Spaces:
Sleeping
Sleeping
File size: 13,608 Bytes
f866820 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 | # RAG Document Assistant - Architecture
> **Version**: 2.0
> **Last Updated**: January 2026
> **Focus**: Zero-Storage Privacy Architecture
---
## System Overview
A privacy-first RAG (Retrieval-Augmented Generation) system where **no document text is ever stored on our servers**. Documents are processed client-side, and text is re-fetched from the user's cloud storage at query time.
### Key Characteristics
- **Zero-Storage**: Document text never persists on servers
- **Client-Side Processing**: Chunking happens in the browser
- **Query-Time Re-fetch**: Text retrieved from user's Dropbox for each search
- **User Control**: Disconnect cloud storage to revoke all access
---
## Privacy Architecture
```
INDEXING (one-time setup)
══════════════════════════════════════════════════════════════════
User's Browser Our Server
────────────── ──────────
1. Connect Dropbox (OAuth)
│
▼
2. Select files from Dropbox
│
▼
3. Files loaded in browser
(never sent to server)
│
▼
4. Text chunked locally ───────────────► 5. Generate embeddings
with position tracking (384-dim vectors)
│ │
▼ ▼
6. Original text 7. Store in Pinecone:
PURGED from memory - Embeddings (irreversible)
- File paths
- Chunk positions
- NO TEXT
══════════════════════════════════════════════════════════════════
QUERY TIME (every search)
══════════════════════════════════════════════════════════════════
User's Question Our Server
─────────────── ──────────
"What does the contract say?"
│
▼
─────────────────────────────────────► 1. Generate query embedding
│
▼
2. Search Pinecone
(find similar chunks)
│
▼
3. Get file paths + positions
│
▼
4. Re-fetch from USER'S Dropbox
using their access token
│
▼
5. Extract chunk text
using stored positions
│
▼
6. Send to LLM for answer
│
▼
Answer + Citations ◄───────────────── 7. Return response
(text never stored)
══════════════════════════════════════════════════════════════════
```
---
## What Gets Stored
| Data | Stored? | Where | Reversible? |
|------|---------|-------|-------------|
| Document files | No | User's Dropbox only | N/A |
| Document text | No | Never stored | N/A |
| Embeddings | Yes | Pinecone | No (one-way transform) |
| File paths | Yes | Pinecone metadata | N/A |
| Chunk positions | Yes | Pinecone metadata | N/A |
| User queries | No | Not logged | N/A |
---
## Technology Stack
### Frontend
- **Framework**: React 18 + Vite
- **Styling**: Tailwind CSS v4
- **Deployment**: Vercel
- **Key Features**:
- Client-side text chunking
- Dropbox OAuth integration
- Position tracking for chunks
### Backend
- **Framework**: FastAPI
- **Deployment**: HuggingFace Spaces (Docker)
- **Key Features**:
- Zero-storage embedding endpoint
- Query-time Dropbox re-fetch
- Multi-provider LLM cascade
### Vector Database
- **Service**: Pinecone Serverless
- **Index**: `rag-semantic-384`
- **Dimensions**: 384
- **Metric**: Cosine similarity
### Embeddings
- **Model**: `all-MiniLM-L6-v2` (sentence-transformers)
- **Dimensions**: 384
- **Processing**: Server-side (text discarded immediately)
### LLM Providers (Cascade)
1. **Gemini 2.5 Flash** (Primary)
2. **Groq** - llama-3.1-8b-instant (Fallback 1)
3. **OpenRouter** - Mistral 7B (Fallback 2)
---
## Component Architecture
```
┌─────────────────────────────────────────────────────────────────┐
│ FRONTEND (React) │
├─────────────────────────────────────────────────────────────────┤
│ │
│ ┌──────────────┐ ┌──────────────┐ ┌──────────────┐ │
│ │ Sidebar │ │ QueryPanel │ │ App.jsx │ │
│ │ │ │ │ │ │ │
│ │ - CloudConnect │ - Search UI │ - State mgmt │ │
│ │ - File select │ - Results │ - Token flow │ │
│ │ - Index button │ - Citations │ - Privacy UI │ │
│ └──────────────┘ └──────────────┘ └──────────────┘ │
│ │
│ ┌──────────────────────────────────────────────────┐ │
│ │ API Layer │ │
│ │ chunker.js │ dropbox.js │ client.js │ │
│ └──────────────────────────────────────────────────┘ │
│ │ │
└──────────────────────────────┼───────────────────────────────────┘
