Skip to content

CSV_AI_GUIDANCE.md

Latest commit

 

History

History
64 lines (49 loc) · 2.19 KB

CSV_AI_GUIDANCE.md

File metadata and controls

64 lines (49 loc) · 2.19 KB

CSV Ticket Guidance For AI Agents

This project exposes ticket data from csv/data.csv through REST, MCP, and agent tools. Use this guide when building prompts or automation for usecase demo idea generation.

1. Data Source Rules

  • Source of truth: csv/data.csv (read-only in current implementation).
  • Backend normalizes many BMC headers into the typed Ticket model (backend/tickets.py).
  • Do not assume every CSV column is mapped. Use only exposed normalized fields.
  • Always treat missing values as unknown, not false.

2. Recommended Tool Order (MCP Or Agent)

  1. Call csv_ticket_fields to discover available fields.
  2. Call csv_ticket_stats to get high-level distribution (status, priority, city, group).
  3. Narrow data with csv_list_tickets filters (status, assigned_group, has_assignee).
  4. Use csv_search_tickets for text scenarios (problem patterns, products, cities, notes).
  5. Call csv_get_ticket only for deep dives on specific IDs.

3. MCP Tools Added In This Repo

  • csv_ticket_fields
  • csv_ticket_stats
  • csv_list_tickets
  • csv_search_tickets
  • csv_get_ticket

All tools are available via POST /mcp (tools/list and tools/call).

4. Prompting Guidance For Usecase Demo Ideas

When generating project ideas:

  • Start from real evidence in tickets (priority, volume, repeated keywords, bottlenecks).
  • Explicitly reference ticket IDs used as evidence.
  • Prefer one menu point per project idea.
  • Return both:
    • Human summary
    • Structured rows (JSON) for table rendering

Suggested output schema:

{
  "rows": [
    {
      "menu_point": "Smart Routing",
      "project_name": "Auto Assignment Optimizer",
      "summary": "Reduces unassigned high-priority incidents.",
      "agent_prompt": "Analyze unassigned critical/high tickets and propose routing rules.",
      "ticket_ids": "id1,id2,id3",
      "csv_evidence": "24 high-priority tickets without assignee in top 2 groups."
    }
  ]
}

5. Guardrails

  • Never invent tickets, IDs, or field values.
  • If the dataset is insufficient, say so clearly.
  • Keep responses deterministic and auditable: include filtering logic used.
  • Prefer concise tables over long prose when showing candidate projects.