RAG Projects — Hybrid Retrieval Pipeline

A multi-stage Retrieval-Augmented Generation system built with a focus on retrieval quality over basic vector search. Each version improves on the last — from naive retrieval to a fully agentic system with autonomous web fallback.

Versions

V1 — Naive RAG

Basic vector search implementation using FAISS embeddings.

• Dense vector retrieval only
• Simple similarity search
• Directories: rag_pdf/, rag_obsidian/

V2 — Advanced RAG

Significant upgrade to retrieval quality using hybrid search and reranking.

• Hybrid Search — combines dense vector retrieval (FAISS) with BM25 keyword matching
• Cross-Encoder Reranking — reranks retrieved chunks for improved context relevance
• Local LLM backend — fully offline using Ollama with Gemma, no external API dependency
• Directory: rag_multi/

V3 — Agentic RAG

ReAct agent with autonomous tool use and a strict local-before-web search protocol.

• Agent reasons through queries using THOUGHT → ACTION → OBSERVATION loops
• Always checks local documents first before escalating to web search
• Web fallback powered by DuckDuckGo — cross-encoder reranking applied to web results too
• ChromaDB for persistent local vector storage with Ollama embeddings
• Directory: agentic_rag_v2/

Tech Stack

• LLM: Ollama + Gemma (local, offline)
• Embeddings: Ollama (Gemma)
• Vector Store: ChromaDB (V3), FAISS (V1, V2)
• Keyword Search: BM25 (rank-bm25) — V2
• Reranking: Cross-encoder (ms-marco-MiniLM-L-6-v2)
• Web Search: DuckDuckGo (ddgs)
• Framework: LangChain
• Language: Python

Why Hybrid Search?

Standard RAG uses only dense vector search — great for semantic similarity but misses exact keyword matches. BM25 catches what vector search misses. Combining both and reranking with a cross-encoder gives significantly better retrieval precision than either method alone.

Why Agentic RAG?

Static RAG pipelines always retrieve — even when the local corpus has nothing relevant. The agentic approach lets the model decide: check local documents first, and only fall back to live web search when local results are insufficient. This keeps responses grounded in your documents while staying useful for out-of-domain queries.

Setup

git clone https://github.com/adithyanum/RAG-Projects
cd RAG-Projects
pip install -r requirements.txt

Make sure Ollama is running locally:

ollama pull gemma
ollama serve

Results

Hybrid search + cross-encoder reranking consistently retrieves more relevant chunks compared to naive vector search, especially for domain-specific queries where exact terminology matters. The agentic layer adds adaptability — handling both in-corpus and open-domain questions in a single pipeline.

---

Built by Adithyan U M — github.com/adithyanum