🧠 AI Document QA System

A production‑style Document Intelligence and Question‑Answering system that ingests files, builds embeddings, performs hybrid retrieval, and answers questions using LLM reasoning pipelines.

---

📚 Table of Contents

• Overview
• Architecture
• Features
• Project Structure
• AI Pipeline
• Notebooks (Start Here)
• Installation
• Running the Project
• Configuration
• Future Improvements

---

🚀 Overview

This project implements a modular AI retrieval architecture designed to bridge research workflows and production systems.

Key capabilities:

• Document ingestion (PDF, DOCX, text, images)
• Intelligent chunking pipeline
• Embedding generation
• Vector search with Chroma
• Keyword search with BM25 (SQLite FTS5)
• LangGraph QA reasoning pipeline
• Config‑driven prompts and models

The system follows clean architecture + adapter pattern to make AI components swappable and testable.

---

🧩 Architecture

                        +-------------------+
                        |       Client      |
                        |  API / Notebook   |
                        +---------+---------+
                                  |
                                  v
                        +-------------------+
                        |    Backend Layer  |
                        |  Request Control  |
                        +---------+---------+
                                  |
                                  v
                      +-----------------------+
                      |      AI Core Layer    |
                      |-----------------------|
                      | Document Parsers      |
                      | Chunking Pipeline     |
                      | Embedding Service     |
                      | Vector Store          |
                      | Keyword Search (BM25) |
                      | QA LangGraph Workflow |
                      +-----------+-----------+
                                  |
                                  v
                           +-------------+
                           |     LLM     |
                           +-------------+

---

✨ Features

Intelligent Document Processing

Supports multiple document types via adapters.

• PDF
• DOCX
• Plain text
• Images

Hybrid Retrieval

Combines:

• Dense vector search
• Sparse keyword search

Result → better recall and accuracy.

LangGraph QA Workflows

Advanced reasoning pipeline capable of:

• Context retrieval
• Multi‑step reasoning
• Structured responses
• Tool orchestration

Modular Adapter Architecture

Adapters isolate external systems:

LLM
Vector Database
Embedding Model
File Parsers
Search Engines

Making the system easy to extend.

---

📂 Project Structure

.
├── notebooks/                 # 🔥 Start here
│
├── configs/
│   ├── configs.yaml
│   └── prompts.yml
│
├── core/
│   ├── adapters/
│   │   ├── chunking/
│   │   ├── embedding/
│   │   ├── keyword_search/
│   │   ├── llm/
│   │   └── vectorstore/
│   │
│   └── workflows/
│       └── langgraph/
│
└── docker/

The architecture follows a ports & adapters design to separate business logic from infrastructure.

---

🧠 AI Pipeline

1️⃣ Document Loading

Files are parsed and normalized using parser adapters.

2️⃣ Chunking

Documents are split into semantic segments using configurable strategies.

3️⃣ Embedding Generation

Each chunk is converted into vectors via HuggingFace embedding models.

4️⃣ Storage

Vectors are stored in ChromaDB.

5️⃣ Retrieval

Two systems run in parallel:

• Vector similarity search
• BM25 keyword search

6️⃣ Reasoning

LangGraph orchestrates:

• Retrieval
• Context filtering
• LLM generation
• Response formatting

---

📓 Notebooks (Start Here)

The entry point of the project is inside the notebooks directory.

Run the notebooks in order:

notebooks/
│
├── 01_Setup_Notebook.ipynb
├── 02_AI_Technical_Notebook.ipynb
├── 03_File_Chunk_Embedding_Search.ipynb
└── 04_QA_Pipeline_LangGraph.ipynb

These notebooks demonstrate the entire AI pipeline step‑by‑step.

They serve as:

• tutorials
• experiments
• debugging tools
• documentation

---

⚙️ Installation

git clone <repo>

cd project

pip install -r requirements.txt

---

▶️ Running

Run notebooks

cd notebooks
jupyter notebook

Start services with Docker

docker compose up --build

---

🔧 Configuration

Main configuration files:

configs/configs.yaml
configs/prompts.yml

You can configure:

• LLM provider
• Embedding models
• Retrieval parameters
• Prompt templates
• Chunking strategies

---

🧪 Design Principles

• Research friendly
• Production oriented
• Highly modular
• Replaceable AI components
• Config‑driven experimentation

---

🔮 Future Improvements

• Evaluation pipelines
• Observability
• Hybrid ranking
• Streaming responses
• UI interface
• Multi‑document reasoning

---

🤝 Contributing

1. Start experimentation in notebooks
2. Move stable logic to core/
3. Add adapters if integrating new systems
4. Expose via backend services

---

⭐ If this project helped you

Consider giving it a star.