NCR Search Chatbot — ITER Organization

Overview

ITER is the world's largest fusion energy project. Over the years of manufacturing and construction it has accumulated roughly 11,000 Non-Conformance Reports (NCRs) — records of things that went wrong along the way.

Procedure says that if the same issue repeats more than twice, it must be escalated. The catch: finding those repeats depended almost entirely on people's memory. My internship project is about integrating AI into ITER's existing systems so engineers can search these reports by meaning, not keywords.

What I Built

Weaviate Prototype

I first prototyped the full pipeline on Weaviate:

• Ingestion of the NCR corpus
• Embedding generation
• Semantic search over the reports

To understand the speed vs accuracy trade-offs, I benchmarked HNSW parameters — ef, efConstruction and maxConnections — and measured how each affected recall and latency.

Azure AI Search Architecture

Together with my supervisor, I then redesigned the architecture on Azure AI Search:

• Hybrid retrieval — BM25 keyword search combined with vector search
• Semantic ranking on top of the hybrid results
• A 78-field index schema that I mapped from the enterprise database

Real-World Constraints

This is where the project stopped feeling like a tutorial:

• Confidential data — an on-premises requirement shaped what infrastructure was even allowed
• Missing long-text fields in the database extract forced a redesign of the ingestion pipeline
• ACL-based security trimming ensures users only ever see the reports they are authorised to access

What I Learned

Real company data is messy, and confidentiality rules shape the architecture more than model choice does. And honestly — explaining vector search to nuclear engineers taught me more than building it.

Links

Internal project (confidential data) — an architecture write-up is available on request.