ForgeRAG

ForgeRAG is an end-to-end RAG system open-sourced by the deeplethe organization, focused on structure-aware reasoning and production-grade usability. The backend is built on Python (FastAPI) with a Vue 3 frontend, released under the MIT license.

Core Architecture#

The system employs a dual-reasoning retrieval mechanism: BM25 and vector retrieval serve as pre-filtering, followed by LLM tree navigation and knowledge graph deep reasoning, with final ranking via RRF (Reciprocal Rank Fusion). The knowledge graph supports three levels of retrieval: local neighborhood traversal, global cross-lingual keyword matching, and relational semantic search. It achieved a 55.48% overall win rate against LightRAG on the UltraDomain benchmark, covering agriculture, CS, law, and mixed domains.

Data Flow#

1. Ingestion: Document upload → PDF parsing (pymupdf/minerU/minerU-vlm) → chunking + LLM tree structure construction → entity/relation extraction → vectorization → distributed persistence across relational DB, vector store, graph store, and blob storage.
2. Retrieval: User query → BM25 + vector pre-filtering → LLM tree navigation + KG dual-level retrieval → RRF fusion ranking.
3. Answering: Fused context + KG-synthesized entity/relation descriptions → LLM generates answers with [c_N] precise citation markers.

Key Features#

• Pixel-precise citations: Each claim carries a citation marker; clicking navigates to the exact source document page with highlighted bounding boxes, suitable for compliance/legal scenarios.
• Full retrieval tracing: End-to-end inspection of path scores, expansion decisions, and merge logic.
• Multi-format document ingestion: Native support for PDF, DOCX, PPTX, XLSX, HTML, Markdown, TXT. PDF parsing supports fast mode (pymupdf), layout-aware mode (mineru), and VLM mode (mineru-vlm for scanned documents).
• Multi-turn conversation: Continuous follow-up questions based on conversational context.
• YAML-first configuration: A single forgerag.yaml controls all behavior with no hidden runtime state.
• Per-request overrides: Dynamic switching of retrieval paths, top-k, and reranking strategies via QueryOverrides, suitable for SDK integration and A/B testing.

Use Cases#

• Enterprise-grade multi-format knowledge base QA
• Compliance/legal scenarios requiring page-level bounding-box traceability
• Cross-document supply-chain multi-hop correlation analysis (e.g., "Which Apple suppliers also supply Samsung?")
• RAG architecture baseline evaluation tool (built-in benchmark module)

Deployment & Engineering#

Supports both local development and Docker deployment, both with interactive configuration wizards. Multi-worker horizontal scaling is supported. Prerequisites: Python 3.10+, Node.js 18+ (frontend build only), LLM API Key (LiteLLM-compatible), recommended 4+ CPU cores and 8GB+ RAM (16GB+ for large documents with KG extraction).

Backend Compatibility#

• Relational databases: SQLite (default), PostgreSQL, MySQL
• Vector stores: ChromaDB (default), pgvector, Qdrant, Milvus, Weaviate
• Graph stores: NetworkX in-memory (default), Neo4j
• Blob storage: Local filesystem (default), Amazon S3, Alibaba Cloud OSS
• LLM / Embeddings: Full compatibility with all LiteLLM-supported providers (OpenAI, Azure, Anthropic, Ollama, DeepSeek, Cohere, etc.)

Key API Endpoints#

Interactive docs: Swagger UI at /docs, ReDoc at /redoc.

Unconfirmed Information#

• Background of the deeplethe organization and DeepLethe Team is not detailed in the repository
• No independent academic paper link provided
• Roadmap has no specific timelines
• No public production use cases or customer references provided

Released under the MIT license, Copyright (c) 2025 DeepLethe Team and contributors.