flockparser 1.0.4

Distributed document RAG system with intelligent GPU/CPU orchestration

FlockParse - Document RAG Intelligence with Distributed Processing

Distributed document RAG system that turns mismatched hardware into a coordinated inference cluster. Auto-discovers Ollama nodes, intelligently routes workloads across heterogeneous GPUs/CPUs, and achieves 60x+ speedups through adaptive load balancing. Privacy-first with local/network/cloud interfaces.

What makes this different: Real distributed systems engineering—not just API wrappers. Handles heterogeneous hardware (RTX A4000 + GTX 1050Ti + CPU laptops working together), network failures, and privacy requirements that rule out cloud APIs.

---

⚠️ Important: Current Maturity

Status: Beta (v1.0.0) - Early adopters welcome, but read this first!

What works well:

• ✅ Core distributed processing across heterogeneous nodes
• ✅ GPU detection and VRAM-aware routing
• ✅ Basic PDF extraction and OCR fallback
• ✅ Privacy-first local processing (CLI/Web UI modes)

Known limitations:

• ⚠️ Limited battle testing - Tested by ~2 developers, not yet proven at scale
• ⚠️ Security gaps - See SECURITY.md for current limitations
• ⚠️ Edge cases - Some PDF types may fail (encrypted, complex layouts)
• ⚠️ Test coverage - ~40% coverage, integration tests incomplete

Read before using: KNOWN_ISSUES.md documents all limitations, edge cases, and roadmap honestly.

Recommended for:

• 🎓 Learning distributed systems
• 🔬 Research and experimentation
• 🏠 Personal projects with non-critical data
• 🛠️ Contributors who want to help mature the project

Not yet recommended for:

• ❌ Mission-critical production workloads
• ❌ Regulated industries (healthcare, finance) without additional hardening
• ❌ Large-scale deployments (>50 concurrent users)

Help us improve: Report issues, contribute fixes, share feedback!

---

📹 Demo Video (76 seconds)

Watch FlockParser in action: 372 seconds → 6 seconds (61.7x speedup) through automatic GPU routing.

What you'll see:

• Single CPU node (372.76s) → Parallel processing (159.79s) → GPU routing (6.04s)
• Real-time document processing with visible timing on screen
• Distributed chat functionality and MCP integration with Claude Desktop
• No editing tricks - all timing shown in real-time

📊 Quick Performance Reference

Workload	Hardware	Time	Speedup	Notes
5 AI papers (~350 pages)	1× RTX A4000 (16GB)	21.3s	17.5×	Real arXiv showcase
12-page PDF (demo video)	1× RTX A4000 (16GB)	6.0s	61.7×	GPU-aware routing
100 PDFs (2000 pages)	3-node cluster (mixed)	3.2 min	13.2×	See BENCHMARKS.md
Embedding generation	RTX A4000 vs i9 CPU	8.2s vs 178s	21.7×	10K chunks

🎯 Try it yourself: pip install flockparser && python showcase/process_arxiv_papers.py

---

🔒 Privacy Model

Interface	Privacy Level	External Calls	Best For
CLI (flockparsecli.py)	🟢 100% Local	None	Personal use, air-gapped systems
Web UI (flock_webui.py)	🟢 100% Local	None	GUI users, visual monitoring
REST API (flock_ai_api.py)	🟡 Local Network	None	Multi-user, app integration
MCP Server (flock_mcp_server.py)	🔴 Cloud	⚠️ Claude Desktop (Anthropic)	AI assistant integration

⚠️ MCP Privacy Warning: The MCP server integrates with Claude Desktop, which sends queries and document snippets to Anthropic's cloud API. Use CLI/Web UI for 100% offline processing.

---

Table of Contents

• Key Features
• 👥 Who Uses This? - Target users & scenarios
• 📐 How It Works (5-Second Overview) - Visual for non-technical evaluators
• Architecture | 📖 Deep Dive: Architecture & Design Decisions
• Quickstart
• Performance & Benchmarks
• 🎓 Showcase: Real-World Example ⭐ Try it yourself
• Usage Examples
• Security & Production
• Troubleshooting
• Contributing

⚡ Key Features

• 🌐 Intelligent Load Balancing - Auto-discovers Ollama nodes, detects GPU vs CPU, monitors VRAM, and routes work adaptively (10x speedup on heterogeneous clusters)
• 🔌 Multi-Protocol Support - CLI (100% local), REST API (network), MCP (Claude Desktop), Web UI (Streamlit) - choose your privacy level
• 🎯 Adaptive Routing - Sequential vs parallel decisions based on cluster characteristics (prevents slow nodes from bottlenecking)
• 📊 Production Observability - Real-time health scores, performance tracking, VRAM monitoring, automatic failover
• 🔒 Privacy-First Architecture - No external API calls required (CLI mode), all processing on-premise
• 📄 Complete Pipeline - PDF extraction → OCR fallback → Multi-format conversion → Vector embeddings → RAG with source citations

---

👥 Who Uses This?

FlockParser is designed for engineers and researchers who need private, on-premise document intelligence with real distributed systems capabilities.

