Industrial RAG & Local Knowledge Agent

Problem

Machine knowledge was spread across hundreds of manuals, electrical documents and working files. Engineers needed a searchable assistant that could operate close to the data and support both local and external LLM experiments.

What I built

• Deployed RAGFlow on ARM64/NVIDIA Jetson with its Elasticsearch, MySQL, MinIO and Redis dependencies in Docker.
• Configured local Qwen embedding and reranking components, plus local Qwen inference and an OpenAI-compatible OpenRouter path.
• Curated, deduplicated and indexed technical documentation into a machine-focused knowledge base.

Architecture

technical documents → parsing/indexing → Elasticsearch knowledge base → embedding + reranking → local or external LLM

Separate profiles supported local document work and broader research queries while keeping the retrieval layer consistent.

Scale

The source collection contained hundreds of technical files. Because intermediate inventory and indexing counts differed during cleanup, the public case intentionally avoids claiming a single exact document total.

Result

• Produced a working industrial documentation search and question-answering environment on ARM64 hardware.
• Validated interchangeable local and OpenAI-compatible inference paths without exposing internal infrastructure details.