Skip to main content
The diagnosis node analyzes detected failures using a combination of LLM reasoning, historical memory, and RAG-based knowledge retrieval to determine root causes and recommend solutions.

​
Overview

When a service failure is detected, the diagnosis node leverages multiple intelligence sources to understand the problem:
  • LLM Analysis: OpenAI GPT models for reasoning
  • Memory System: Historical successes and failures
  • Knowledge Base: RAG-based documentation retrieval
  • Context: Prior diagnosis attempts and retry counts
The diagnosis node is designed to learn from past incidents, avoiding previously failed solutions and prioritizing known working fixes.

​
Architecture

llm = ChatOpenAI(model=config.MODEL_NAME, temperature=config.TEMPERATURE)
The node uses LangChain’s ChatOpenAI integration for structured LLM interactions.

​
Diagnosis Workflow

1

Context Gathering

Extracts error information from the agent state:
error = state.get("current_error", "")
service = state.get("affected_service", "desconocido")
prior_logs = state.get("diagnosis_log", [])
retry_count = state.get("retry_count", 0)
2

Memory Consultation

Queries the memory system for relevant historical data:
failed_commands = memory.get_failed_commands(error)
if failed_commands:
    memory_consulted = True
    memory_context += "COMANDOS QUE YA FALLARON (NO repetir):\n"
    for cmd in failed_commands:
        memory_context += f"- {cmd}\n"
Also searches for similar successful resolutions:
similar = memory.find_similar(error)
if similar and similar["success"]:
    memory_context += f"\nSolucion exitosa previa: {similar['command']}\n"
3

Knowledge Base Query

Retrieves relevant documentation using RAG:
if kb:
    log("diagnose", "Consultando base de conocimiento (RAG)...")
    rag_context = kb.query(f"How to fix: {error}")
4

LLM Analysis

Constructs a comprehensive prompt and requests diagnosis:
messages = [
    SystemMessage(content=(
        "Eres Sentinel AI, un agente DevOps autonomo.\n"
        "Analiza el error y proporciona un diagnostico BREVE (maximo 3 lineas).\n"
        f"\nServicio afectado: {service}\n"
        f"\n{memory_context}"
        f"\nDocumentacion tecnica:\n{rag_context[:1000]}\n"
    )),
    HumanMessage(content=f"Error: {error}")
]

response = llm.invoke(messages)
diagnosis = response.content.strip()

​
Memory Integration

The diagnosis node heavily relies on the memory system to avoid repeating mistakes:

Failed Commands

Retrieves commands that previously failed for the same error and explicitly instructs the LLM not to repeat them.

Successful Solutions

Finds similar past incidents that were successfully resolved and prioritizes those approaches.
The system explicitly forbids the LLM from suggesting commands that have already failed, as defined in the prompt:“REGLAS CRITICAS: NO sugieras comandos que ya fallaron (listados arriba).”

​
System Prompt Design

The diagnosis prompt includes critical constraints: 
SystemMessage(content=(
    "Eres Sentinel AI, un agente DevOps autonomo.\n"
    "Analiza el error y proporciona un diagnostico BREVE (maximo 3 lineas).\n"
    "Indica la causa probable y la solucion recomendada.\n"
    f"\nHistorial de intentos:\n{chr(10).join(prior_logs[-3:]) if prior_logs else 'Primer intento.'}\n"
    "\nREGLAS CRITICAS:"
    "\n1. NO sugieras comandos que ya fallaron (listados arriba)."
    "\n2. Si 'service' o 'apt-get' fallan, prueba alternativas como 'systemctl', 'dmesg', o verificar ficheros de log especificos."
))
The diagnosis is intentionally kept brief (maximum 3 lines) to focus on actionable insights rather than verbose explanations.

​
State Updates

The diagnosis node returns updated state: 
return {
    "current_step": "diagnose",
    "diagnosis_log": prior_logs + [diagnosis],
    "memory_consulted": memory_consulted
}

diagnosis_log

Appends the new diagnosis to the history

current_step

Marks the workflow position

memory_consulted

Tracks whether historical data was available

​
Retry Context

The diagnosis adapts based on retry attempts: 
f"\nHistorial de intentos:\n{chr(10).join(prior_logs[-3:]) if prior_logs else 'Primer intento.'}\n"
This provides the LLM with awareness of previous attempts, enabling progressive problem-solving strategies.

​
RAG Integration

When available, the knowledge base provides technical documentation: 
if kb:
    rag_context = kb.query(f"How to fix: {error}")
else:
    log("warning", "Base de conocimiento no disponible.")
The RAG context is truncated to 1000 characters to stay within LLM token limits while providing relevant documentation.

​
Implementation Location

Source: src/agent/nodes/diagnose.py:14

​
Next Steps

After diagnosis, the workflow moves to planning where the diagnosis is translated into executable remediation commands.

Build docs developers (and LLMs) love

Get started for freeTalk to us