Ideal Users

User Type	Use Case	Why FlockParser?
🔬 ML/AI Engineers	Process research papers, build knowledge bases, experiment with RAG systems	GPU-aware routing, 21× faster embeddings, full pipeline control
📊 Data Scientists	Extract insights from large document corpora (100s-1000s of PDFs)	Distributed processing, semantic search, production observability
🏢 Enterprise Engineers	On-premise document search for regulated industries (healthcare, legal, finance)	100% local processing, no cloud APIs, privacy-first architecture
🎓 Researchers	Build custom RAG systems, experiment with distributed inference patterns	Full source access, extensible architecture, real benchmarks
🛠️ DevOps/Platform Engineers	Set up document intelligence infrastructure for teams	Multi-node setup, health monitoring, automatic failover
👨‍💻 Students/Learners	Understand distributed systems, GPU orchestration, RAG architectures	Real working example, comprehensive docs, honest limitations

Real-World Scenarios

✅ "I have 500 research papers and a spare GPU machine" → Process your corpus 20× faster with distributed nodes ✅ "I can't send medical records to OpenAI" → 100% local processing (CLI/Web UI modes) ✅ "I want to experiment with RAG without cloud costs" → Full pipeline, runs on your hardware ✅ "I need to search 10,000 internal documents" → ChromaDB vector search with sub-20ms latency ✅ "I have mismatched hardware (old laptop + gaming PC)" → Adaptive routing handles heterogeneous clusters

Not Ideal For

❌ Production SaaS with 1000+ concurrent users → Current SQLite backend limits concurrency (~50 users) ❌ Mission-critical systems requiring 99.9% uptime → Still in Beta, see KNOWN_ISSUES.md ❌ Simple one-time PDF extraction → Overkill; use pdfplumber directly ❌ Cloud-first deployments → Designed for on-premise/hybrid; cloud works but misses GPU routing benefits

Bottom line: If you're building document intelligence infrastructure on your own hardware and need distributed processing with privacy guarantees, FlockParser is for you.

---

📐 How It Works (5-Second Overview)

For recruiters and non-technical evaluators:

┌─────────────────────────────────────────────────────────────────┐
│                         INPUT                                    │
│  📄 Your Documents (PDFs, research papers, internal docs)       │
└────────────────────────┬────────────────────────────────────────┘
                         │
                         ▼
┌─────────────────────────────────────────────────────────────────┐
│                     FLOCKPARSER                                  │
│                                                                  │
│  1. Extracts text from PDFs (handles scans with OCR)           │
│  2. Splits into chunks, creates vector embeddings              │
│  3. Distributes work across GPU/CPU nodes (auto-discovery)     │
│  4. Stores in searchable vector database (ChromaDB)            │
│                                                                  │
│  ⚡ Distributed Processing: 3 nodes → 13× faster               │
│  🚀 GPU Acceleration: RTX A4000 → 61× faster than CPU          │
│  🔒 Privacy: 100% local (no cloud APIs)                        │
└────────────────────────┬────────────────────────────────────────┘
                         │
                         ▼
┌─────────────────────────────────────────────────────────────────┐
│                        OUTPUT                                    │
│  🔍 Semantic Search: "Find all mentions of transformers"        │
│  💬 AI Chat: "Summarize the methodology section"                │
│  📊 Source Citations: Exact page/document references            │
│  🌐 4 Interfaces: CLI, Web UI, REST API, Claude Desktop         │
└─────────────────────────────────────────────────────────────────┘

Key Innovation: Auto-detects GPU nodes, measures performance, and routes work to fastest hardware. No manual configuration needed.

---

🏗️ Architecture

┌─────────────────────────────────────────────────────────────┐
│             Interfaces (Choose Your Privacy Level)           │
│  CLI (Local) | REST API (Network) | MCP (Claude) | Web UI   │
└──────────────────────┬──────────────────────────────────────┘
                       │
                       ▼
┌─────────────────────────────────────────────────────────────┐
│                  FlockParse Core Engine                      │
│  ┌─────────────┐  ┌──────────────┐  ┌──────────────┐       │
│  │   PDF       │  │  Semantic    │  │     RAG      │       │
│  │ Processing  │→ │   Search     │→ │   Engine     │       │
│  └─────────────┘  └──────────────┘  └──────────────┘       │
│         │                │                    │              │
│         ▼                ▼                    ▼              │
│  ┌───────────────────────────────────────────────────┐      │
│  │        ChromaDB Vector Store (Persistent)         │      │
│  └───────────────────────────────────────────────────┘      │
└──────────────────────┬──────────────────────────────────────┘
                       │ Intelligent Load Balancer
                       │ • Health scoring (GPU/VRAM detection)
                       │ • Adaptive routing (sequential vs parallel)
                       │ • Automatic failover & caching
                       ▼
    ┌──────────────────────────────────────────────┐
    │       Distributed Ollama Cluster              │
    │  ┌──────────┐  ┌──────────┐  ┌──────────┐   │
    │  │ Node 1   │  │ Node 2   │  │ Node 3   │   │
    │  │ GPU A    │  │ GPU B    │  │ CPU      │   │
    │  │16GB VRAM │  │ 8GB VRAM │  │ 16GB RAM │   │
    │  │Health:367│  │Health:210│  │Health:50 │   │
    │  └──────────┘  └──────────┘  └──────────┘   │
    └──────────────────────────────────────────────┘
         ▲ Auto-discovery | Performance tracking

Want to understand how this works? Read the 📖 Architecture Deep Dive for detailed explanations of:

• Why distributed AI inference solves real-world problems
• How adaptive routing decisions are made (sequential vs parallel)
• MCP integration details and privacy implications
• Technical trade-offs and design decisions

🚀 Quickstart (3 Steps)

Requirements:

• Python 3.10 or later
• Ollama 0.1.20+ (install from ollama.com)
• 4GB+ RAM (8GB+ recommended for GPU nodes)

# 1. Install FlockParser
pip install flockparser

# 2. Start Ollama and pull models
ollama serve  # In a separate terminal
ollama pull mxbai-embed-large    # Required for embeddings
ollama pull llama3.1:latest       # Required for chat

# 3. Run your preferred interface
flockparse-webui                     # Web UI - easiest (recommended) ⭐
flockparse                           # CLI - 100% local
flockparse-api                       # REST API - multi-user
flockparse-mcp                       # MCP - Claude Desktop integration

💡 Pro tip: Start with the Web UI to see distributed processing with real-time VRAM monitoring and node health dashboards.

