Searches Past Incidents: Retrieves relevant historical incident data using semantic search
  • Provides Contextual Answers: Generates responses based on verified knowledge and tool results
  • Learns from Examples: Uses golden examples to improve response quality
  • Maintains Conversation State: Remembers context across multiple turns in a conversation

​
Key Capabilities

​
Intelligent Query Understanding

The agent analyzes user queries and automatically determines the best approach:
  • Recognizes specific incident IDs (e.g., INC-2025-08-24-001)
  • Understands problem descriptions and error messages
  • Identifies application/system-specific queries
  • Interprets time-based requests (e.g., “incidents from last week”)
The agent has access to four specialized search tools:

lookup_incident_by_id

Fast direct lookup for specific incident IDs

search_similar_incidents

Semantic similarity search across all incident reports

get_incidents_by_application

Find incidents affecting specific applications or systems

get_recent_incidents

Retrieve incidents from a specific timeframe

​
Query Rewriting

Before calling tools, the agent automatically rewrites conversational queries into optimized search terms: 
User: "hey how to solve the issue with loan emi?"
Agent Query: "Loan EMI issue"

User: "tell me about this issue INC-2025-08-24-001"
Agent Query: incident_id="INC-2025-08-24-001"

​
Golden Example Enhancement

The agent searches for similar past conversations (golden examples) and uses them to improve response quality. This ensures consistent, high-quality answers based on verified resolutions.

​
Conversation Memory

Using PostgreSQL-backed checkpointing, the agent maintains conversation state:
  • Remembers previous messages in the conversation
  • Can answer follow-up questions without re-searching
  • Maintains session continuity across API calls

​
Use Cases

​
Incident Research

User: "What caused the payment gateway timeout last week?"
Agent: [Searches recent incidents → Returns relevant incident details]

​
Problem Resolution

User: "How do I fix HTTP 403 errors in PayU?"
Agent: [Searches similar incidents → Provides resolution steps]

​
Specific Incident Lookup

User: "Show me incident INC-2025-08-24-001"
Agent: [Direct lookup → Returns complete incident report]

​
Application Monitoring

User: "What incidents affected the Settlement & Reporting system?"
Agent: [Application search → Lists all related incidents]

​
Architecture Highlights

The agent is built on LangGraph, a framework for building stateful, multi-actor applications with LLMs. Learn more about the architecture in the Architecture section.

​
Key Components

  1. State Management: TypedDict-based state schema with message history
  2. Node Functions: Modular processing units (LLM calls, tool execution, title generation)
  3. Conditional Routing: Smart decision-making about when to use tools
  4. Persistent Checkpointing: PostgreSQL-backed conversation memory

​
Technology Stack

  • LangGraph: State graph orchestration
  • LangChain: LLM integration and tool binding
  • Qdrant: Vector database for semantic search
  • PostgreSQL: Conversation state persistence
  • Langfuse: Optional observability and tracing

​
Configuration

The agent supports multiple LLM providers:
  • Anthropic (Claude models)
  • OpenAI (GPT models)
  • Google (Gemini models)
  • Custom (Any OpenAI-compatible API)
  • Ollama (Local models)
Default settings are defined in src/copilot/config.py:51: 
DEFAULT_LLM_TEMPERATURE = 0.33
DEFAULT_LLM_MAX_RETRIES = 2

​
Response Guidelines

The agent follows strict response guidelines:
  • Cite Sources: Always references incident IDs when using tool data
  • No Fabrication: Never makes up information
  • Concise Output: Provides clear, factual responses
  • Table Limits: Never generates tables with more than 4 columns
The agent will not commit files that likely contain secrets (.env, credentials.json, etc.) and will warn users if they attempt to do so.

​
Next Steps

Architecture

Explore the LangGraph state machine and node structure

Workflow

Understand how queries are processed end-to-end

Build docs developers (and LLMs) love

Get started for freeTalk to us