RAG Search Engine

A full RAG pipeline with keyword, semantic, hybrid, LLM-aided, multimodal search, and augmented generation capabilities.

Setup

Prerequisites

• Bun installed
• Google Gemini API key (for LLM features)

Installation

1. Install dependencies:

bun install

2. Set up environment variables:

cp .env.example .env
# Edit .env and add your GEMINI_API_KEY

Usage

Important: Due to native dependencies (ONNX Runtime), the application must be run directly with Bun:

bun start <command> [options]
# or
bun run cli/index.ts <command> [options]

Commands

build — Build search index

bun start build -t keyword
bun start build -t vector
bun start build -t vector --chunked

Options:

• -t, --type <type> — Index type: keyword, vector
• -c, --chunked — Build chunked vector index (for semantic search)

---

keyword-search — Keyword-based search

bun start keyword-search -t basic "family movie about bears"
bun start keyword-search -t tf-idf "family movie about bears"
bun start keyword-search -t bm25 "family movie about bears"

Options:

• -t, --type <type> — Search type: basic, tf-idf, bm25
• -l, --limit <number> — Number of results (default: 5)

---

semantic-search — Embedding-based semantic search

bun start semantic-search "family movie about bears"
bun start semantic-search --chunked "family movie about bears"

Options:

• -l, --limit <number> — Number of results (default: 5)
• -c, --chunked — Use chunked vector index

---

hybrid-search — Hybrid search (BM25 + semantic)

# Weighted combination
bun start hybrid-search -t weighted -a 0.5 "family movie about bears"

# Reciprocal Rank Fusion
bun start hybrid-search -t ranked "family movie about bears"

Options:

• -t, --type <type> — Search type: weighted, ranked
• -a, --alpha <number> — Alpha weight for BM25 in weighted mode (default: 0.5)
• -k, --k <number> — K constant for RRF ranked search (default: 60)
• -l, --limit <number> — Number of results (default: 5)

---

llm-search — LLM-aided hybrid search

# Query enhancement
bun start llm-search -e spell "famly moovie about bares"
bun start llm-search -e rewrite "something heartwarming with animals"
bun start llm-search -e expand "family movie about bears"

# Re-ranking
bun start llm-search -r llm "family movie about bears"
bun start llm-search -r cross-encoder "family movie about bears"

# Judge results
bun start llm-search --judge "family movie about bears"

# Combined
bun start llm-search -e expand -r cross-encoder "family movie about bears"

Options:

• -e, --enhanced <type> — Query enhancement: spell, rewrite, expand
• -r, --reRank <type> — Re-ranking method: llm, cross-encoder
• -j, --judge — Judge and score results using LLM (scores out of 3)
• -k, --k <number> — K constant for RRF (default: 60)
• -l, --limit <number> — Number of results (default: 5)

---

rag — Retrieval-Augmented Generation

bun start rag -t answer "What is a good family movie about bears?"
bun start rag -t summary "What is a good family movie about bears?"
bun start rag -t citation "What is a good family movie about bears?"
bun start rag -t detailed_answer "What is a good family movie about bears?"

# With an image (for multimodal RAG)
bun start rag -t answer -i dataset/movie_poster.png "What movie is this?"

Options:

• -t, --type <type> — RAG type: answer, summary, citation, detailed_answer
• -i, --image <image> — Path to image (optional, for multimodal RAG)
• -l, --limit <number> — Number of results to retrieve (default: 5)

---

multimodal-search — Search by image

bun start multimodal-search dataset/movie_poster.png

Options:

• -l, --limit <number> — Number of results (default: 5)

---

evaluate — Evaluate hybrid search with golden dataset

bun start evaluate
bun start evaluate -l 10

Options:

• -l, --limit <number> — Number of results per query (default: 5)

---

get-index — Inspect index values for a term

bun start get-index -t tf -d 1 "bears"
bun start get-index -t idf "bears"
bun start get-index -t tf-idf -d 1 "bears"
bun start get-index -t bm25-idf "bears"
bun start get-index -t bm25-tf -d 1 "bears"

Options:

• -t, --type <type> — Index type: tf, idf, tf-idf, bm25-idf, bm25-tf
• -d, --docId <docId> — Document ID (required for tf, tf-idf, bm25-tf)
• -k, --k1 <k1> — BM25 k1 parameter
• -b, --b <b> — BM25 b parameter

Why Not Compile?

The bun build --compile option cannot be used because:

• The cross-encoder re-ranker depends on @huggingface/transformers
• This library requires native ONNX Runtime libraries (.dylib files)
• Native libraries cannot be bundled into compiled binaries
• For compiled binaries, use the LLM re-ranker instead (-r llm)