---

Alternative: Install from Source

If you want to contribute or modify the code:

git clone https://github.com/BenevolentJoker-JohnL/FlockParser.git
cd FlockParser
pip install -e .  # Editable install

Quick Test (30 seconds)

# Start the CLI
python flockparsecli.py

# Process the sample PDF
> open_pdf testpdfs/sample.pdf

# Chat with it
> chat
🙋 You: Summarize this document

First time? Start with the Web UI (streamlit run flock_webui.py) - it's the easiest way to see distributed processing in action with a visual dashboard.

---

🐳 Docker Deployment (One Command)

Quick Start with Docker Compose

# Clone and deploy everything
git clone https://github.com/BenevolentJoker-JohnL/FlockParser.git
cd FlockParser
docker-compose up -d

# Access services
# Web UI: http://localhost:8501
# REST API: http://localhost:8000
# Ollama: http://localhost:11434

What Gets Deployed

Service	Port	Description
Web UI	8501	Streamlit interface with visual monitoring
REST API	8000	FastAPI with authentication
CLI	-	Interactive terminal (docker-compose run cli)
Ollama	11434	Local LLM inference engine

Production Features

✅ Multi-stage build - Optimized image size ✅ Non-root user - Security hardened ✅ Health checks - Auto-restart on failure ✅ Volume persistence - Data survives restarts ✅ GPU support - Uncomment deploy section for NVIDIA GPUs

Custom Configuration

# Set API key
export FLOCKPARSE_API_KEY="your-secret-key"

# Set log level
export LOG_LEVEL="DEBUG"

# Deploy with custom config
docker-compose up -d

GPU Support (NVIDIA)

Uncomment the GPU section in docker-compose.yml:

deploy:
  resources:
    reservations:
      devices:
        - driver: nvidia
          count: all
          capabilities: [gpu]

Then run: docker-compose up -d

CI/CD Pipeline

graph LR
    A[📝 Git Push] --> B[🔍 Lint & Format]
    B --> C[🧪 Test Suite]
    B --> D[🔒 Security Scan]
    C --> E[🐳 Build Multi-Arch]
    D --> E
    E --> F[📦 Push to GHCR]
    F --> G[🚀 Deploy]

    style A fill:#4CAF50
    style B fill:#2196F3
    style C fill:#2196F3
    style D fill:#FF9800
    style E fill:#9C27B0
    style F fill:#9C27B0
    style G fill:#F44336

Automated on every push to main:

Stage	Tools	Purpose
Code Quality	black, flake8, mypy	Enforce formatting & typing standards
Testing	pytest (Python 3.10/3.11/3.12)	78% coverage across versions
Security	Trivy	Vulnerability scanning & SARIF reports
Build	Docker Buildx	Multi-architecture (amd64, arm64)
Registry	GitHub Container Registry	Versioned image storage
Deploy	On release events	Automated production deployment

Pull the latest image:

docker pull ghcr.io/benevolentjoker-johnl/flockparser:latest

View pipeline runs: https://github.com/BenevolentJoker-JohnL/FlockParser/actions

---

🌐 Setting Up Distributed Nodes

Want the 60x speedup? Set up multiple Ollama nodes across your network.

Quick Multi-Node Setup

On each additional machine:

# 1. Install Ollama
curl -fsSL https://ollama.com/install.sh | sh

# 2. Configure for network access
export OLLAMA_HOST=0.0.0.0:11434
ollama serve

# 3. Pull models
ollama pull mxbai-embed-large
ollama pull llama3.1:latest

# 4. Allow firewall (if needed)
sudo ufw allow 11434/tcp  # Linux

FlockParser will automatically discover these nodes!

Check with:

python flockparsecli.py
> lb_stats  # Shows all discovered nodes and their capabilities

📖 Complete Guide: See DISTRIBUTED_SETUP.md for:

• Step-by-step multi-machine setup
• Network configuration and firewall rules
• Troubleshooting node discovery
• Example setups (budget home lab to professional clusters)
• GPU router configuration for automatic optimization

---

🔒 Privacy Levels by Interface:

• Web UI (flock_webui.py): 🟢 100% local, runs in your browser
• CLI (flockparsecli.py): 🟢 100% local, zero external calls
• REST API (flock_ai_api.py): 🟡 Local network only
• MCP Server (flock_mcp_server.py): 🔴 Integrates with Claude Desktop (Anthropic cloud service)

Choose the interface that matches your privacy requirements!

