AlphaNote

A personal RAG-powered investment thesis journal with local LLM integration.

Features

• Chat with your Notes - Query historical investment theses using semantic search
• Log New Thesis - Record asset/ticker, market stance (Bullish/Neutral/Bearish), and reasoning
• Accountability - Each entry requires an author name
• Recent Intelligence - Preview your 6 most recent entries
• Incremental RAG - New entries are indexed instantly without full rebuild
• Local LLM - Private AI analysis via Ollama (llama3.2)
• Persistent Storage - Markdown files + ChromaDB vector database

Quick Start

# 1. Install dependencies
pip install -r requirements.txt

# 2. Install & start Ollama (required for chat)
# Download from: https://ollama.ai
ollama pull llama3.2
ollama pull nomic-embed-text
ollama serve

# 3. Run AlphaNote
python app.py

Open http://127.0.0.1:7860

If port 7860 is occupied, either:

ALPHANOTE_PORT=7861 python app.py

or stop old process first:

lsof -nP -iTCP:7860 -sTCP:LISTEN
kill <PID>

File Structure

personal-RAG/
├── app.py                  # Main application
├── requirements.txt        # Python dependencies
├── knowledge-base/         # Markdown storage
│   └── entries_001.md      # Investment thesis journal
└── vector_db/              # ChromaDB (auto-generated)

Entry Format

Entries are stored in knowledge-base/entries_XXX.md:

---
## GOLD | Bullish | 2026-01-31 14:30 | Oliver

Gold remains attractive as a hedge against uncertainty...

When a file exceeds 500KB, a new file (entries_002.md, etc.) is created automatically.

Configuration

Edit app.py constants:

Variable	Default	Description
MAX_FILE_SIZE_KB	500	File size limit before rotation
USE_LOCAL_LLM	True	Use Ollama (True) or OpenAI (False)
LOCAL_MODEL	llama3.2	Ollama model name
OPENAI_MODEL	gpt-4o-mini	OpenAI model (if USE_LOCAL_LLM=False)
EMBEDDING_MODEL	nomic-embed-text	Ollama embedding model

Startup Issue: Root Cause and Fix

Root Cause

1. Port conflict: older AlphaNote processes were still listening on 7860, so Gradio failed with: Cannot find empty port in range....
2. Environment drift with latest dependency versions: eager imports in the original startup path (langchain_huggingface, langchain_openai, and heavy splitters/loaders) could block startup in this local Python/venv setup.

Fix Implemented

1. Made launch port configurable via ALPHANOTE_PORT; if not set, app uses Gradio default behavior.
2. Reduced startup-time heavy imports:
   • switched embeddings to OllamaEmbeddings (nomic-embed-text)
   • replaced LangChain ChatOpenAI call path with direct Ollama HTTP call
   • replaced heavy text-splitter/doc-loader path with lightweight local chunking
3. Deferred vector DB initialization until first request, so UI can start quickly and reliably.

Performance Optimizations

• Incremental Updates: New entries are added to vector DB without rebuilding
• Persistent Embeddings: Model loaded once at startup
• Disk-based VectorDB: ChromaDB persists between restarts

Requirements

• Python 3.10+
• Ollama running locally (ollama serve)
• ~2GB disk space for llama3.2 model