NLP Agent Creation

Overview

Users describe what they want in plain English. The system extracts the intent, maps it to tools and a schedule, and presents a review screen before creating anything.

Creation Flow

sequenceDiagram
  participant User
  participant UI as AgentBuilder UI
  participant BE as Backend :8080
  participant AIS as AI Service :9300
  participant LLM as Groq LLM

  User->>UI: "Monitor payment errors\nand alert me daily"
  UI->>AIS: POST /api/v1/agent-builder/parse\n{description: "..."}
  AIS->>LLM: System prompt + user description
  LLM-->>AIS: Structured JSON (name, tools, schedule, alert)
  AIS-->>UI: Parsed agent config

  UI->>User: Review screen\n(editable fields)
  User->>UI: Adjust schedule → "Every 4 hours"\nClick Confirm

  UI->>BE: POST /api/v1/agents\n{...finalConfig}
  BE-->>UI: Agent created {id: 42}
  UI-->>User: "Agent created!\nFirst run in 4 hours."

AgentBuilder Parse Endpoint

POST /api/v1/agent-builder/parse accepts a free-text description and returns a draft agent config.

Example Request

{
  "description": "Every morning, check if there were any payment timeout errors overnight and send me a summary"
}

Example Response

{
  "name": "Overnight Payment Timeout Monitor",
  "description": "Checks for payment timeout errors from the previous night and summarizes findings",
  "systemPrompt": "You are a monitoring agent for the Logrite platform. Your job is to check for payment timeout errors that occurred overnight (between midnight and 8am). Search the payment_queries logs for timeout-related errors, count occurrences, identify the most affected payment methods or endpoints, and provide a concise summary with recommendations.",
  "toolsAllowed": ["semantic_search", "opensearch_query"],
  "schedule": "0 8 * * *",
  "scheduleLabel": "Every day at 8:00 AM",
  "alertCondition": null,
  "suggestedQuery": "Find payment timeout errors from last night (midnight to 8am)",
  "confidence": 0.92
}

LLM Extraction Prompt

The AI Service uses a structured extraction prompt:

You are an agent configuration extractor for Logrite, a log analysis platform.

Given a user's description of what they want an AI agent to do, extract:
1. name: Short, descriptive agent name (max 60 chars)
2. description: One sentence explaining the agent's purpose
3. systemPrompt: Full system prompt the agent will use when executing
4. toolsAllowed: Array of tools from: ["semantic_search", "opensearch_query", "get_scan_results", "analyze_logs"]
5. schedule: Cron expression (or "manual" if no schedule mentioned)
6. scheduleLabel: Human-readable schedule description
7. alertCondition: {metric, operator, threshold} or null
8. suggestedQuery: The first query the agent should ask on each run
9. confidence: Float 0.0-1.0 confidence in your extraction

Respond ONLY with valid JSON matching this schema.

Available indexes: payment_queries, external_service_calls, database_queries, exceptions
Available tools: semantic_search (neural log search), opensearch_query (aggregations/filters),
                 get_scan_results (compliance scan data), analyze_logs (AI analysis)

User description: {userDescription}

Review Screen

The AgentBuilder UI presents the parsed config in an editable form before creation:

┌─────────────────────────────────────────────┐
│ Review Your Agent                           │
│                                             │
│ Name: [Overnight Payment Timeout Monitor  ] │
│                                             │
│ Schedule: [Every day at 8:00 AM      ▼]    │
│                                             │
│ Tools allowed:                              │
│  ✅ semantic_search                         │
│  ✅ opensearch_query                        │
│  ☐ get_scan_results                        │
│  ☐ analyze_logs                            │
│                                             │
│ System prompt:                              │
│ [You are a monitoring agent... (editable)]  │
│                                             │
│ Alert condition: None                       │
│                                             │
│ [Cancel]              [Create Agent →]     │
└─────────────────────────────────────────────┘

Confidence Handling

Confidence	UI Behavior
≥ 0.85	Parsed config shown with green indicator
0.60–0.84	Parsed config shown with yellow warning — “Please review carefully”
< 0.60	Low confidence banner — “We couldn’t confidently parse this. Please fill in manually.”

Edge Cases

Ambiguous Schedule

“Check periodically” → Defaults to hourly, user must confirm.

No Tools Matched

“Alert me about problems” → Shows tool browser for manual selection.

Multiple Possible Interpretations

“Monitor errors” → Prompts: “Which service? payment, database, external calls, or all?”

Overly Broad Request

“Do everything” → Returns error: “Please describe a specific monitoring task.”