Grover + LLM Integration for Enhanced RAG Systems

Overview

This project implements a RAG system enhanced with Grover's algorithm for intelligent context selection. Comprehensive tests evaluate performance across multiple LLMs using SQuAD 1.1 benchmark data.

---

Key Tests & Findings

🔍 Test 1: Answer Quality Evaluation

Objective: Assess impact of:

• Grover vs. classic context selection
• LLM models (llama-3-8b, mixtral-8x7b, phi-3.5)
• Context variants (no_context, top1, top3)

Parameters:

• 63 SQuAD questions
• Metrics: Word overlap and cosine similarity

Key Insights:

• 3-context setups outperform: llama-3-8b achieved highest cosine similarity (0.80)
• Grover ≈ Classic: Quality differences <0.5%
• No-context fails: All models showed significant quality drop

---

⚡ Test 2: Full System Performance

Parameters:

• 56 SQuAD questions
• End-to-end latency measurement

Results:

⏱️ Timing Analysis

Component	Time
Context retrieval (top-10)	0.297 s
Grover selection (top-3)	0.030 s
Answer generation:
mixtral-8x7b	1.18 s 🚀
llama-3-8b	2.56 s
phi-3.5	2.65 s

🔍 Context Consistency

• 99.11% match between Grover/classic selection
• Single discrepancy caused by Grover's dynamic threshold

📊 Quality Validation

• Confirmed Test 1 trends: top3 > top1 > no_context
• llama-3-8b dominated quality metrics (cosine: 0.80)


---

Key Conclusions

1. Grover is efficient:

   • Adds only 30ms latency vs classic selection
   • Maintains context selection quality

2. Context is critical:

   • 3 contexts boost accuracy by 38% vs no context

3. Model tradeoffs:

   • mixtral-8x7b: Fastest inference (1.18s)
   • llama-3-8b: Highest accuracy with contexts

---

GUI Demonstration

Interactive interface features:

• Multi-model comparisons (llama, mixtral, phi-3.5)
• Context inspection (Grover vs classic)
• Collapsible answers

Screen	Preview
Home
Answers
Top Contexts