│ HTTPS
▼
┌─────────────────────────────────────────────────────────────────┐
│ BACKEND (FastAPI) │
├─────────────────────────────────────────────────────────────────┤
│ │
│ ┌────────────────────────────────────────────────────┐ │
│ │ API Routes │ │
│ │ │ │
│ │ POST /embed-chunks - Generate embeddings │ │
│ │ POST /query-secure - Zero-storage query │ │
│ │ POST /dropbox/token - OAuth token exchange │ │
│ │ POST /dropbox/folder - List folder contents │ │
│ │ POST /dropbox/file - Download file content │ │
│ │ DELETE /clear-index - Clear Pinecone index │ │
│ └────────────────────────────────────────────────────┘ │
│ │ │
└──────────────────────────────┼───────────────────────────────────┘
│
┌────────────────┼────────────────┐
│ │ │
▼ ▼ ▼
┌──────────┐ ┌──────────┐ ┌──────────┐
│ Pinecone │ │ Dropbox │ │ LLM │
│ (vectors)│ │ (files) │ │ Providers│
└──────────┘ └──────────┘ └──────────┘
```
---
## Data Flow: Indexing
```python
# 1. User selects files in browser
files = [
{id: "abc123", name: "contract.pdf", path: "/Documents/contract.pdf"}
]
# 2. Files fetched from Dropbox (via backend proxy)
content = await fetch("/api/dropbox/file", {path: file.path, access_token})
# 3. Text chunked CLIENT-SIDE with position tracking
chunks = chunkText(content, {chunkSize: 1000, overlap: 100})
# Result:
# {text: "...", startChar: 0, endChar: 1000}
# {text: "...", startChar: 900, endChar: 1900}
# 4. Chunks sent to backend for embedding
await fetch("/api/embed-chunks", {
chunks: [{
text: "...", // Used for embedding only
metadata: {
filename: "contract.pdf",
filePath: "/Documents/contract.pdf",
fileId: "abc123",
startChar: 0,
endChar: 1000
}
}]
})
# 5. Backend generates embeddings, stores in Pinecone
# TEXT IS IMMEDIATELY DISCARDED
pinecone.upsert({
id: "abc123::0",
values: [0.123, -0.456, ...], # 384-dim embedding
metadata: {
filename: "contract.pdf",
file_path: "/Documents/contract.pdf",
file_id: "abc123",
start_char: 0,
end_char: 1000
# NO TEXT STORED
}
})
```
---
## Data Flow: Query
```python
# 1. User submits query with access token
request = {
query: "What is the payment term?",
access_token: "user_dropbox_token"
}
# 2. Generate query embedding
query_embedding = sentence_transformer.encode(query)
# 3. Search Pinecone
results = pinecone.query(
vector=query_embedding,
top_k=3,
include_metadata=True
)
# Returns: file paths + positions (NO TEXT)
# 4. Re-fetch files from USER'S Dropbox
for file_path in unique_file_paths:
content = dropbox.download(file_path, access_token)
# 5. Extract chunks using stored positions
for chunk in chunks_from_file:
text = content[chunk.start_char:chunk.end_char]
# 6. Build prompt with re-fetched text
prompt = f"""
Context:
1. {chunk1_text}
2. {chunk2_text}
Question: {query}
"""
# 7. Call LLM
answer = llm.generate(prompt)
# 8. Return answer (text never stored)
return {answer, citations}
```
---
## Security & Privacy
### User Control
- **OAuth Scopes**: Read-only access to user-selected files
- **Token Storage**: Access token stored only in browser session
- **Revocation**: Disconnect Dropbox = immediate access revocation
### Server Security
- **No Persistent Storage**: Text never written to disk or database
- **Memory Only**: Text exists in memory only during processing
- **Immediate Purge**: Explicit deletion after embedding generation
### Data Protection
- **Embeddings**: One-way transformation, cannot reconstruct text
- **Positions**: Only useful with original file access
- **File Paths**: Dropbox paths, require valid access token
---
## Deployment
### Frontend (Vercel)
- Automatic deploys from GitHub
- Environment: `VITE_API_URL` pointing to backend
### Backend (HuggingFace Spaces)
- Docker-based deployment
- Environment variables for API keys:
- `PINECONE_API_KEY`
- `DROPBOX_APP_KEY`
- `DROPBOX_APP_SECRET`
- `GEMINI_API_KEY`
- `GROQ_API_KEY`
---
## References
- **Live Demo**: https://rag-document-assistant.vercel.app/
- **Backend API**: https://vn6295337-rag-document-assistant.hf.space/
- **GitHub**: https://github.com/vn6295337/RAG-document-assistant
|