Enterprise-KB-AI

An enterprise-grade internal knowledge and incident handling platform built on top of RAG.

Instead of a generic "chat with docs" demo, this project focuses on a real operational workflow:

Intake -> Triage -> Investigation -> Escalation -> Resolution -> Knowledge Capture

Repositories

• GitHub (main): Koala-la-la/AI-RAG-System
• GitCode (mirror): 2501_94493648/KB-AI

Why This Project

In real teams, large volumes of API docs, runbooks, SOPs, and FAQs create friction:

• Slow onboarding for new teammates
• Inconsistent troubleshooting responses
• Repeated "same issue, same manual search" cycles
• Weak incident traceability and poor knowledge retention

This project turns RAG into a workflow-centric product, not just a chatbot.

Product Scope

Target Users

• Technical support / customer support teams
• SRE / operations teams
• On-call engineers
• Product operations teams

Core Product Goal

Provide a single internal workspace to:

1. Register and process cases
2. Retrieve grounded answers from internal knowledge
3. Enforce case lifecycle transitions
4. Track SLA and audit trails
5. Evaluate quality and iterate continuously

Key Features

1) Case Workflow Engine (Core)

Cases are first-class objects with strict status transitions:

• new
• triaged
• investigating
• pending_escalation
• escalated
• resolved
• archived

Illegal transitions are rejected by backend validation.

2) SLA Tracking

SLA deadlines are auto-generated by priority:

Priority	First Response SLA	Resolution SLA
low	240 minutes	72 hours
medium	120 minutes	24 hours
high	30 minutes	8 hours
critical	10 minutes	2 hours

The system exposes breach flags for both response and resolution.

3) Audit Timeline

Every important action is logged per case:

• case created
• case updated
• status transitioned
• messages appended
• case deleted

Each event includes actor, timestamp, event type, and details.

4) Enterprise Access Control

• Roles: member, admin
• Document visibility: private, team, org
• Retrieval and evaluation are both permission-aware

5) Advanced RAG Pipeline

• Query rewrite (context-aware)
• Case-aware routing
• Hybrid retrieval (semantic + lexical)
• Reranking
• Grounded-answer constraints and citation exposure

6) Evaluation Loop

• Generated multi-document benchmark suite
• JSON / Markdown report outputs
• Pass rate and quality trend tracking

System Architecture

flowchart LR
    UI[React UI] --> API[FastAPI API]

    API --> AUTH[Auth and RBAC]
    API --> WF[Case Workflow Engine]
    API --> RAG[RAG Pipeline]
    API --> EVAL[Evaluation Engine]

    WF --> STORE[Conversation and Case Store]
    WF --> AUDIT[Audit Timeline Store]

    RAG --> REWRITE[Query Rewrite]
    RAG --> ROUTER[Case Router]
    RAG --> RETRIEVE[Hybrid Retrieval]
    RETRIEVE --> MILVUS[Milvus]
    RAG --> RERANK[Rerank and Grounding]
    RAG --> LLM[LLM Response]

    API --> REG[Document Registry]
    REG --> INGEST[Upload and Chunking]
    INGEST --> MILVUS

    API --> REDIS[Redis Cache]

Loading

Tech Stack

• Backend: FastAPI, Pydantic, LangGraph, LangChain
• Frontend: React, Vite, Axios
• Vector DB: Milvus
• Cache: Redis
• Model Serving: Ollama (default), optional HuggingFace embeddings
• Document Processing: PyPDF
• Evaluation: custom benchmark runner

Repository Structure

app/
  api/
    auth_api.py
    chat_api.py
    conversation_api.py
    document_api.py
    evaluation_api.py
    upload_api.py
  auth/
  cache/
  conversation/
    store.py
  evaluation/
    benchmark.py
  knowledge/
    access.py
    document_manager.py
    document_registry.py
    embedder.py
    milvus_store.py
    splitter.py
  llm/
    ollama_client.py
  rag/
    graph.py
    prompt.py
    ranking.py
    retriever.py
    router.py
web/react-ui/src/
  pages/
    Dashboard.jsx
    Chat.jsx
    Documents.jsx
    Evaluation.jsx
    Login.jsx
data/
  auth_users.json
  auth_sessions.json
  document_registry.json
  conversations/
  uploads/
  eval/

Quick Start

1) Prerequisites

• Python 3.10+
• Node.js 18+
• Docker Desktop

2) Install Dependencies

pip install -r requirements.txt

cd web/react-ui
npm install

3) Start Infrastructure (Milvus + Redis)

docker compose up -d redis etcd minio milvus

4) Configure Environment

Use .env.example as baseline. Key vars:

• MILVUS_HOST, MILVUS_PORT
• REDIS_HOST, REDIS_PORT
• OLLAMA_BASE_URL, OLLAMA_MODEL
• EMBED_PROVIDER (ollama or huggingface)

5) Run Backend

python -m uvicorn app.main:app --reload

6) Run Frontend

cd web/react-ui
npm run dev

Open: http://localhost:5173

Main API Endpoints

Auth

• POST /api/auth/register
• POST /api/auth/login
• POST /api/auth/logout

Documents

• POST /api/upload
• GET /api/documents
• DELETE /api/documents
• POST /api/documents/reindex

Cases and Workflow

• GET /api/conversations
• POST /api/conversations
• PATCH /api/conversations/{conversation_id}
• POST /api/conversations/{conversation_id}/transition
• GET /api/conversations/{conversation_id}/messages
• GET /api/conversations/{conversation_id}/timeline
• DELETE /api/conversations/{conversation_id}

Chat

• POST /api/chat

Evaluation

• GET /api/evaluation/latest-report
• POST /api/evaluation/run-generated-suite

Benchmark Commands

List indexed sources:

python -m app.evaluation.benchmark --list-sources --user-id user001 --kb-id defaultkb

Generate suite template:

python -m app.evaluation.benchmark --generate-suite-template --user-id user001 --kb-id defaultkb --benchmarks-dir data/eval/benchmarks/generated --suite-output data/eval/generated_suites/user001_defaultkb.json

Run suite:

python -m app.evaluation.benchmark --suite data/eval/generated_suites/user001_defaultkb.json

Troubleshooting

Milvus connection error on upload

If you see:

Fail connecting to server on 127.0.0.1:19530

Run:

docker compose ps
docker compose up -d redis etcd minio milvus

HuggingFace download reset (WinError 10054)

• Retry downloads
• Prefer EMBED_PROVIDER=ollama for stable local flow
• Pre-cache model assets if needed

Chinese text appears garbled in PowerShell

Usually this is terminal encoding, not data corruption. Verify with UTF-8 readers.

Roadmap

• Assignee and escalation target management (R&D / Ops)
• SLA alerts and notification center
• Auto-promote failed cases into FAQ/SOP candidates
• Org-level knowledge operations dashboard

Contribution

Issues and PRs are welcome.

If this project helps you, consider giving it a star.