Moxie RAG

Multi-tenant RAG (Retrieval-Augmented Generation) system with hybrid SQLite + ChromaDB memory storage. Built for GPU-accelerated semantic search on AMD ROCm.

Features

• Multi-tenant document storage — Isolated collections per user/tenant
• Hybrid memory system — SQLite for structured data + ChromaDB for vector embeddings
• GPU-accelerated embeddings — Runs on AMD ROCm (RX 7900 XTX tested)
• Progressive disclosure — Token-efficient retrieval pattern
• File ingestion — PDF, DOCX, Excel, TXT, MD, CSV support
• Audio/video transcription — Whisper integration for media files
• Email polling — Auto-ingest attachments from configured IMAP accounts

Quick Start

# Clone and start
git clone https://git.oe74.net/adelorenzo/moxie-rag.git
cd moxie-rag
docker compose up -d

# Check health
curl http://localhost:8899/health

API Endpoints

Document Operations

Endpoint	Method	Description
/	GET	Service info and collections
/health	GET	Health check
/collections	GET	List all collections
/ingest	POST	Ingest text content
/ingest-file	POST	Upload and ingest file
/query	POST	Semantic search
/documents	GET	List indexed documents
/documents/{id}	DELETE	Delete document
/transcribe	POST	Transcribe audio/video

Memory Operations

Endpoint	Method	Description
/memory/save	POST	Save observation (auto-embeds)
/memory/query	POST	Semantic search with progressive disclosure
/memory/get	POST	Fetch full observations by IDs
/memory/timeline	POST	Chronological context around time/ID
/memory/preference	POST	Save user preference
/memory/preferences/{user_id}	GET	Get all preferences
/memory/stats/{user_id}	GET	Memory statistics

Memory System

The memory system provides user-isolated observation storage with vector search:

# Save an observation
curl -X POST http://localhost:8899/memory/save \
  -H "Content-Type: application/json" \
  -d '{
    "user_id": "alice",
    "content": "Learned that the API requires auth header X-API-Key",
    "type": "learning",
    "title": "API Auth Discovery"
  }'

# Query memory (progressive disclosure - index first)
curl -X POST http://localhost:8899/memory/query \
  -H "Content-Type: application/json" \
  -d '{
    "user_id": "alice",
    "query": "API authentication",
    "include_content": false
  }'

# Get full details by IDs
curl -X POST http://localhost:8899/memory/get \
  -H "Content-Type: application/json" \
  -d '{
    "user_id": "alice",
    "ids": [1, 2, 3]
  }'

Observation Types

• general — Default type
• learning — Learned information
• decision — Decisions made
• preference — User preferences
• tool_use — Tool usage logs
• system_event — System events

Configuration

Environment variables (set in docker-compose.yml):

Variable	Default	Description
CHROMA_DIR	/app/data/chromadb	ChromaDB storage path
MEMORY_DB	/app/data/memory.db	SQLite database path
WHISPER_URL	http://host.docker.internal:8081/transcribe	Whisper API endpoint
UPLOAD_DIR	/app/data/uploads	File upload storage
LOG_DIR	/app/logs	Log directory

Architecture

┌─────────────────────────────────────────────────────┐
│                    FastAPI Server                    │
│                    (port 8899)                       │
├─────────────────────────────────────────────────────┤
│  Document Endpoints  │  Memory Endpoints            │
│  /ingest, /query     │  /memory/save, /memory/query │
├──────────────────────┴──────────────────────────────┤
│                    RAG Engine                        │
│         (sentence-transformers + ChromaDB)          │
├─────────────────────────────────────────────────────┤
│  ChromaDB (vectors)  │  SQLite (structured data)    │
│  /app/data/chromadb  │  /app/data/memory.db         │
└─────────────────────────────────────────────────────┘

Email Pollers

Auto-ingest attachments from IMAP accounts:

• zeus-email-poller — Polls zeus@zz11.net → zeus_docs collection
• moxie-email-poller — Polls moxie@zz11.net → adolfo_docs collection

Requirements

• Docker with compose v2
• AMD GPU with ROCm 7.2+ (or modify Dockerfile for CUDA/CPU)
• Optional: Whisper API for transcription

License

Private repository.