Project 2 : Demo

Deep Search Assistant

A local multimodal RAG system for document and video understanding, enabling semantic search and grounded question-answering using optimized inference.

---

Models

• Embeddings: BAAI/bge-small-en-v1.5
• LLM: Gemma 7B (via Ollama / OpenVINO)
• Reranker: BAAI/bge-reranker-base
• Multimodal: CLIP ViT-B/32

---

Key Components

• LocalInferenceEngine – LLM inference
• KeywordRetriever – hybrid retrieval (BM25 + dense)
• RelevanceReranker – cross-encoder ranking
• DocumentPassageUnit – structured retrieval output

---

Quick Start

python3 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt

python cli.py ingest --dataset docvqa
python cli.py ask "your question"

---

Architecture

Modular pipeline:

• Core: embeddings, retrieval, LLM
• Processing: ingestion & preprocessing
• Utils: indexing, monitoring, benchmarking

---

Features

Multimodal RAG • Hybrid retrieval • Re-ranking • Hardware-aware inference • Edge-ready