neuro-dotcom/enterprise-rag-agent

🤖 Enterprise RAG Support Agent

An autonomous, cloud-deployed AI Support Agent built for Telegram. This bot utilizes Retrieval-Augmented Generation (RAG) to provide highly accurate, hallucination-free technical support strictly grounded in official corporate documentation.

⚙️ Architecture & Tech Stack

• —AI Engine: Google Gemini 2.5 Flash (for rapid, reasoning-based generation)
• —Embeddings: Gemini-Embedding-001
• —Vector Database: ChromaDB (for semantic memory and document retrieval)
• —Interface: Telegram Bot API (pyTelegramBotAPI)
• —Infrastructure: Docker, Hugging Face Spaces
• —State Management: Local persistent JSON ledger for user state and RBAC.
• —Network Ops: Flask-based health checks, IPv4 forced routing, and hardcoded DNS bypass for cloud resilience.

🚀 Key Features

1. 1.Strict Context Grounding: The bot is constrained to answer only using the provided knowledge_base.md. Out-of-domain questions are respectfully escalated to human support.
2. 2.Semantic Chunking & Ingestion: The system automatically splits manuals into logical chunks, converts them to high-dimensional vectors, and stores them in ChromaDB at runtime. Updating the AI's knowledge simply requires modifying the markdown file.
3. 3.Role-Based Access Control (RBAC): Secured via Telegram Chat ID verification to prevent unauthorized API quota drain. Includes a built-in /admin dashboard.
4. 4.Persistent Demo Quota System: Admins can securely issue temporary "Guest Passes." This allows recruiters or tech leads exactly 2 queries (ideal for testing 1 in-domain and 1 out-of-domain request) before automatically revoking access via the JSON state ledger.
5. 5.Zero-Downtime Resilience: Utilizes custom socket patching and background HTTP ping servers to bypass strict cloud DNS limitations.

🛠️ How It Works (The RAG Pipeline)

1. 1.Load & Embed: The system reads knowledge_base.md and checks if the ChromaDB collection is empty. If so, it embeds the chunks and stores them locally.
2. 2.Query: A user sends a support question via Telegram.
3. 3.Retrieve: The query is embedded, and ChromaDB performs a similarity search to find the closest matching technical manual chunks.
4. 4.Generate: The matching text is injected into a strict system prompt, forcing Gemini to generate an accurate, localized response based only on the retrieved context.

Developed by neuro-dotcom as an AI Ops portfolio project.