🧠 Company Knowledge Assistant

A Retrieval-Augmented Generation (RAG) system that enables intelligent Q&A over company documents. Built with modern AI stack to provide accurate, source-grounded answers from your knowledge base.

✨ Features

• 📚 Multi-format Document Support: PDF, Markdown, Word documents, and plain text
• 🔍 Semantic Search: Vector embeddings with HNSW indexing for fast retrieval
• 🎯 Intelligent Reranking: Cohere reranking for improved relevance
• 💾 Semantic Caching: Redis-based caching to reduce API calls for similar questions
• 🌐 Web Interface: Clean, interactive UI for asking questions
• 📊 RAG Evaluation: Built-in Ragas metrics for system assessment
• 🗂️ Category Organization: Documents organized by type (FAQs, guides, policies, etc.)

🛠️ Tech Stack

• 🐍 Python + FastAPI: High-performance async API framework
• 🔗 LangChain: Orchestration framework for LLM applications
• 🤖 OpenAI API: GPT-4o-mini for generation, embeddings for vector search
• 🐘 PostgreSQL + pgvector: Vector database with HNSW indexing
• 📦 Redis: Semantic caching layer
• 🐳 Docker Compose: Container orchestration
• 🎯 Cohere Rerank: Advanced reranking for retrieval quality
• 📈 Ragas: RAG evaluation framework

🏗️ Pipeline

Data Indexing

1. Load Documents

   • Recursively scans data/ directory
   • Supports: .pdf, .md, .docx, .doc, .txt

2. Chunk Documents

   • Splits documents into chunks with RecursiveCharacterTextSplitter
   • Preserves metadata (category, source)

3. Embed & Store

   • OpenAI embeddings for vector representation
   • Stores in PostgreSQL with pgvector
   • Builds HNSW index for fast similarity search

Retrieval & Generation

1. Vector Retrieval

   • Retrieves top-k documents (default: 5) based on cosine similarity
   • Optional category filtering

2. Reranking

   • Cohere rerank-multilingual-v3.0 narrows to top 3 most relevant chunks

3. Redis Semantic Cache Check

   • Skips LLM call if cached answer exists for similar questions

4. Answer Generation

   • GPT-4o-mini generates answer grounded in retrieved context
   • Returns answer, sources, and contexts

🚀 Getting Started

Prerequisites

• Docker
• API Keys: OpenAI, Cohere (optional), LangSmith (optional)

Setup

1. Clone the repository

   git clone <repo-url>
   cd rag-company-knowledge-assistant

2. Configure environment variables

   Create a .env file:

   # Database
   DATABASE_URL=postgresql+asyncpg://postgres:postgres@postgres:5432/postgres
   
   # OpenAI
   OPENAI_API_KEY=your_openai_key
   
   # LangSmith (optional, for tracing)
   LANGSMITH_TRACING_V2=true
   LANGSMITH_PROJECT=cka
   LANGSMITH_API_KEY=your_langsmith_key
   
   # Data & Configuration
   DATA_DIR=data
   RETRIEVAL_K=5
   
   # Redis
   REDIS_URL=redis://cka-redis:6379/0
   
   # Cohere (optional, for reranking)
   CO_API_KEY=your_cohere_key

3. Start services

   docker compose up

4. Access the application

   • Web UI: http://localhost:8000

5. Ingest your documents

   • Place documents in data/ directory (organized by category)
   • Click "Ingest Data" in the web UI, or POST to /ingest

Cleanup

To stop services and remove volumes:

docker compose down -v

🔧 API Endpoints

• GET / - Web UI
• POST /ingest - Trigger document ingestion
• GET /ingest/status - Check ingestion status
• POST /ask - Ask a question

  {
    "question": "How many PTO days are full-time employees entitled to in a calendar year?"
  }

🧪 Evaluation

Run RAG evaluation using Ragas metrics:

python app/eval_ragas.py

Metrics evaluated:

• Faithfulness: Answer grounded in context
• Answer Relevancy: Relevance of answer to question
• Context Precision: Precision of retrieved contexts
• Context Recall: Recall of relevant information

📝 License

See LICENSE file for details.