Project Description

An enterprise-style AI Customer Support Copilot that uses Hybrid RAG to provide grounded customer support responses from a verified knowledge base. The system combines vector search, BM25 keyword search, and Cross-Encoder reranking to improve retrieval quality, while guardrails help prevent prompt injection, jailbreaks, off-topic requests, and unsupported responses.

The platform also supports human-in-the-loop escalation, ticket management, conversation history, and employee-side support workflows.

Project Architecture

                     ┌──────────────────────┐
                     │      React UI        │
                     │ Customer / Employee  │
                     └──────────┬───────────┘
                                │
                                ▼
                     ┌──────────────────────┐
                     │      FastAPI         │
                     │      Backend         │
                     └──────────┬───────────┘
                                │
                ┌───────────────┴────────────────┐
                │                                │
                ▼                                ▼
         ┌─────────────┐                 ┌──────────────┐
         │  Guardrails │                 │ Conversation │
         │ Input/Output│                 │   Service    │
         └──────┬──────┘                 └──────────────┘
                │
                ▼
         ┌──────────────┐
         │ Hybrid RAG   │
         └──────┬───────┘
                │
      ┌─────────┼─────────┐
      ▼         ▼         ▼
  ┌───────┐ ┌────────┐ ┌─────────────┐
  │Qdrant │ │Elastic-│ │Cross-Encoder│
  │Vector │ │search  │ │  Reranker   │
  │Search │ │ BM25   │ │             │
  └───────┘ └────────┘ └──────┬──────┘
                              │
                              ▼
                     ┌─────────────────┐
                     │   Confidence    │
                     │    Scoring      │
                     └────────┬────────┘
                              │
                     ┌────────┴────────┐
                     ▼                 ▼
              ┌─────────────┐  ┌──────────────┐
              │  Groq LLM   │  │   Human      │
              │  Response   │  │  Escalation  │
              └─────────────┘  └──────┬───────┘
                                     │
                                     ▼
                              ┌──────────────┐
                              │ Ticket System│
                              └──────────────┘
                                     │
                                     ▼
                              ┌──────────────┐
                              │  PostgreSQL  │
                              └──────────────┘

Current Development

Completed

• FastAPI backend with modular service architecture
• JWT-based authentication and protected routes
• PostgreSQL database integration
• Customer chat sessions and conversation history
• Hybrid RAG pipeline
  • Qdrant vector retrieval
  • Elasticsearch BM25 keyword retrieval
  • Reciprocal Rank Fusion / result merging
  • Cross-Encoder reranking
• Grounded LLM responses using retrieved knowledge
• Groq LLM integration
• Human support ticket management
• Employee support workflow
• Confidence score calculation for retrieval results
• Initial confidence threshold (0.60) for identifying potentially low-confidence responses
• Automatically generating support tickets from escalated conversations

The results folder contain the retireval results and score

In Progress

• Automatic human escalation based on confidence score
• Automatic conversation summarization during escalation
• Improving confidence-score calibration using real retrieval results
• Guardrails failed-Hindering the response(Need to fix)
• Knowledge-base learning from verified human resolutions
• Customer feedback and AI response evaluation
• Better Frontend