title: Zaza Semantic Engine emoji: 🧠 colorFrom: indigo colorTo: blue sdk: gradio sdk_version: 4.44.0 app_file: app.py pinned: false license: mit short_description: Local-first multilingual semantic search (50+ languages)

Zaza Semantic Engine

Local-first multi-format document ingestion engine with real semantic search.

Why Zaza?

Most document tools fall into two camps: cloud-based SaaS (your docs leave your machine) or dumb keyword search (finds exact word matches, misses the point). Zaza does both locally and semantically.

• Local-first — your documents never leave your machine. No API keys, no data leaks.
• Semantic search — find documents by meaning, not just keywords. Search "budget" and it finds "financial analysis", "quarterly results".
• Multi-format — TXT, PDF, Markdown, DOCX, JSON, YAML, EPUB, CSV, HTML, XML. Ingest anything.
• 50+ languages — built on paraphrase-multilingual-MiniLM-L12-v2. Search in French, English, Arabic, or any supported language.
• Zero config — zaza ingest ./docs/ and you're done.

Installation

# Core package
pip install -e .

# With API support
pip install -e ".[api]"

# With semantic search (embeddings + multilingual model)
pip install -e ".[semantic]"

# Full installation
pip install -e ".[all]"

Quick Start

# Ingest documents
zaza ingest ./my-documents/

# Keyword search (by filename)
zaza search "report"

# Semantic search (by meaning)
zaza search-semantic "financial analysis quarterly results" --top 5

# View stats
zaza stats

# Start API server (V3: either form works)
zaza api
zaza server

Semantic Search in Action

This project uses sentence-transformers (paraphrase-multilingual-MiniLM-L12-v2) to generate embeddings and ChromaDB for vector storage.

Unlike keyword search, semantic search finds documents with related concepts even when the exact words differ:

Demo

Try it live on Hugging Face: Zaza Semantic Search Space

CLI Commands

API Endpoints

Supported Formats

Model Caching (V3)

The embedding model is cached globally within a single process. zaza ingest + zaza search-semantic doesn't reload the model — it reuses the cached instance. Startup time drops significantly.

Configuration

Edit config.yaml to customize paths, embedding models, and search settings.

semantic:
  enabled: true                    # Set false to disable embeddings
  model_name: "sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2"
  embed_dir: "./data/embeddings"   # ChromaDB persist directory
  max_search_results: 10

License

MIT