Problem
Large collections of manuals, SOPs and training documents are searchable in theory but expensive to navigate in practice. Generic file-chat tools can answer simple questions, but production use often needs stronger evidence handling, document boundaries, retrieval control, multilingual access and private source policies.
System objective
FactoryMind was designed as a reusable knowledge layer: retrieve the most relevant evidence, synthesize only what the sources support, show confidence, and let the user inspect the exact excerpt used.
What the live demo proves
- Cross-manual retrieval across a real 666-page corpus.
- Persian questions over English source material with Persian grounded answers.
- Page-level citations and inspectable evidence excerpts.
- Conservative handling when the corpus does not support the requested claim.
- Original PDFs remain outside the public download surface.
- Production deployment with health checks, rate limits, HTTPS and restricted public endpoints.
Public disclosure boundary: the production source code, prompts, ranking implementation, deployment secrets, source PDFs and licensed assets are intentionally private. The GitHub repository is a public case study, not an open-source release.
Why multilingual retrieval matters
Question language, source language and answer language do not have to match. The current demo can take a Persian question, retrieve the relevant English passages and return a Persian answer while preserving the original English evidence for verification.
Product direction
The same engine is intended to support isolated libraries for technical manuals, internal SOPs, engineering knowledge, training collections and private company documentation โ with per-library permissions, retrieval policies and usage limits.