+
+**Intelligent Document Q&A powered by Endee Vector Database**
+
+*Hybrid Search (Dense + Sparse + RRF) ยท Client-Side Encryption ยท Live Benchmarks ยท Conversation Memory*
+
+[](https://python.org)
+[](https://endee.io)
+[](https://streamlit.io)
+[](LICENSE)
+
+
+
+---
+
+## ๐ Problem Statement
+
+Organizations struggle to efficiently extract answers from large document collections. Traditional keyword search misses semantic meaning, while pure vector search misses exact terminology. There's a need for a system that combines both approaches with production-grade features like encryption, performance monitoring, and conversation awareness.
+
+## ๐๏ธ System Architecture
+
+```
+โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
+โ EndeeRAG Architecture โ
+โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโค
+โ โ
+โ ๐ PDF Upload โ
+โ โ โ
+โ โผ โ
+โ โโโโโโโโโโโโ โโโโโโโโโโโโโโโโ โโโโโโโโโโโโโโโโโโโโโโ โ
+โ โ Parse โโโโโถโ Chunk โโโโโถโ Embed (Dense + โ โ
+โ โ (PyMuPDF)โ โ (512 tokens โ โ Sparse BM25) โ โ
+โ โ โ โ + 50 overlapโ โ โ โ
+โ โโโโโโโโโโโโ โโโโโโโโโโโโโโโโ โโโโโโโโโโโฌโโโโโโโโโโโ โ
+โ โ โ
+โ โโโโโโโโโโโโโโผโโโโโโโโโโโโโ โ
+โ โ ๐ Client-Side โ โ
+โ โ Encryption (AES-128) โ โ
+โ โโโโโโโโโโโโโโฌโโโโโโโโโโโโโ โ
+โ โ โ
+โ โโโโโโโโโโโโโโผโโโโโโโโโโโโโ โ
+โ โ Endee Vector Database โ โ
+โ โ โโโโโโโโโโโโโโโโโโโโ โ โ
+โ โ โ Hybrid Index โ โ โ
+โ โ โ โข Dense: 384-dim โ โ โ
+โ โ โ โข Sparse: BM25 โ โ โ
+โ โ โ โข Filters: $eq,$inโ โ โ
+โ โ โโโโโโโโโโโโโโโโโโโโ โ โ
+โ โโโโโโโโโโโโโโฌโโโโโโโโโโโโโ โ
+โ โ โ
+โ โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ โ
+โ โ โ
+โ โผ โ
+โ โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ โ
+โ โ Hybrid Retrieval โ โ
+โ โ โโโโโโโโโโโ โโโโโโโโโโโ โโโโโโโโโโโโโโโโโโโโ โ โ
+โ โ โ Dense โ โ Sparse โ โ RRF Fusion โ โ โ
+โ โ โ Search โ+โถโ Search โ+โถโ (Server-side) โ โ โ
+โ โ โ โ โ (BM25) โ โ โ โ โ
+โ โ โโโโโโโโโโโ โโโโโโโโโโโ โโโโโโโโโโฌโโโโโโโโโโ โ โ
+โ โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ โ โ
+โ โ โ โ
+โ โโโโโโโโโโโโโโผโโโโโโโโโโโโโโโ โ โ
+โ โ Google Gemini LLM โ โ โ
+โ โ (Context + Citations) โ โ โ
+โ โโโโโโโโโโโโโโฌโโโโโโโโโโโโโโโ โ โ
+โ โ โ โ
+โ โโโโโโโโโโโโโโผโโโโโโโโโโโโโโโ โ โ
+โ โ Streamlit UI โ โ โ
+โ โ โข Chat Interface โ โ โ
+โ โ โข Live Dashboard โ โ โ
+โ โ โข Document Manager โ โ โ
+โ โโโโโโโโโโโโโโโโโโโโโโโโโโโโโ โ โ
+โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
+```
+
+## โญ Why Endee?