🏆 Why FlockParse? Comparison to Competitors

Feature	FlockParse	LangChain	LlamaIndex	Haystack
100% Local/Offline	✅ Yes (CLI/JSON)	⚠️ Partial	⚠️ Partial	⚠️ Partial
Zero External API Calls	✅ Yes (CLI/JSON)	❌ No	❌ No	❌ No
Built-in GPU Load Balancing	✅ Yes (auto)	❌ No	❌ No	❌ No
VRAM Monitoring	✅ Yes (dynamic)	❌ No	❌ No	❌ No
Multi-Node Auto-Discovery	✅ Yes	❌ No	❌ No	❌ No
CPU Fallback Detection	✅ Yes	❌ No	❌ No	❌ No
Document Format Export	✅ 4 formats	❌ Limited	❌ Limited	⚠️ Basic
Setup Complexity	🟢 Simple	🔴 Complex	🔴 Complex	🟡 Medium
Dependencies	🟢 Minimal	🔴 Heavy	🔴 Heavy	🟡 Medium
Learning Curve	🟢 Low	🔴 Steep	🔴 Steep	🟡 Medium
Privacy Control	🟢 High (CLI/JSON)	🔴 Limited	🔴 Limited	🟡 Medium
Out-of-Box Functionality	✅ Complete	⚠️ Requires config	⚠️ Requires config	⚠️ Requires config
MCP Integration	✅ Native	❌ No	❌ No	❌ No
Embedding Cache	✅ MD5-based	⚠️ Basic	⚠️ Basic	⚠️ Basic
Batch Processing	✅ Parallel	⚠️ Sequential	⚠️ Sequential	⚠️ Basic
Performance	🚀 60x+ faster with GPU auto-routing	⚠️ Varies by config	⚠️ Varies by config	⚠️ Varies by config
Cost	💰 Free	💰💰 Free + Paid	💰💰 Free + Paid	💰💰 Free + Paid

Key Differentiators:

1. Privacy by Design: CLI and JSON interfaces are 100% local with zero external calls (MCP interface uses Claude Desktop for chat)
2. Intelligent GPU Management: Automatically finds, tests, and prioritizes GPU nodes
3. Production-Ready: Works immediately with sensible defaults
4. Resource-Aware: Detects VRAM exhaustion and prevents performance degradation
5. Complete Solution: CLI, REST API, MCP, and batch interfaces - choose your privacy level

📊 Performance

📹 76-Second Demo Video - Watch 6 minutes become 6 seconds

Real-Time Demo Results (unedited timing shown on screen):

Processing Mode	Time	Speedup	What It Shows
Single CPU node	372.76s (~6 min)	1x baseline	Sequential CPU processing
Parallel (multi-node)	159.79s (~2.5 min)	2.3x faster	Distributed across cluster
GPU node routing	6.04s (~6 sec)	61.7x faster	Automatic GPU detection & routing

Why the Massive Speedup?

• GPU processes embeddings in milliseconds vs seconds on CPU
• Adaptive routing detected GPU was 60x+ faster and sent all work there
• Avoided bottleneck of waiting for slower CPU nodes to finish
• No network overhead (local cluster, no cloud APIs)

Demo Contents:

• 0:00 - Single node baseline (372.76s)
• 0:30 - Auto-discover cluster nodes on network
• 0:45 - Parallel processing across nodes (159.79s)
• 1:00 - GPU routing with adaptive decision (6.04s)
• 1:10 - Document chat with RAG + source citations
• 1:15 - MCP integration with Claude Desktop

Key Insight: The system automatically detected performance differences and made routing decisions - no manual GPU configuration needed.

Hardware (Demo Cluster):

• Node 1 (10.9.66.90): Intel i9-12900K, 32GB DDR5-6000, 6TB NVMe Gen4, RTX A4000 16GB - routed here
• Node 2 (10.9.66.159): AMD Ryzen 7 5700X, 32GB DDR4-3600, GTX 1050Ti (CPU-mode)
• Node 3: Intel i7-12th gen (laptop), 16GB DDR5, CPU-only
• Software: Python 3.10, Ollama, Ubuntu 22.04

Reproducibility:

• Timing shown on-screen in real-time (not edited)
• Commands visible in terminal output
• Full source code available in this repo
• Test with your own hardware - results will vary based on GPU

The project offers four main interfaces:

1. flock_webui.py - 🎨 Beautiful Streamlit web interface (NEW!)
2. flockparsecli.py - Command-line interface for personal document processing
3. flock_ai_api.py - REST API server for multi-user or application integration
4. flock_mcp_server.py - Model Context Protocol server for AI assistants like Claude Desktop

---

🎓 Showcase: Real-World Example

Processing influential AI research papers from arXiv.org

Want to see FlockParser in action on real documents? Run the included showcase:

pip install flockparser
python showcase/process_arxiv_papers.py

What It Does

Downloads and processes 5 seminal AI research papers:

• Attention Is All You Need (Transformers) - arXiv:1706.03762
• BERT - Pre-training Deep Bidirectional Transformers - arXiv:1810.04805
• RAG - Retrieval-Augmented Generation for NLP - arXiv:2005.11401
• GPT-3 - Language Models are Few-Shot Learners - arXiv:2005.14165
• Llama 2 - Open Foundation Language Models - arXiv:2307.09288

Total: ~350 pages, ~25 MB of PDFs

Expected Results

Configuration	Processing Time	Speedup
Single CPU node	~90s	1.0× baseline
Multi-node (1 GPU + 2 CPU)	~30s	3.0×
Single GPU node (RTX A4000)	~21s	4.3×

What You Get

After processing, the script demonstrates:

1. Semantic Search across all papers:

   # Example queries that work immediately:
   "What is the transformer architecture?"
   "How does retrieval-augmented generation work?"
   "What are the benefits of attention mechanisms?"
   

