Machinery AI

Technical support knowledge lives across manuals, scanned pages, tables and the memories of specialists. Keyword search alone misses the relationship between a symptom, a component and the right procedure, while a generic chatbot cannot show where an answer came from.

For an enterprise workflow, retrieval quality was only half the problem. The system also needed controlled access, isolated user data, document operations and enough telemetry to diagnose a weak answer in production.

I built a React and TypeScript workspace around a FastAPI backend, with server-sent events for responsive chat and dedicated surfaces for documents, profiles and administration.

The ingestion pipeline accepts PDF, DOCX, PPTX, XLSX, JPG and PNG files, then combines OCR, table extraction, semantic chunking and metadata enrichment. At query time, Weaviate blends vector and BM25 search before Cohere reranking and context compression; every composed answer retains source attribution.

Argon2 password hashing, JWT sessions, email verification, role-based access and Weaviate multi-tenancy protect the workflow. Docker Compose, Prometheus, Grafana, Sentry and structured logs make the complete system operable instead of leaving it as a retrieval demo.

The repository delivers a container-ready support platform spanning streaming chat, document intelligence, authentication, admin workflows and monitoring. It demonstrates how a serious RAG product is assembled around retrieval — not just how to call a model with a few chunks.