Edge AI · MI-EAI-05
On-Premise RAG Pipeline over Plant Documents
Local RAG indexes 500+ plant documents and answers with citations, e.g. how to reset drive fault F07 on line 3.
Edge AI · MI-EAI-05
Index every manual, SOP, drawing and incident report your plant owns into a local vector store, and let engineers ask questions in plain language — with citations — from a box that never sends a byte off site.
- ₹1,54,000pilot BOM · hardware only
- 6 – 10 daysto a live pilot
- Advancedintegration level
- 3validated providers
Why plant knowledge is locked in PDFs and people
The problem
The answer to 'how do we reset drive fault F07 on line 3' exists — in a PDF on a shared drive, in a scanned OEM manual, or in the head of the one engineer who is on leave. Public AI assistants could help, but drawings, SOPs and incident reports cannot go to a cloud API.
- Documents scattered across drives, email and paper binders
- Answers depend on who is on shift; onboarding takes months
- Cloud LLMs are off-limits for confidential drawings and process data
- Scanned manuals and drawings are not searchable at all
The solution
A compact GPU/NPU box on the plant LAN runs an open-weight LLM and a retrieval pipeline: documents are ingested with OCR, chunked and embedded into a local vector database. Engineers ask questions in a browser; the model answers from the plant's own documents and shows the source page. No internet connection is required.
- 500+ PDFs, DOCs and scanned drawings ingested with OCR
- Answers with citations to the exact page and document
- Role-based access so teams see only their document sets
- Runs fully on-premise; zero outbound traffic
Validated service providers
Validated by TwoElectrons and ranked by validation score: verified credentials, completed jobs on the platform and client ratings.
- Sample Provider S — Bengaluru · On-prem LLM & RAG · ★ 4.9 · 12 jobs
- Sample Provider T — Hyderabad · TinyML & edge inference · ★ 4.7 · 9 jobs
- Sample Provider U — Pune · Secure AI infrastructure · ★ 4.6 · 7 jobs
Provider listings are samples pending live registrations.
Pilot BOM and cost
Reference hardware to run the pilot. Unit costs are indicative Indian market prices for the component class.
| Qty | Item | Unit cost | Line |
|---|---|---|---|
| 1× | Edge AI compute box GPU/NPU class, 16–32 GB unified memory | ₹1,40,000 | ₹1,40,000 |
| 1× | NVMe SSD 2 TB document store + vector DB | ₹14,000 | ₹14,000 |
| 1× | Local embedding + vector DB stack e.g. nomic/bge embeddings with Qdrant or Chroma | Free tier / open source | — |
| 1× | Document ingestion & OCR pipeline for scanned manuals and drawings | Free tier / open source | — |
| 1× | RAG orchestration & LAN UI LangChain / LlamaIndex with citation display | Free tier / open source | — |
| Pilot BOM total | ₹1,54,000 |
POC BOM cost only — hardware for a pilot. Installation, integration and provider services are quoted separately. Request for Quote
How on-premise RAG works
- Collect. Point the ingestion pipeline at the document folders; OCR handles scans.
- Index. Chunk and embed into the local vector store on the edge box.
- Ask. Engineers query in a LAN browser; answers cite the source page.
- Tune. Review answer quality with the team; adjust chunking and document sets.
Pilot architecture
Plant documents (PDF · DOC · scanned drawings) → Ingestion + OCR (chunking · embeddings) → Local vector DB + LLM (Qdrant / Chroma · 7–8B model) → LAN chat UI (citations · RBAC · audit log)
Everything inside the dotted line runs on one edge AI box on the plant network. No data leaves the site; internet access is not required.
Business case: time to answer, onboarding, data security
Ranges are typical figures reported for this class of solution; your pilot establishes the numbers for your plant.
- 15–30 min → < 1 min time to find a procedure (typical for OEM manuals and SOP lookups)
- 0 bytes leave the plant (air-gap capable; no cloud API)
- 500+ documents in the pilot index (PDF, DOC and scanned drawings)
- 6–10 days to a working pilot (including OCR and quality review)
Who this is for: Plant engineering heads, Maintenance and reliability teams, IT / OT security leads, Pharma, chemicals and defence suppliers, Multi-OEM plants with large manual libraries, Plants onboarding new engineers.
Illustrative case study
Illustrative scenario · not a client reference
A specialty-chemicals plant with strict data controls in Maharashtra (illustrative)
Set-up. One edge AI box in the server room, 640 documents ingested including 120 scanned OEM manuals. Access limited to the maintenance and process teams; IT verified zero outbound traffic on the firewall.
What happened. Maintenance engineers used it 40–60 times a day within two weeks, mostly for fault codes and isolation procedures. The team reported that a drive fault that previously took a call to the OEM was resolved from the manual in minutes. The plant went on to add its incident-report archive.
- Pilot budget: ₹1,54,000 hardware
- Documents indexed: 640
- Daily queries (week 2): 40–60
FAQ
Is any data sent to OpenAI, Google or another cloud?
No. The model, the embeddings and the vector database all run on the edge box on your LAN. It works without an internet connection.
How accurate are the answers?
The system answers only from your documents and shows the source page, so engineers can verify every answer. Quality is reviewed with your team during the pilot.
Can it read scanned manuals and drawings?
Yes. The ingestion pipeline includes OCR for scanned PDFs; drawings are indexed by their text and labels.
What size model runs on the box?
A quantised 7–8B parameter open-weight model, which is sufficient for document Q&A on a compact GPU/NPU box. Larger models can run on a bigger box in the scale-up phase.
What does the ₹1,54,000 cover?
Only the edge compute box and storage. The software stack is open source; ingestion, tuning, security hardening and provider services are quoted after a Request for Quote.