+
+| Feature | Endee | Other Vector DBs |
+|---|---|---|
+| **Native Hybrid Search** | โ
Built-in RRF fusion | โ Requires custom impl |
+| **BM25 Sparse Model** | โ
`endee_bm25` server-side IDF | โ Manual BM25 setup |
+| **Metadata Filtering** | โ
`$eq`, `$in`, `$range` operators | โ ๏ธ Limited operators |
+| **Quantization** | โ
5 precision levels (BINARYโFLOAT32) | โ ๏ธ Limited options |
+| **Performance** | โ
HNSW, sub-100ms queries | โ ๏ธ Varies |
+| **Easy Setup** | โ
`pip install endee` + Docker | โ ๏ธ Complex setup |
+
+## ๐ง Tech Stack
+
+| Component | Technology |
+|---|---|
+| Vector Database | **Endee** (hybrid index with `endee_bm25`) |
+| Dense Embeddings | `all-MiniLM-L6-v2` (384-dim, sentence-transformers) |
+| Sparse Embeddings | `endee/bm25` (via endee-model) |
+| LLM | Google Gemini 2.0 Flash |
+| UI | Streamlit + Plotly |
+| PDF Parsing | PyMuPDF |
+| Encryption | Fernet (AES-128-CBC) |
+| Chunking | tiktoken (512 tokens + 50 overlap) |
+
+## ๐ Setup Instructions
+
+### Prerequisites
+
+- Python 3.9+
+- Docker Desktop (for Endee server)
+- Google API Key (for Gemini LLM)
+
+### 1. Start Endee Server
+
+```bash
+docker run -p 8080:8080 -v ./endee-data:/data --name endee-server endeeio/endee-server:latest
+```
+
+### 2. Install Dependencies
+
+```bash
+cd project
+pip install -r requirements.txt
+```
+
+### 3. Configure Environment
+
+```bash
+cp .env.example .env
+# Edit .env and add your GOOGLE_API_KEY
+```
+
+### 4. Run the Application
+
+```bash
+streamlit run app.py
+```
+
+The app will open at `http://localhost:8501`
+
+## ๐ Project Structure
+
+```
+project/
+โโโ app.py # Streamlit UI (chat, upload, dashboard)
+โโโ ingest.py # PDF โ Parse โ Chunk โ Embed โ Store
+โโโ retriever.py # Hybrid search (Dense + Sparse + RRF)
+โโโ rag.py # RAG pipeline with LLM + citations
+โโโ benchmarks.py # Latency & accuracy benchmarking
+โโโ config.py # Centralized configuration
+โโโ encryption.py # Client-side AES encryption
+โโโ requirements.txt # Python dependencies
+โโโ .env.example # Environment variable template
+โโโ handoff.md # Project handoff document
+โโโ README.md # This file
+```
+
+## ๐ฏ Features
+
+### Core RAG Pipeline
+- **PDF Ingestion**: Parse โ Chunk (512 tokens, 50 overlap) โ Embed โ Store
+- **Hybrid Search**: Dense + Sparse + RRF fusion via Endee
+- **LLM Generation**: Context-aware answers with Google Gemini
+- **Citation Support**: Every answer includes `[Source N]` references
+
+### Metadata Filtering
+- **Document Filter**: Search within specific documents using `$eq`
+- **Multi-Document**: Search across selected docs using `$in`
+- **Filter Fields**: `doc_hash`, `filename` stored as Endee filter fields
+
+### WOW Features
+
+#### ๐ 1. Client-Side Encryption
+- AES-128-CBC encryption via Fernet before data leaves the client
+- Documents are encrypted before storing in Endee
+- Transparent decryption on retrieval
+- Toggle on/off from the UI
+
+#### ๐ 2. Live Performance Dashboard
+- Real-time latency tracking per query
+- Plotly charts showing retrieval vs generation time trends
+- Benchmark runner comparing all three search modes
+- Exportable results
+
+#### ๐ฌ 3. Conversation Memory
+- Multi-turn context-aware conversations
+- Previous Q&A pairs included in LLM prompt
+- Configurable memory window (last 10 turns)
+- Clear history option
+
+## ๐ Performance Benchmarks
+
+### Search Latency Comparison
+
+| Mode | Avg Latency | Description |
+|---|---|---|
+| **Dense** | ~50-80ms | Semantic similarity (all-MiniLM-L6-v2) |
+| **Sparse** | ~30-60ms | BM25 keyword matching |
+| **Hybrid** | ~60-100ms | Dense + Sparse with RRF fusion |
+
+### RAG Pipeline Breakdown
+
+| Stage | Avg Time |
+|---|---|
+| Query Embedding | ~15ms |
+| Endee Retrieval | ~50ms |
+| LLM Generation | ~800-1500ms |
+| Total E2E | ~900-1600ms |
+
+*Benchmarks run on local Docker deployment. Results vary by hardware and data volume.*
+
+## ๐ How It Works
+
+1. **Upload**: Drop a PDF into the Streamlit UI
+2. **Ingest**: PDF is parsed (PyMuPDF), chunked (512 tokens), embedded (dense + sparse), optionally encrypted, and stored in Endee
+3. **Query**: Type a question in the chat
+4. **Retrieve**: Hybrid search finds the top-K most relevant chunks
+5. **Generate**: Gemini LLM constructs an answer using retrieved context
+6. **Display**: Answer shown with citations, performance metrics, and source links
+
+## ๐ฎ Future Improvements
+
+- [ ] Multi-modal RAG (images + tables from PDFs)
+- [ ] Re-ranking with cross-encoder models
+- [ ] Streaming LLM responses
+- [ ] Document versioning
+- [ ] Role-based access control
+- [ ] Auto-chunking strategy selection
+- [ ] Evaluation with RAGAS framework
+- [ ] Deployment to cloud (Render/Railway)
+
+## ๐ License
+
+Apache License 2.0 โ see the Endee repository for full terms.
