🧠 Intelligent Knowledge Assistant (RAG + Agents)

📘 Overview

A retrieval‑augmented AI system that answers natural‑language questions over custom documents. It combines text ingestion, embedding, retrieval, and LLM‑based reasoning through an API and optional UI.

---

🧩 System Flow

Stage	Description
1. Documents / PDFs	Source material: policies, papers, notes, etc.
2. Ingestion & Chunking	Converts PDFs or text into manageable chunks using LangChain’s splitter.
3. Embedding & Vector DB	Generates embeddings with OpenAI or Hugging Face and stores them in FAISS or Pinecone.
4. Retriever	Performs similarity search to find the most relevant document segments.
5. LLM Generation	Uses GPT‑4 or Claude to compose coherent answers using retrieved context.
6. FastAPI Layer	Exposes a /ask endpoint for client or UI requests.
7. Streamlit / React UI	Optional web interface for interactive use or demos.

---

⚙️ Configuration

The system supports both Ollama (free, local) and OpenAI for LLM generation. Configure in .env:

# Use Ollama (FREE - runs locally)
LLM_PROVIDER=ollama
OLLAMA_BASE_URL=http://127.0.0.1:11434
OLLAMA_MODEL=deepseek-r1:8b

# Or use OpenAI (requires API credits)
# LLM_PROVIDER=openai
# OPENAI_API_KEY=sk-your-key-here

⚙️ Run Locally

Step 1: Ingest Documents and Create Embeddings

python src/ingest.py          # Load and chunk PDFs from data/
python src/embed_store.py     # Create FAISS vector store (uses free local embeddings)

Step 2: Choose Your Interface

Option A: Streamlit UI (Recommended - with conversational memory)

streamlit run ui/app.py
# Or use the shortcut:
./run_ui.sh

Features:

• 💬 Interactive chat interface
• 🧠 Conversational memory (remembers past questions)
• 📜 Conversation history sidebar
• 📄 Source document citations

Option B: FastAPI (REST API)

uvicorn src.api:app --reload

Then open http://127.0.0.1:8000/docs for the Swagger API UI.

Option C: Command Line

python src/generator.py

---

🚀 Next Steps

• Add embedding cache for performance.
• Integrate evaluation (RAGAS or LLM‑as‑a‑judge).
• Add a LangChain Agent for summarization + reasoning.
• Deploy via Render or Hugging Face Spaces.

💡 Features

• ✅ 100% Free Operation: Uses local Ollama + HuggingFace embeddings (no API costs)
• ✅ Conversational Memory: Streamlit UI remembers conversation history
• ✅ Flexible LLM Support: Switch between Ollama and OpenAI via config
• ✅ Source Citations: Shows which documents were used to generate answers
• ✅ Multiple Interfaces: Streamlit UI, REST API, or command line
• ✅ Easy Setup: Simple environment variable configuration

---

This project demonstrates practical RAG implementation skills for AI Engineer roles.