Five self-contained Python examples demonstrating different Retrieval-Augmented Generation (RAG) architectures. Every script runs fully offline — no API keys required. All LLM responses and embeddings are mocked or generated with deterministic random vectors so the focus stays on the pipeline logic.

pip install -r requirements.txt

Combines dense vector search (fake CLIP-style embeddings + cosine similarity store) with sparse keyword search (hand-rolled BM25). Both retrievers run in parallel on the same query, then results are merged via Reciprocal Rank Fusion (RRF). The fused top-K chunks are passed to the LLM.

Query → Dense Retrieval ─┐
      → BM25 Retrieval  ─┴→ RRF → Top-K chunks → LLM → Answer

python 01_hybrid_rag.py

Builds a Knowledge Graph (using networkx) from the document corpus. Named entities (Person, Company, Technology, Location) become nodes; typed relationships (founded, ceo_of, invested_in, …) become directed edges. At query time, query entities seed a multi-hop subgraph traversal; the retrieved subgraph is summarised into a community description fed to the LLM.

Query → Entity Extraction → KG Subgraph Retrieval → Community Summary → LLM → Answer

python 02_graph_rag.py

A Planner Agent analyses the query and selects an ordered sequence of tools to invoke. Three tools are available: VectorSearchTool, WebSearchTool, and SQLDatabaseTool. The agent loop executes each tool, tracks accumulated confidence, and short-circuits when confidence is high enough. A Reasoner Agent then synthesises all tool results into a final answer.

Query → Planner → [Tool Loop: Vector / Web / SQL] → Reasoner → Answer

python 03_agentic_rag.py

After initial retrieval an Evaluator/Grader scores relevance via keyword overlap and emits one of three verdicts:

• CORRECT → pass chunks directly to the LLM.
• AMBIGUOUS → rewrite the query and re-retrieve, then go to LLM.
• INCORRECT → fall back to a simulated web search, then go to LLM.

Three demo queries exercise all three branches in one run.

Query → Retriever → Evaluator ──CORRECT──→ LLM → Answer
                             ├─AMBIGUOUS─→ Rewriter → Re-retrieve → LLM → Answer
                             └─INCORRECT─→ Web Search → LLM → Answer

python 04_corrective_rag.py

Indexes three modality types — text chunks, images/charts (represented by captions and tags), and tables — into a single unified vector index using a mock CLIP-style embedder that maps every modality to the same 64-dimensional space. At query time, modality preferences are detected ("chart", "table", …) and a post-retrieval modality boost re-ranks results before the mock Multimodal LLM generates an answer referencing text, images, and tables together.

Text Chunks ─┐
Images      ─┼→ Multimodal Embedder → Unified Index → Retrieval → Boost → MLLM → Answer
Tables      ─┘

python 05_multimodal_rag.py

BM25 is implemented from scratch in 01_hybrid_rag.py — no extra package needed.