Skip to content

README.md

Latest commit

 

History

History
55 lines (35 loc) · 3.04 KB

README.md

File metadata and controls

55 lines (35 loc) · 3.04 KB
Raw

AI Support Triage Agent - /code

This directory contains the core logic for the Support Ticket Triage AI Agent. It implements a highly optimized Hybrid RAG (Retrieval-Augmented Generation) pipeline using Azure OpenAI to process, analyze, and resolve customer support tickets intelligently.

1 2

🏗️ Architecture & Modules:

The pipeline is designed to be modular, separating data orchestration, retrieval, LLM interaction, and risk assessment into distinct files.

Core Execution

  • main.py
    The primary entry point for the application. It initializes the environment, authenticates with Azure OpenAI, and triggers the ticket processing loop.
  • processor.py
    The orchestration engine. It reads the raw support_tickets.csv, sanitizes the inputs (handling NaNs and empty rows), coordinates the flow between the retriever and the LLM, and formats the final output.csv.

Intelligence & RAG:

  • retriever.py
    The intelligence backbone implementing the Hybrid RAG system. It combines scikit-learn's TF-IDF for exact-keyword (sparse) retrieval with Azure OpenAI's text-embedding-3-large for semantic (dense) retrieval. It features intelligent chunking, 8,000-character truncation to prevent token overflow, and automatic caching logic.

  • build_index.py
    A standalone utility script to generate and locally cache the document embeddings (embeddings.npy). Running this script once saves thousands of API tokens and drastically speeds up execution times.

  • llm.py
    Manages communication with the Azure OpenAI Chat model (gpt-4o). Contains the heavily engineered system prompt that forces the agent to extract structured JSON, securely reason through sparse context, and aggressively minimize unnecessary human escalations.

Analytics & Processing

  • risk.py
    A safety net module containing rule-based keyword scanning logic. It instantly detects critical issues (e.g., fraud, security breaches, legal threats) and forces an immediate escalation, bypassing the LLM to ensure safety.

  • classifier.py
    A lightweight helper module containing logic to determine product/module responsibilities and basic keyword-based request typing.

  • utils.py
    Contains essential utility functions, primarily focusing on robust JSON parsing. It gracefully handles malformed LLM responses and ensures the pipeline never crashes during string-to-dictionary conversion, providing a safe "escalated" fallback if parsing fails.

🚀 Usage

To execute the pipeline and process the support tickets, ensure your .env file is properly configured at the root of the project, then run:

python code/main.py

(Optional) If the data/ document corpus has been modified, you should update the embedding cache by running:

python code/build_index.py