2. Performance Metrics (showcase/results.json):

   {
     "total_time": 21.3,
     "papers": [
       {
         "title": "Attention Is All You Need",
         "processing_time": 4.2,
         "status": "success"
       }
     ],
     "node_info": [...]
   }
   

3. Human-Readable Summary (showcase/RESULTS.md) with:

   • Per-paper processing times
   • Hardware configuration used
   • Fastest/slowest/average performance
   • Replication instructions

Why This Matters

This isn't a toy demo - it's processing actual research papers that engineers read daily. It demonstrates:

✅ Real document processing - Complex PDFs with equations, figures, multi-column layouts ✅ Production-grade pipeline - PDF extraction → embeddings → vector storage → semantic search ✅ Actual performance gains - Measurable speedups on heterogeneous hardware ✅ Reproducible results - Run it yourself with pip install, compare your hardware

Perfect for portfolio demonstrations: Show this to hiring managers as proof of real distributed systems work.

---

🔧 Installation

1. Clone the Repository

git clone https://github.com/yourusername/flockparse.git
cd flockparse

2. Install System Dependencies (Required for OCR)

⚠️ IMPORTANT: Install these BEFORE pip install, as pytesseract and pdf2image require system packages

For Better PDF Text Extraction:

• Linux:

  sudo apt-get update
  sudo apt-get install poppler-utils
  

• macOS:

  brew install poppler
  

• Windows: Download from Poppler for Windows

For OCR Support (Scanned Documents):

FlockParse automatically detects scanned PDFs and uses OCR!

• Linux (Ubuntu/Debian):

  sudo apt-get update
  sudo apt-get install tesseract-ocr tesseract-ocr-eng poppler-utils
  

• Linux (Fedora/RHEL):

  sudo dnf install tesseract poppler-utils
  

• macOS:

  brew install tesseract poppler
  

• Windows:
  1. Install Tesseract OCR - Download the installer
  2. Install Poppler for Windows
  3. Add both to your system PATH

Verify installation:

tesseract --version
pdftotext -v

3. Install Python Dependencies

pip install -r requirements.txt

Key Python dependencies (installed automatically):

• fastapi, uvicorn - Web server
• pdfplumber, PyPDF2, pypdf - PDF processing
• pytesseract - Python wrapper for Tesseract OCR (requires system Tesseract)
• pdf2image - PDF to image conversion (requires system Poppler)
• Pillow - Image processing for OCR
• chromadb - Vector database
• python-docx - DOCX generation
• ollama - AI model integration
• numpy - Numerical operations
• markdown - Markdown generation

How OCR fallback works:

1. Tries PyPDF2 text extraction
2. Falls back to pdftotext if no text
3. Falls back to OCR if still no text (<100 chars) - Requires Tesseract + Poppler
4. Automatically processes scanned documents without manual intervention

4. Install and Configure Ollama

1. Install Ollama from ollama.com
2. Start the Ollama service:

   ollama serve
   

3. Pull the required models:

   ollama pull mxbai-embed-large
   ollama pull llama3.1:latest
   

📜 Usage

🎨 Web UI (flock_webui.py) - Easiest Way to Get Started!

Launch the beautiful Streamlit web interface:

streamlit run flock_webui.py

The web UI will open in your browser at http://localhost:8501

Features:

• 📤 Upload & Process: Drag-and-drop PDF files for processing
• 💬 Chat Interface: Interactive chat with your documents
• 📊 Load Balancer Dashboard: Real-time monitoring of GPU nodes
• 🔍 Semantic Search: Search across all documents
• 🌐 Node Management: Add/remove Ollama nodes, auto-discovery
• 🎯 Routing Control: Switch between routing strategies

Perfect for:

• Users who prefer graphical interfaces
• Quick document processing and exploration
• Monitoring distributed processing
• Managing multiple Ollama nodes visually

---

CLI Interface (flockparsecli.py)

Run the script:

python flockparsecli.py

Available commands:

📖 open_pdf <file>   → Process a single PDF file
📂 open_dir <dir>    → Process all PDFs in a directory
💬 chat              → Chat with processed PDFs
📊 list_docs         → List all processed documents
🔍 check_deps        → Check for required dependencies
🌐 discover_nodes    → Auto-discover Ollama nodes on local network
➕ add_node <url>    → Manually add an Ollama node
➖ remove_node <url> → Remove an Ollama node from the pool
📋 list_nodes        → List all configured Ollama nodes
⚖️  lb_stats          → Show load balancer statistics
❌ exit              → Quit the program

Web Server API (flock_ai_api.py)

Start the API server:

# Set your API key (or use default for testing)
export FLOCKPARSE_API_KEY="your-secret-key-here"

# Start server
python flock_ai_api.py

The server will run on http://0.0.0.0:8000 by default.

🔒 Authentication (NEW!)

All endpoints except / require an API key in the X-API-Key header:

# Default API key (change in production!)
X-API-Key: your-secret-api-key-change-this

# Or set via environment variable
export FLOCKPARSE_API_KEY="my-super-secret-key"

Available Endpoints:

Endpoint	Method	Auth Required	Description
/	GET	❌ No	API status and version info
/upload/	POST	✅ Yes	Upload and process a PDF file
/summarize/{file_name}	GET	✅ Yes	Get an AI-generated summary
/search/?query=...	GET	✅ Yes	Search for relevant documents

Example API Usage:

Check API status (no auth required):

curl http://localhost:8000/

Upload a document (with authentication):

curl -X POST \
  -H "X-API-Key: your-secret-api-key-change-this" \
  -F "file=@your_document.pdf" \
  http://localhost:8000/upload/

Get a document summary:

curl -H "X-API-Key: your-secret-api-key-change-this" \
  http://localhost:8000/summarize/your_document.pdf

Search across documents:

curl -H "X-API-Key: your-secret-api-key-change-this" \
  "http://localhost:8000/search/?query=your%20search%20query"

⚠️ Production Security:

• Always change the default API key
• Use environment variables, never hardcode keys
• Use HTTPS in production (nginx/apache reverse proxy)
• Consider rate limiting for public deployments

MCP Server (flock_mcp_server.py)

The MCP server allows FlockParse to be used as a tool by AI assistants like Claude Desktop.

