RAGForge

Multi-tenant knowledge-base retrieval and RAG service with a native FastAPI API, OpenAI-compatible endpoints, a Next.js console, and a synchronous Python SDK.

This README describes the current main branch as of 2026-08-19. Read the Chinese capability map for the complete, implementation-verified feature inventory, dependencies, API routes, and limitations.

Highlights

• Tenant isolation, API-key roles/scopes, document ACL security trimming, audit logging, rate limiting, and admin operations.
• Knowledge-base CRUD and text/URL/file/batch ingestion with document status, processing logs, chunk inspection, and interruption.
• Pluggable parsers, chunkers, enrichers, indexers and retrievers: dense, BM25, hybrid, fusion, HyDE, multi-query, self-query, parent-document, ensemble, LlamaIndex and RAPTOR.
• Traceable multimodal RAG: MinerU PDF parsing, image/table/formula content units, Qwen VL embedding, Qwen VL rerank provider wiring, /v2/retrieve, and chat source panels.
• Synchronous/SSE RAG, persisted conversations, OpenAI-compatible Chat Completions and Embeddings, plus multi-provider LLM/embedding/rerank configuration.
• PostgreSQL pgvector, in-memory BM25 or OpenSearch, Redis cache/rate limiting/Celery broker, and MinIO/S3-compatible object storage.
• Docker Compose deployment, health/readiness/metrics endpoints, maintenance scripts, console, and Python SDK.

Quick start

# Default: PostgreSQL pgvector + Redis + MinIO + in-memory BM25
./deploy.sh up

# Shared sparse index for multi-instance or production-like search
./deploy.sh up search

curl http://localhost:8020/health
curl http://localhost:8020/ready

Configure a strong ADMIN_TOKEN and at least the required model provider credentials before creating a tenant. The API listens on 8020 and the console on 3003.

File uploads are queued through Celery with Redis; run the API and worker services together, then poll GET /v1/documents/{id} until processing_status is completed.

Documentation

• 中文 README
• Capability map
• Getting started
• Multimodal RAG deployment
• API reference
• Pipeline architecture
• Operations
• Python SDK

Development

uv sync
uv run pytest tests/ -v
uv run ruff check .
npm --prefix docs run build

Production deployments must run Alembic migrations, replace all sample credentials, restrict CORS and network access to internal dependencies, configure TLS, and back up persistent volumes. For MinerU + Qwen VL embedding/rerank, keep secrets outside git, configure QWEN_RERANK_API_BASE, and run the smoke scripts documented in Multimodal RAG deployment.