A practical multi-modal RAG system for learning advanced retrieval techniques with PDF documents.

Create a database named hybridRag with pgvector extension in postgres database

# First time (creates tables)
python main_ingest.py --init-db --pdf sample_data/your_file.pdf

# Subsequent documents
python main_ingest.py --pdf sample_data/another_file.pdf

Chunking options:

• --strategy recursive (default, recommended)
• --strategy semantic (paragraph-based)
• --strategy fixed (fixed size with overlap)

# Single question
python main_query.py --query "What is the total salary?"

# Interactive mode (recommended)
python main_query.py --interactive

├── config
│   └── settings.py
├── .env
├── .gitignore
├── ingestion
│   ├── chunker.py
│   ├── embedder.py
│   └── parser.py
├── main_ingest.py
├── main_query.py
├── readme.md
├── requirements.txt
├── storage
│   └── schema.sql
└── stores
    ├── graph_store.py
    ├── media_store.py
    └── postgres_store.py

PDF → Parse (text + images) → Chunk text → Generate embeddings → 
Store in Postgres + Build knowledge graph → Done!

Stores created:

• Text chunks with vector embeddings
• Images with GPT-4V descriptions
• Knowledge graph (document → chunks → media)
• Relationships between all entities

In pgAdmin, run:

-- See all documents
SELECT * FROM documents;

-- See text chunks
SELECT chunk_id, page_number, LEFT(text_content, 100) as preview 
FROM chunks;

-- See images with descriptions
SELECT media_id, page_number, file_path, LEFT(caption, 100) as description 
FROM media;

-- See relationships
SELECT * FROM relationships;

Edit config/settings.py for:

Models (cost-optimized):

EMBEDDING_MODEL = "text-embedding-3-small"
CHAT_MODEL = "gpt-4o-mini"
VISION_MODEL = "gpt-4o-mini"
PROCESS_IMAGES = True  # Set False to skip images