Inspecly's AI Pipeline Puts Structured Data Before RAG and LLMs

The team behind automotive diagnostics app Inspecly shared lessons from building an AI pipeline that deliberately avoids starting with trendy tools like RAG or multimodal models. Instead of routing all inputs into a single large language model, they first normalize heterogeneous data — including OBD codes, photos, voice transcriptions, and vehicle metadata — recognizing that each carries a different level of reliability. Structured data such as OBD diagnostic codes is queried directly against a curated internal database before any semantic retrieval is attempted, since exact lookups outperform RAG for well-defined identifiers. Tool-using agents are only introduced when validated structured knowledge is absent, and retrieved information is explicitly tagged as unverified to preserve data provenance. The core architectural lesson is that knowing what information you have and how trustworthy it is should precede any decision about which AI model or retrieval method to use.