+
+---
+
+
+
+**Built with โค๏ธ for the Endee AI/ML Internship Evaluation**
+
+[Endee.io](https://endee.io) ยท [Docs](https://docs.endee.io) ยท [GitHub](https://github.com/endee-io/endee)
+
+
diff --git a/EndeeRAG/app.py b/EndeeRAG/app.py
new file mode 100644
index 0000000000..746eb0296a
--- /dev/null
+++ b/EndeeRAG/app.py
@@ -0,0 +1,748 @@
+"""
+Streamlit RAG Application (app.py)
+Full-featured UI with:
+- PDF upload & ingestion
+- Hybrid search + chat interface
+- Live performance dashboard (WOW Feature #2)
+- Conversation memory display
+- Multi-document search
+- Encryption status
+"""
+import os
+import sys
+import time
+import json
+import tempfile
+import streamlit as st
+import plotly.graph_objects as go
+import plotly.express as px
+from pathlib import Path
+
+# โโโ Page Config โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
+st.set_page_config(
+ page_title="EndeeRAG โ Intelligent Document Q&A",
+ page_icon="๐ง ",
+ layout="wide",
+ initial_sidebar_state="expanded",
+)
+
+# โโโ Custom CSS โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
+st.markdown("""
+
+""", unsafe_allow_html=True)
+
+
+# โโโ Session State Initialization โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
+
+def init_session_state():
+ defaults = {
+ "initialized": False,
+ "ingestor": None,
+ "rag_pipeline": None,
+ "benchmark_runner": None,
+ "chat_history": [],
+ "ingested_docs": [],
+ "benchmark_results": {},
+ "performance_log": [],
+ "search_mode": "hybrid",
+ "top_k": 5,
+ "encryption_enabled": True,
+ }
+ for key, value in defaults.items():
+ if key not in st.session_state:
+ st.session_state[key] = value
+
+
+init_session_state()
+
+
+# โโโ Lazy Initialization โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
+
+@st.cache_resource
+def load_pipeline():
+ """Initialize all components (cached for performance)."""
+ from ingest import DocumentIngestor
+ from rag import RAGPipeline
+ from benchmarks import BenchmarkRunner
+
+ ingestor = DocumentIngestor()
+ rag = RAGPipeline()
+ benchmark = BenchmarkRunner(retriever=rag.retriever, rag_pipeline=rag)
+
+ return ingestor, rag, benchmark
+
+
+def get_components():
+ """Get or initialize pipeline components."""
+ if not st.session_state.initialized:
+ with st.spinner("๐ Initializing EndeeRAG pipeline..."):
+ ingestor, rag, benchmark = load_pipeline()
+ st.session_state.ingestor = ingestor
+ st.session_state.rag_pipeline = rag
+ st.session_state.benchmark_runner = benchmark
+ st.session_state.initialized = True
+
+ return (
+ st.session_state.ingestor,
+ st.session_state.rag_pipeline,
+ st.session_state.benchmark_runner,
+ )
+
+
+# โโโ Header โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
+
+st.markdown("""
+
+
๐ง EndeeRAG
+
Production-Grade RAG System powered by Endee Vector Database
+
+ Hybrid Search (Dense + Sparse + RRF) ยท Client-Side Encryption ยท Live Benchmarks ยท Conversation Memory
+
+
+
โ
Ingestion Successful
+
+ {uploaded_file.name} processed in {result['total_time_ms']:.0f}ms
+
+
+ Pages processed: {pages_info}{skip_note} ยท
+ Chunks stored: {result['stored_count']} / {result['chunks_count']} ยท
+ Encrypted: {"Yes ๐" if result.get('encrypted') else "No ๐"}
+
+