ML System Design · Engineering Foundry challenge level: Advanced

Near-Duplicate Knowledge Base Detector

Design a retrieval and review system that finds semantically overlapping documents without collapsing legitimately distinct guidance.

All challengesSuggested practice time: 85 minutes

Original prompt

What you are building

Design a system that identifies likely duplicate or substantially overlapping knowledge-base articles across multiple business units. It should generate review candidates, explain why a pair was surfaced, and learn from reviewer decisions without automatically deleting content.

Engineering Foundry originalNo company association

Deliverables

Product decision, candidate definition, and human review path.
Evaluation set and ranking metrics.
Candidate generation, features, and model stages.
Indexing, serving, and permission enforcement.
Reviewer feedback and retraining design.
Monitoring and failure-mode analysis.

Constraints

The system may recommend review but cannot delete or merge content.
Unauthorized document content must not leak through candidates or explanations.
New and edited documents should become searchable within an hour.
Multilingual behavior must be measured rather than assumed.

Success criteria

Candidate generation and final ranking are separated.
Permission filtering is enforced before exposure.
Evaluation reflects reviewer utility and false-positive cost.
Index freshness, representation drift, and language performance are observable.

Suggested workflow

Define duplicate categories and reviewer action.
Build a judged evaluation set with hard negatives.
Choose candidate generators and ranking signals.
Design permission-aware indexing and explanations.
Trace feedback, freshness, and model migration.

Attempt before reveal

Build your deliverable independently.

Use the prompt, constraints, and workflow first. Guidance offers considerations and questions - not a canonical solution.

Reveal guidanceDirections and questions to test

One direction to explore

Combine high-recall lexical and embedding retrieval before a more expensive pairwise ranker.
Use section-level evidence to make reviewer explanations concrete.
Treat permissions as retrieval constraints, not a UI-only filter.

Reveal common mistakesFailure patterns worth checking

Using one similarity threshold as the entire system.
Evaluating only on easy exact duplicates.
Leaking titles or snippets across permission boundaries.
Treating reviewer disagreement as simple label noise.

Reveal stretch goalsOptional ways to extend the deliverable

Suggest a canonical article without making the merge decision.
Support a staged embedding-model index migration.

Optional solution worksheet

Prepare your approach for discussion.

Capture your reasoning, then copy a plain-text summary you control.

Session only - not saved.Refreshing or leaving this page clears every worksheet field and self-review choice. Nothing is uploaded or officially submitted.

Approach summary

Key decision

Important tradeoff

Failure mode or edge case

Validation or testing plan

Reflection

Optional solution URLThis link remains in the current browser session and is included only if you copy the summary.

Nothing is sent when you copy.

Optional community discussion

Compare reasoning, not rankings.

Discuss this challenge in the Engineering Foundry community if you choose. Share your own work, explain assumptions, critique ideas rather than people, and remove confidential or proprietary material.

Discuss this challenge

Loading Engineering Foundry…

ML System Design · Engineering Foundry challenge level: Advanced

Near-Duplicate Knowledge Base Detector

Design a retrieval and review system that finds semantically overlapping documents without collapsing legitimately distinct guidance.

All challengesSuggested practice time: 85 minutes

Original prompt

What you are building

Engineering Foundry originalNo company association

Deliverables

Product decision, candidate definition, and human review path.
Evaluation set and ranking metrics.
Candidate generation, features, and model stages.
Indexing, serving, and permission enforcement.
Reviewer feedback and retraining design.
Monitoring and failure-mode analysis.

Constraints

The system may recommend review but cannot delete or merge content.
Unauthorized document content must not leak through candidates or explanations.
New and edited documents should become searchable within an hour.
Multilingual behavior must be measured rather than assumed.

Success criteria

Candidate generation and final ranking are separated.
Permission filtering is enforced before exposure.
Evaluation reflects reviewer utility and false-positive cost.
Index freshness, representation drift, and language performance are observable.

Suggested workflow

Define duplicate categories and reviewer action.
Build a judged evaluation set with hard negatives.
Choose candidate generators and ranking signals.
Design permission-aware indexing and explanations.
Trace feedback, freshness, and model migration.

Attempt before reveal

Build your deliverable independently.

Use the prompt, constraints, and workflow first. Guidance offers considerations and questions - not a canonical solution.

Reveal guidanceDirections and questions to test

One direction to explore

Combine high-recall lexical and embedding retrieval before a more expensive pairwise ranker.
Use section-level evidence to make reviewer explanations concrete.
Treat permissions as retrieval constraints, not a UI-only filter.

Reveal common mistakesFailure patterns worth checking

Using one similarity threshold as the entire system.
Evaluating only on easy exact duplicates.
Leaking titles or snippets across permission boundaries.
Treating reviewer disagreement as simple label noise.

Reveal stretch goalsOptional ways to extend the deliverable

Suggest a canonical article without making the merge decision.
Support a staged embedding-model index migration.

Optional solution worksheet

Prepare your approach for discussion.

Capture your reasoning, then copy a plain-text summary you control.

Session only - not saved.Refreshing or leaving this page clears every worksheet field and self-review choice. Nothing is uploaded or officially submitted.

Approach summary

Key decision

Important tradeoff

Failure mode or edge case

Validation or testing plan

Reflection

Optional solution URLThis link remains in the current browser session and is included only if you copy the summary.

Nothing is sent when you copy.

Optional community discussion

Compare reasoning, not rankings.

Discuss this challenge

Near-Duplicate Knowledge Base Detector

What you are building

Deliverables

Constraints

Success criteria

Suggested workflow

Build your deliverable independently.

One direction to explore

ML System Design deliverable self-review

Prepare your approach for discussion.

Compare reasoning, not rankings.

Near-Duplicate Knowledge Base Detector

What you are building

Deliverables

Constraints

Success criteria

Suggested workflow

Build your deliverable independently.

One direction to explore

ML System Design deliverable self-review

Prepare your approach for discussion.

Compare reasoning, not rankings.