RAG System Configuration

Memory and knowledge retrieval for 1ai-skills

Overview

This RAG system provides:

• Local vector storage using ZVec (fast, lightweight)
• Self-learning memory using RuVector (GNN-powered)
• Automatic context retrieval on session start
• Skill-aware memory for different domains

Installation

# Install ZVec (fast local vector DB)
pip install zvec

# Or install RuVector (self-learning)
npm install ruvector

Collections

strategic_decisions

{
    "name": "strategic_decisions",
    "dimension": 384,
    "description": "Key strategic decisions and rationale",
    "index_type": "hnsw",
    "metadata_fields": ["date", "topic", "outcome", "decision_maker"]
}

lessons_learned

{
    "name": "lessons_learned", 
    "dimension": 384,
    "description": "Failures, fixes, and insights",
    "index_type": "hnsw", 
    "metadata_fields": ["date", "category", "severity", "fixed"]
}

market_insights

{
    "name": "market_insights",
    "dimension": 384,
    "description": "Competitor analysis, market trends, opportunities",
    "index_type": "hnsw",
    "metadata_fields": ["date", "source", "confidence", "action_items"]
}

team_context

{
    "name": "team_context",
    "dimension": 384,
    "description": "User preferences, team strengths, working style",
    "index_type": "flat",
    "metadata_fields": ["person", "role", "last_updated"]
}

project_memory

{
    "name": "project_memory",
    "dimension": 384,
    "description": "Active projects, milestones, blockers",
    "index_type": "hnsw",
    "metadata_fields": ["project", "status", "priority", "owner"]
}

Usage

Python API

import zvec

# Initialize
db = zvec.open("1ai-memory.db")

# Create collections
db.create_collection("strategic_decisions", dimension=384)
db.create_collection("lessons_learned", dimension=384)

# Add memory
db.insert("strategic_decisions", {
    "id": "decision_2026_02_27",
    "vector": embedding,
    "metadata": {
        "topic": "cashflow_strategy",
        "date": "2026-02-27",
        "outcome": "approved",
        "decision_maker": "Vilona"
    }
})

# Retrieve
results = db.search("strategic_decisions", query_embedding, top_k=5)

With RuVector (Self-Learning)

from ruvector import VectorDB

# Initialize with GNN
db = VectorDB("memory.db", enable_gnn=True)

# Add with learning
db.insert("strategic_decisions", {
    "vector": embedding,
    "metadata": {...},
    "learn": True  # GNN learns from this query
})

# Search improves over time
results = db.search(query, top_k=5)  # Gets smarter

Integration Points

Session Start

# Auto-retrieve relevant context
def on_session_start(user_id, query):
    # Query all collections
    context = []
    for collection in ["strategic_decisions", "lessons_learned", "project_memory"]:
        results = db.search(collection, query, top_k=3)
        context.extend(results)
    
    return context

Memory Update

# After significant decision
def remember_decision(decision_text, metadata):
    embedding = model.encode(decision_text)
    db.insert("strategic_decisions", {
        "vector": embedding,
        "metadata": metadata
    })

Auto-Configuration

The RAG system auto-configures based on 1ai-skills installation:

1. Install → Creates memory/ folder in workspace
2. Setup → Initializes ZVec database
3. First Run → Loads existing memories if any
4. Ongoing → Auto-indexes new decisions

File Structure

workspace/
├── memory/
│   ├── 1ai-memory.db      # ZVec database
│   ├── 2026-02-27.md      # Daily notes
│   └── ...
├── SOUL.md                 # AI identity
├── USER.md                 # User context
└── IDENTITY.md             # Character

---

This RAG system powers Vilona's memory. Every significant decision, lesson learned, and insight gets stored and retrieved automatically.