All case studies
Knowledge Systems/Manufacturing

Thirty years of engineering knowledge, searchable in seconds

A permission-aware knowledge platform over 1.2 million design documents, test reports and quality records. Engineers ask questions in Japanese or English and get answers with citations to the exact page of the source.

questions asked per week
2,400
questions asked per week
answer faithfulness on the evaluation set
93%
answer faithfulness on the evaluation set
documents indexed across 4 systems
1.2M
documents indexed across 4 systems

The challenge

Design reviews, defect analyses and test reports from three decades were spread across file servers, a legacy document management system, SharePoint and the quality system. Much of the know-how lived with senior engineers approaching retirement.

Finding a comparable past design or an earlier defect countermeasure typically took half a day. Many documents were restricted by department or by customer contract, so any tool had to respect existing access rights exactly.

Our approach

  1. 01

    Permissions first

    Access rights are synced from the source systems and applied at retrieval time. If a user cannot open a document, the platform will not use it in their answers.

  2. 02

    Built for technical documents

    Engineering documents are full of tables, drawings and scanned pages in Japanese. We built parsing that preserves tables and section structure, instead of flattening everything to text.

  3. 03

    Answers you can verify

    Every claim links to its source passage. If the documents do not support an answer, the system states this explicitly.

What we built

Connectors

Incremental sync from file servers, SharePoint, the legacy DMS and the quality system, including their permission models.

Hybrid retrieval

Japanese-aware keyword search and semantic search combined with re-ranking, filtered by product line, plant and year.

Answer engine

Grounded generation with inline citations and a “not enough information” fallback.

Search app & Teams bot

A web app for deep research and a chat bot in Microsoft Teams for quick questions.

How it flows

  1. 4 source systems
  2. Parse & chunk
  3. Index + ACLs
  4. Hybrid retrieval
  5. Grounded answer
  6. Cited response

How we measured it

A set of 600 real questions, with reference answers written by senior engineers, is scored for retrieval recall, faithfulness and citation accuracy in both languages. A permission-leak test suite runs on every release.

Results

  • Around 2,400 questions a week from roughly 650 active users.
  • Finding a comparable past design went from half a day to a few minutes.
  • No permission leaks found in pre-launch red-teaming or in production audits.
  • Quality teams now regularly reuse past defect countermeasures in new design reviews.

Timeline

  1. Weeks 1–2

    Discovery

    Source inventory, permission model review, question collection

  2. Weeks 3–5

    Prototype

    Two source systems, retrieval baseline, first eval run

  3. Weeks 6–14

    Build

    All connectors, ACL sync, answer engine, apps

  4. Weeks 15–16

    Rollout

    Pilot with 2 plants, then company-wide

Contact

Have an AI product in mind?

Share what you would like to build, and we will help turn the idea into a production-ready product.

No technical specification required.

What happens next

  1. 1

    Share your idea

    Describe the problem, the users and the outcome you need.

  2. 2

    Discovery call

    We review your requirements together and clarify scope and constraints.

  3. 3

    Proposal

    You receive a clear scope, technical approach and delivery plan.