Setting up with Claude Desktop

1. Start the MCP server:

   python flock_mcp_server.py
   

2. Configure Claude Desktop: Add to your Claude Desktop config file (~/Library/Application Support/Claude/claude_desktop_config.json on macOS, or %APPDATA%\Claude\claude_desktop_config.json on Windows):

   {
     "mcpServers": {
       "flockparse": {
         "command": "python",
         "args": ["/absolute/path/to/FlockParser/flock_mcp_server.py"]
       }
     }
   }
   

3. Restart Claude Desktop and you'll see FlockParse tools available!

Available MCP Tools:

• process_pdf - Process and add PDFs to the knowledge base
• query_documents - Search documents using semantic search
• chat_with_documents - Ask questions about your documents
• list_documents - List all processed documents
• get_load_balancer_stats - View node performance metrics
• discover_ollama_nodes - Auto-discover Ollama nodes
• add_ollama_node - Add an Ollama node manually
• remove_ollama_node - Remove an Ollama node

Example MCP Usage:

In Claude Desktop, you can now ask:

• "Process the PDF at /path/to/document.pdf"
• "What documents do I have in my knowledge base?"
• "Search my documents for information about quantum computing"
• "What does my research say about black holes?"

💡 Practical Use Cases

Knowledge Management

• Create searchable archives of research papers, legal documents, and technical manuals
• Generate summaries of lengthy documents for quick review
• Chat with your document collection to find specific information without manual searching

Legal & Compliance

• Process contract repositories for semantic search capabilities
• Extract key terms and clauses from legal documents
• Analyze regulatory documents for compliance requirements

Research & Academia

• Process and convert academic papers for easier reference
• Create a personal research assistant that can reference your document library
• Generate summaries of complex research for presentations or reviews

Business Intelligence

• Convert business reports into searchable formats
• Extract insights from PDF-based market research
• Make proprietary documents more accessible throughout an organization

🌐 Distributed Processing with Load Balancer

FlockParse includes a sophisticated load balancer that can distribute embedding generation across multiple Ollama instances on your local network.

Setting Up Distributed Processing

Option 1: Auto-Discovery (Easiest)

# Start FlockParse
python flockparsecli.py

# Auto-discover Ollama nodes on your network
⚡ Enter command: discover_nodes

The system will automatically scan your local network (/24 subnet) and detect any running Ollama instances.

Option 2: Manual Node Management

# Add a specific node
⚡ Enter command: add_node http://192.168.1.100:11434

# List all configured nodes
⚡ Enter command: list_nodes

# Remove a node
⚡ Enter command: remove_node http://192.168.1.100:11434

# View load balancer statistics
⚡ Enter command: lb_stats

Benefits of Distributed Processing

• Speed: Process documents 2-10x faster with multiple nodes
• GPU Awareness: Automatically detects and prioritizes GPU nodes over CPU nodes
• VRAM Monitoring: Detects when GPU nodes fall back to CPU due to insufficient VRAM
• Fault Tolerance: Automatic failover if a node becomes unavailable
• Load Distribution: Smart routing based on node performance, GPU availability, and VRAM capacity
• Easy Scaling: Just add more machines with Ollama installed

Setting Up Additional Ollama Nodes

On each additional machine:

# Install Ollama
curl -fsSL https://ollama.com/install.sh | sh

# Pull the embedding model
ollama pull mxbai-embed-large

# Start Ollama (accessible from network)
OLLAMA_HOST=0.0.0.0:11434 ollama serve

Then use discover_nodes or add_node to add them to FlockParse.

GPU and VRAM Optimization

FlockParse automatically detects GPU availability and VRAM usage using Ollama's /api/ps endpoint:

• 🚀 GPU nodes with models loaded in VRAM get +200 health score bonus
• ⚠️ VRAM-limited nodes that fall back to CPU get only +50 bonus
• 🐢 CPU-only nodes get -50 penalty

To ensure your GPU is being used:

1. Check GPU detection: Run lb_stats command to see node status
2. Preload model into GPU: Run a small inference to load model into VRAM

   ollama run mxbai-embed-large "test"
   

3. Verify VRAM usage: Check that size_vram > 0 in /api/ps:

   curl http://localhost:11434/api/ps
   

4. Increase VRAM allocation: If model won't load into VRAM, free up GPU memory or use a smaller model

Dynamic VRAM monitoring: FlockParse continuously monitors embedding performance and automatically detects when a GPU node falls back to CPU due to VRAM exhaustion during heavy load.

🔄 Example Workflows

CLI Workflow: Research Paper Processing

1. Check Dependencies:

   ⚡ Enter command: check_deps
   

2. Process a Directory of Research Papers:

   ⚡ Enter command: open_dir ~/research_papers
   

3. Chat with Your Research Collection:

   ⚡ Enter command: chat
   🙋 You: What are the key methods used in the Smith 2023 paper?
   

API Workflow: Document Processing Service

1. Start the API Server:

   python flock_ai_api.py
   

2. Upload Documents via API:

   curl -X POST -F "file=@quarterly_report.pdf" http://localhost:8000/upload/
   

3. Generate a Summary:

   curl http://localhost:8000/summarize/quarterly_report.pdf
   

4. Search Across Documents:

   curl http://localhost:8000/search/?query=revenue%20growth%20Q3
   

🔧 Troubleshooting Guide

Ollama Connection Issues

Problem: Error messages about Ollama not being available or connection failures.

