RagLeap Core Open-Sources a Self-Hosted AI Business Engine

Summary

RagLeap Core is the MIT-licensed, self-hosted engine behind the RagLeap AI business platform. It provides a document-grounded chat pipeline that ingests files or URLs, creates embeddings, retrieves relevant passages, and generates cited answers on infrastructure controlled by the operator. The repository includes 46 role-based AI Employee definitions, spanning nine generalist roles and 37 vertical roles, plus persistent business-context memory, automatic workflow triggers, escalation, and off, semi-autonomous, and full-autonomy modes. Its core stack supports PostgreSQL with pgvector, optional Neo4j entity-graph retrieval, hybrid dense and full-text search using Reciprocal Rank Fusion, streaming responses, per-call generation controls, token reporting, context trimming, and automatic fallback across configured LLM providers. The project uses a bring-your-own-key model: generation can use providers including Gemini, OpenAI, Anthropic, Mistral, DeepSeek, Moonshot, Zhipu, and OpenAI-compatible custom endpoints, while Ollama can run locally without an API key; embeddings currently use Gemini. Single-tenant WhatsApp, Telegram, Discord, web chat, and Twilio voice adapters are included, although the voice Whisper/TTS round trip has not been fully tested and OpenAI voice support is the current v1 limitation. Optional integrations cover nine external data sources, with encrypted credentials and owner-configured SQL actions, including writes. Neo4j graph retrieval, language detection, ingestion canary checks, provider fallback, and token trimming are described as end-to-end verified, but several capabilities remain limited or incomplete. The open repository does not include the hosted platform’s multi-tenancy, billing, audit history, advanced memory controls, and other commercial features. It can be run with Docker Compose or installed as separate MIT-licensed Python packages, with users supplying their own infrastructure and provider credentials.