Antfly: A Distributed Search Engine for Multimodal AI Data

Antfly is a distributed search engine written in Zig, focusing on multimodal data such as text, images, audio, and video. It provides hybrid search capabilities, combining full-text search (BM25), dense vector similarity, and graph traversal within a single query. The system automatically generates embeddings, chunks data, and creates graph edges as data is ingested.

Key features include built-in RAG (Retrieval Augmented Generation) agents with streaming, multi-turn chat, tool calling (web search, graph traversal), and confidence scoring. It supports multimodal indexing and search using models like CLIP, CLAP, and vision-language models. For enhanced search quality, Antfly incorporates cross-encoder reranking with score-based pruning.

The architecture is distributed, utilizing a multi-Raft design for metadata and storage, ensuring ACID transactions, automatic sharding, replication, and horizontal scaling. It also offers S3 storage integration for cost savings and faster shard splits. Antfly provides enrichment pipelines for embeddings, summaries, and custom fields, and allows users to bring their own models from platforms like Ollama, OpenAI, Bedrock, and Google, or run models locally with Antfly inference.

Additional functionalities include authentication, backup and restore options, a Kubernetes operator for cluster management, and a PostgreSQL extension (`pgaf`) to integrate Antfly search into Postgres. The project also includes React components for building search UIs and an inference runtime for various ML tasks.