Solution:

1. Verify Ollama is running: ps aux | grep ollama
2. Restart the Ollama service:

   killall ollama
   ollama serve
   

3. Check that you've pulled the required models:

   ollama list
   

4. If models are missing:

   ollama pull mxbai-embed-large
   ollama pull llama3.1:latest
   

PDF Text Extraction Failures

Problem: No text extracted from certain PDFs.

Solution:

1. Check if the PDF is scanned/image-based:

   • Install OCR tools: sudo apt-get install tesseract-ocr (Linux)
   • For better scanned PDF handling: pip install ocrmypdf
   • Process with OCR: ocrmypdf input.pdf output.pdf

2. If the PDF has unusual fonts or formatting:

   • Install poppler-utils for better extraction
   • Try using the -layout option with pdftotext manually:

     pdftotext -layout problem_document.pdf output.txt
     

Memory Issues with Large Documents

Problem: Application crashes with large PDFs or many documents.

Solution:

1. Process one document at a time for very large PDFs
2. Reduce the chunk size in the code (default is 512 characters)
3. Increase your system's available memory or use a swap file
4. For server deployments, consider using a machine with more RAM

API Server Not Starting

Problem: Error when trying to start the API server.

Solution:

1. Check for port conflicts: lsof -i :8000
2. If another process is using port 8000, kill it or change the port
3. Verify FastAPI is installed: pip install fastapi uvicorn
4. Check for Python version compatibility (requires Python 3.7+)

---

🔐 Security & Production Notes

REST API Security

⚠️ The default API key is NOT secure - change it immediately!

# Set a strong API key via environment variable
export FLOCKPARSE_API_KEY="your-super-secret-key-change-this-now"

# Or generate a random one
export FLOCKPARSE_API_KEY=$(openssl rand -hex 32)

# Start the API server
python flock_ai_api.py

Production Checklist:

• ✅ Change default API key - Never use your-secret-api-key-change-this
• ✅ Use environment variables - Never hardcode secrets in code
• ✅ Enable HTTPS - Use nginx or Apache as reverse proxy with SSL/TLS
• ✅ Add rate limiting - Use nginx limit_req or FastAPI middleware
• ✅ Network isolation - Don't expose API to public internet unless necessary
• ✅ Monitor logs - Watch for authentication failures and abuse

Example nginx config with TLS:

server {
    listen 443 ssl;
    server_name your-domain.com;

    ssl_certificate /path/to/cert.pem;
    ssl_certificate_key /path/to/key.pem;

    location / {
        proxy_pass http://127.0.0.1:8000;
        proxy_set_header Host $host;
        proxy_set_header X-Real-IP $remote_addr;
    }
}

MCP Privacy & Security

What data leaves your machine:

• 🔴 Document queries - Sent to Claude Desktop → Anthropic API
• 🔴 Document snippets - Retrieved context chunks sent as part of prompts
• 🔴 Chat messages - All RAG conversations processed by Claude
• 🟢 Document files - Never uploaded (processed locally, only embeddings stored)

To disable MCP and stay 100% local:

1. Remove FlockParse from Claude Desktop config
2. Use CLI (flockparsecli.py) or Web UI (flock_webui.py) instead
3. Both provide full RAG functionality without external API calls

MCP is safe for:

• ✅ Public documents (research papers, manuals, non-sensitive data)
• ✅ Testing and development
• ✅ Personal use where you trust Anthropic's privacy policy

MCP is NOT recommended for:

• ❌ Confidential business documents
• ❌ Personal identifiable information (PII)
• ❌ Regulated data (HIPAA, GDPR sensitive content)
• ❌ Air-gapped or classified environments

Database Security

SQLite limitations (ChromaDB backend):

• ⚠️ No concurrent writes from multiple processes
• ⚠️ File permissions determine access (not true auth)
• ⚠️ No encryption at rest by default

For production with multiple users:

# Option 1: Separate databases per interface
CLI:     chroma_db_cli/
API:     chroma_db_api/
MCP:     chroma_db_mcp/

# Option 2: Use PostgreSQL backend (ChromaDB supports it)
# See ChromaDB docs: https://docs.trychroma.com/

VRAM Detection Method

FlockParse detects GPU usage via Ollama's /api/ps endpoint:

# Check what Ollama reports
curl http://localhost:11434/api/ps

# Response shows VRAM usage:
{
  "models": [{
    "name": "mxbai-embed-large:latest",
    "size": 705530880,
    "size_vram": 705530880,  # <-- If >0, model is in GPU
    ...
  }]
}

Health score calculation:

• size_vram > 0 → +200 points (GPU in use)
• size_vram == 0 but GPU present → +50 points (GPU available, not used)
• CPU-only → -50 points

This is presence-based detection, not utilization monitoring. It detects if the model loaded into VRAM, not how efficiently it's being used.

