How Sound Suite Works
From raw court PDFs to AI-powered intelligence in four automated steps.
Monitor Your Case Folders
Point Sound Suite at your case directories. The built-in file watcher monitors for new PDFs in real-time. When a new document arrives, it's automatically queued for processing.
Technical Details
- • Configurable watch paths via
WATCH_PATHS - • Recursive directory monitoring with chokidar
- • SHA-256 hash deduplication prevents reprocessing
- • Creates Document records with QUEUED status
Automated Processing
Each document flows through an intelligent processing pipeline: text extraction, page-by-page density analysis, OCR for scanned content, and exhibit image extraction. All processing happens locally.
Technical Details
- • Job queue with configurable concurrency (p-queue)
- • Automatic retry with exponential backoff
- • Status: QUEUED → PROCESSING → INDEXED
- • Job logging for full audit trail
AI Indexing
Processed text is split into semantic chunks and converted to vector embeddings — numerical representations that capture meaning, not just keywords. Vectors are stored in LanceDB for millisecond-fast similarity search.
Embedding Providers
Local
Transformers — fully offline
OpenAI
High-quality via API
Anthropic
Claude-powered
Ollama
Local LLM-based
Vector Space Visualization
Based on the Service Agreement (Exhibit A), the defendant has the following obligations:
- Monthly reporting due by the 15th (Section 4.2)
- Maintain insurance coverage of $1M (Section 7.1)
- Non-compete for 24 months post-termination (Section 9.3)
Sources: exhibit-a.pdf (p.3, 7, 12)
Search & Analyze
Now your documents are alive. Use any of 14 MCP analysis tools through your preferred AI assistant — ask questions in natural language, find contradictions, extract timelines, screen for privileged content, and more.
Compatible AI Assistants
- • Claude (Anthropic)
- • ChatGPT (OpenAI)
- • Any MCP-compatible AI client
Draft Documents with AI Assistance
Sound Suite doesn't just analyze your case files — it helps you write. The built-in draft editor connects your indexed documents to the writing process.
Auto-Suggest
As you type, Sound Suite searches your indexed case documents and suggests relevant text — citations, contract clauses, deposition excerpts, dates with sources.
You type: "The contract requires..."
...monthly reporting by the 15th (Section 4.2, Exhibit A p.3) and maintenance of $1M insurance coverage (Section 7.1)...
Tab to accept · Esc to dismiss
Search Speed Presets
Balance speed vs. depth. Ultra searches more documents with larger context windows.
AI Chat Panel
A side panel where you ask questions about your case while writing. The AI searches your indexed documents and returns sourced answers you can insert directly.
"What did the plaintiff claim about the delivery date?"
Plaintiff states delivery was promised by March 1 (Complaint, para. 14) but actually occurred April 12 (Exhibit B, p.2). The 42-day delay is referenced in three filings.
Sources: complaint.pdf (p.5), exhibit-b.pdf (p.2), reply-brief.pdf (p.8)
Full Legal Editor
Not a toy editor — a production-ready tool built for legal document workflows.
- Import/export Word (.docx)
- PDF export with headers/footers
- Track changes with accept/reject
- Version history with compare
- Format Painter & heading nav
- Font sizes in points (legal standard)
- Document minimap
Architecture
Everything runs on your machine. Here's how the pieces fit together.
Your Machine
Case Folders
PDFs on disk
Job Queue
p-queue + retry
Ingestion Pipeline
OCR + extraction + chunking
SQLite
Prisma ORM
LanceDB
Vector embeddings
MCP Server
14 analysis tools · port 3001
Next.js Dashboard
port 3000
AI Clients
Claude, ChatGPT, etc.
Sidecar: Local AI via Docker
Sound Suite includes a GPU sidecar — a self-contained process that orchestrates Docker containers running local AI models. It handles embeddings, OCR, reranking, and completions without any cloud API keys.
The sidecar auto-detects your GPU, pulls the right Docker images, and manages VRAM allocation across containers. It communicates with the main server via WebSocket (with HTTP polling fallback), so it works across network boundaries and NAT.
How It Works
- 1 Download & launch — The sidecar ships as a tarball or runs via Docker itself. On startup it detects Node.js 18+ and GPU availability.
- 2 Auto-provision — Pulls Ollama and vLLM images, creates containers for each role (embedding, completion, OCR, reranker), checks VRAM before starting.
- 3 Connect to server — Opens a WebSocket to the main Sound Suite instance. The server dispatches embedding, OCR, and reranking jobs to the sidecar.
- 4 Mode switching — Swaps between "indexing" mode (embedding + OCR containers) and "searching" mode (embedding + reranker + completion) to fit within available VRAM.
The sidecar runs on port 8098 with its own dashboard for monitoring container health and GPU utilization.
Docker Containers
Embedding
ollama · qwen3-embedding:0.6b
Completion
ollama · qwen3.5:9b
OCR
ollama · olmocr2:7b-q8
Reranker
vLLM · Qwen3-Reranker-8B
Sidecar Communication
Main Server
:3000
GPU Sidecar
:8098
Primary: WebSocket (:3002) · Fallback: HTTP polling · Heartbeat every 5s
VRAM-Aware Mode Switching
Indexing Mode
- Embedding container
- OCR container
Searching Mode
- Embedding container
- Reranker container
- Completion container
Containers are stopped and started as needed to stay within available VRAM.
What Makes It Different
| Traditional Cloud Solutions | Sound Suite |
|---|---|
| Documents uploaded to third-party servers | All processing on your local machine |
| Monthly subscription fees | Source-available, free for pro se use |
| Vendor lock-in | Your data in standard formats |
| Internet required | Works fully offline |
| Limited AI analysis tools | 14 specialized MCP tools |
| Generic document processing | Built for legal documents |
| Complex enterprise setup | Install and run in minutes |
Ready to See It In Action?
Download Sound Suite and index your first case in under 5 minutes.