Technical Manuals Turned Into a Cited AI Assistant
A multi-tenant RAG platform that answers technicians from text, tables and diagrams in their own equipment manuals, one QR scan from the machine
Client
Manual Minds AI
Timeline
7 months
Technology
RAG SYSTEM
3
content types searched per question, text, tables and diagrams, with the source page cited
1
isolated knowledge base per machine, opened by scanning its QR code
89
API routes across chat, ingestion, assets, workspaces and usage reporting
7 months
of build, from the first commit in May 2025 to January 2026
The challenge
Critical procedures, specifications and wiring diagrams locked inside long PDF manuals, read by technicians who are not always specialists and need the right page for the right machine, fast
The solution
A FastAPI and Next.js platform that extracts text, tables and diagrams from every manual, indexes each machine separately in ChromaDB and streams GPT-4o answers with page-level citations, safety notes first
The results
- 3 content types searched per question, text, tables and diagrams, with the source page cited
- 1 isolated knowledge base per machine, opened by scanning its QR code
- 89 API routes across chat, ingestion, assets, workspaces and usage reporting
- 7 months of build, from the first commit in May 2025 to January 2026
The Answer Is in the Manual. Finding It Is the Problem.
Manual Minds AI is an industrial document intelligence platform that turns static technical manuals into a conversational assistant for manufacturing, maintenance and technical support teams. EdgeFirm designed and built it end to end, from PDF ingestion to the technician's phone.
Industrial equipment ships with long PDF manuals: operating procedures, error code tables, part numbers, tolerances, exploded views and wiring diagrams. When a machine stops, the technician in front of it is often not the person who knows that manual best. They need the procedure for this machine, the right table row and the right diagram, and they need to trust where it came from.
Generic chatbots do not solve that. Half of what matters in a manual is not prose: it sits in tables and drawings, and a wrong answer about a pressurised or electrical system is a safety problem, not an inconvenience. Answers also had to stay inside the right boundary. A supplier serving several customers cannot let one site's documents leak into another's answers, and a question about one machine should not be answered from the manual of the machine next to it.
The brief was a platform that reads manuals the way an engineer does, across text, tables and visuals, answers in plain language with page-level citations, and puts that assistant one scan away from the equipment itself.
4 capabilities carried the build
We built a Python back end on FastAPI with a separate Celery processing tier, and a Next.js web app for administrators, workspace owners and technicians. Four workstreams carried the build.
01
Multimodal manual ingestion
Uploaded PDFs are processed in the background with live progress in the interface. PyMuPDF extracts page text in parallel, pdfplumber pulls tables, Docling adds captions and layout structure, and diagrams are captured both as embedded images and as vector drawings clustered from the page. Duplicate images are removed with perceptual hashing, every image is described by GPT-4o vision with its part numbers and labels, and everything is chunked, embedded and stored in ChromaDB with its file and page reference.
02
Cited, safety-first answers
A LangChain ReAct agent on GPT-4o works with dedicated text, table, image and cross-reference search tools, then streams the answer to the browser over server-sent events. Off-topic questions are screened out before any retrieval runs, follow-up questions carry the conversation context, and repeated questions can reuse an earlier answer. Every answer cites the document and page, opens the PDF at that page, and puts safety notes such as PPE and lockout and tagout first when a task involves electrical, pressurised or moving parts. Answers can close with the workspace's supplier contact details.
03
Assets and QR access
Administrators create assets, the physical machines, and assign them to customer workspaces. Each asset gets its own ChromaDB collection, its own chat history and a generated QR code label stored on S3. Scanning the code opens a mobile chat scoped to that machine's documents only, and access tokens can be rotated with a fresh code. Workspace owners can request more assets within a quota, and administrators review each request.
04
Multi-tenant administration and usage
Workspaces, roles for administrators, owners and users, and email invitations run on Clerk, with webhooks keeping user state in sync. Token and embedding usage is logged with its cost per workspace, storage is tracked, and billing reports are generated as PDFs. An admin dashboard covers users, workspaces, assets, invitations and usage, with soft deletion and recovery built in.
The stack behind it
OpenAI GPT-4o, LangChain, ChromaDB, FastAPI, Celery and Redis, PostgreSQL, Next.js, Clerk, Docling, AWS S3 and EC2, Docker
OpenAI
LangChain
AWS
PostgreSQL
RedisAI and retrieval
OpenAI GPT-4o for answers and image analysis, text-embedding-ada-002, LangChain ReAct agents with custom search tools, and ChromaDB with a separate collection per asset
Document processing
PyMuPDF, pdfplumber and Docling for text, tables, captions and layout, plus Pillow and imagehash for image extraction and deduplication
Back end
Python and FastAPI with async SQLAlchemy, PostgreSQL and Alembic migrations, streaming answers over server-sent events
Background processing
Celery workers on Redis with separate document and image queues, and Flower for queue monitoring
Front end
Next.js with React, TypeScript, Redux Toolkit and Tailwind CSS, an in-app PDF viewer and a mobile chat for QR access
Identity and infrastructure
Clerk for sign-in and invitations, AWS S3 for files and QR codes, and Docker Compose on AWS EC2 deployed through GitHub Actions
Our development process
Months 1 to 2
Core retrieval
We started with PDF upload, parallel text extraction and a first conversational agent over the manuals, then added streamed answers, follow-up handling, table ingestion and page-level references that open the source PDF.
Month 3
Platform foundations
We moved identity to Clerk and built workspaces, invitations, an admin dashboard, background ingestion with live status and usage tracking for AI and storage costs.
Month 4
Visuals and tenancy
We added image extraction with GPT-4o vision analysis, embedding cost tracking, soft deletion and supplier details per workspace, so answers can point technicians to the right support contact.
Month 5
Assets and production hardening
We introduced assets with their own ChromaDB collections, QR codes and a mobile scan interface, moved processing to Celery workers with separate document and image queues, and added Docling for captions, tables and layout.
Months 6 to 7
Stabilisation
We tuned the containers for the production server's memory, improved retrieval quality for asset-scoped questions, hardened ChromaDB collection naming and updated the web app's build tooling.
What changed after launch
Answers from the whole manual
Questions are answered from text, tables and diagrams together, so a specification table or an exploded view is part of the answer instead of something the technician has to find on their own.
Every claim traceable to a page
Each answer lists the document and page it came from, and the PDF opens at that page, so a technician can check the source before acting on it.
One machine, one context
Scanning the QR code on a machine opens a chat that only searches that machine's documents, inside its own workspace, so answers never mix one machine's manual with another's.
Large manuals without blocking the app
Heavy PDFs are processed by background workers on separate queues, with progress shown in the interface, and jobs interrupted by a restart are found and recovered automatically.
Usage visible per customer
Token, embedding and storage usage is recorded per workspace with its cost, and PDF billing reports come straight from that data.
Manual Minds AI turns a shelf of PDF manuals into an assistant that answers from text, tables and diagrams, cites the page it used and puts safety first. Each machine gets its own knowledge base behind a QR code, each customer gets an isolated workspace, and the platform tracks what every workspace uses.
More case studies
Get AI live in months, not years.
Book a 30-minute strategy call with the founders who will write your code.