---

💡 Features

Feature	Description
Multi-method PDF Extraction	Uses both PyPDF2 and pdftotext for best results
Format Conversion	Converts PDFs to TXT, Markdown, DOCX, and JSON
Semantic Search	Uses vector embeddings to find relevant information
Interactive Chat	Discuss your documents with AI assistance
Privacy Options	Web UI/CLI: 100% offline; REST API: local network; MCP: Claude Desktop (cloud)
Distributed Processing	Load balancer with auto-discovery for multiple Ollama nodes
Accurate VRAM Monitoring	Real GPU memory tracking with nvidia-smi/rocm-smi + Ollama API (NEW!)
GPU & VRAM Awareness	Automatically detects GPU nodes and prevents CPU fallback
Intelligent Routing	4 strategies (adaptive, round_robin, least_loaded, lowest_latency) with GPU priority
Flexible Model Matching	Supports model name variants (llama3.1, llama3.1:latest, llama3.1:8b, etc.)
ChromaDB Vector Store	Production-ready persistent vector database with cosine similarity
Embedding Cache	MD5-based caching prevents reprocessing same content
Model Weight Caching	Keep models in VRAM for faster repeated inference
Parallel Batch Processing	Process multiple embeddings simultaneously
Database Management	Clear cache and clear DB commands for easy maintenance (NEW!)
Filename Preservation	Maintains original document names in converted files
REST API	Web server for multi-user/application integration
Document Summarization	AI-generated summaries of uploaded documents
OCR Processing	Extract text from scanned documents using image recognition

Comparing FlockParse Interfaces

Feature	flock_webui.py	flockparsecli.py	flock_ai_api.py	flock_mcp_server.py
Interface	🎨 Web Browser (Streamlit)	Command line	REST API over HTTP	Model Context Protocol
Ease of Use	⭐⭐⭐⭐⭐ Easiest	⭐⭐⭐⭐ Easy	⭐⭐⭐ Moderate	⭐⭐⭐ Moderate
Use case	Interactive GUI usage	Personal CLI processing	Service integration	AI Assistant integration
Document formats	Creates TXT, MD, DOCX, JSON	Creates TXT, MD, DOCX, JSON	Stores extracted text only	Creates TXT, MD, DOCX, JSON
Interaction	Point-and-click + chat	Interactive chat mode	Query/response via API	Tool calls from AI assistants
Multi-user	Single user (local)	Single user	Multiple users/applications	Single user (via AI assistant)
Storage	Local file-based	Local file-based	ChromaDB vector database	Local file-based
Load Balancing	✅ Yes (visual dashboard)	✅ Yes	❌ No	✅ Yes
Node Discovery	✅ Yes (one-click)	✅ Yes	❌ No	✅ Yes
GPU Monitoring	✅ Yes (real-time charts)	✅ Yes	❌ No	✅ Yes
Batch Operations	⚠️ Multiple upload	❌ No	❌ No	❌ No
Privacy Level	🟢 100% Local	🟢 100% Local	🟡 Local Network	🔴 Cloud (Claude)
Best for	🌟 General users, GUI lovers	Direct CLI usage	Integration with apps	Claude Desktop, AI workflows

📁 Project Structure

• /converted_files - Stores the converted document formats (flockparsecli.py)
• /knowledge_base - Legacy JSON storage (backwards compatibility only)
• /chroma_db_cli - ChromaDB vector database for CLI (flockparsecli.py) - Production storage
• /uploads - Temporary storage for uploaded documents (flock_ai_api.py)
• /chroma_db - ChromaDB vector database (flock_ai_api.py)

🚀 Recent Additions

• ✅ GPU Auto-Optimization - Background process ensures models use GPU automatically (NEW!)
• ✅ Programmatic GPU Control - Force models to GPU/CPU across distributed nodes (NEW!)
• ✅ Accurate VRAM Monitoring - Real GPU memory tracking across distributed nodes
• ✅ ChromaDB Production Integration - Professional vector database for 100x faster search
• ✅ Clear Cache & Clear DB Commands - Manage embeddings and database efficiently
• ✅ Model Weight Caching - Keep models in VRAM for 5-10x faster inference
• ✅ Web UI - Beautiful Streamlit interface for easy document management
• ✅ Advanced OCR Support - Automatic fallback to OCR for scanned documents
• ✅ API Authentication - Secure API key authentication for REST API endpoints
• ⬜ Document versioning - Track changes over time (Coming soon)

📚 Complete Documentation

Core Documentation

• 📖 Architecture Deep Dive - System design, routing algorithms, technical decisions
• 🌐 Distributed Setup Guide - ⭐ Set up your own multi-node cluster
• 📊 Performance Benchmarks - Real-world performance data and scaling tests
• ⚠️ Known Issues & Limitations - 🔴 READ THIS - Honest assessment of current state
• 🔒 Security Policy - Security best practices and vulnerability reporting
• 🐛 Error Handling Guide - Troubleshooting common issues
• 🤝 Contributing Guide - How to contribute to the project
• 📋 Code of Conduct - Community guidelines
• 📝 Changelog - Version history

Technical Guides

• ⚡ Performance Optimization - Tuning for maximum speed
• 🔧 GPU Router Setup - Distributed cluster configuration
• 🤖 GPU Auto-Optimization - Automatic GPU management
• 📊 VRAM Monitoring - GPU memory tracking
• 🎯 Adaptive Parallelism - Smart workload distribution
• 🗄️ ChromaDB Production - Vector database scaling
• 💾 Model Caching - Performance through caching
• 🖥️ Node Management - Managing distributed nodes
• ⚡ Quick Setup - Fast track to getting started

Additional Resources

• 📹 Demo Video (76 seconds) - Watch FlockParser in action
• 📦 Docker Setup - Containerized deployment
• ⚙️ Environment Config - Configuration template
• 🧪 Tests - Test suite and CI/CD

📝 Development Process

This project was developed iteratively using Claude and Claude Code as coding assistants. All design decisions, architecture choices, and integration strategy were directed and reviewed by me.

🤝 Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.