AI Search Engine Development

Description

Our documentation site has 4,000 pages and a search box that returns results ordered by how many times a word appears. Our support tickets are full of questions answered on page two of our own docs. We want semantic search that understands the question, returns the passage rather than the page, and handles the versioning problem: we support four product versions and an answer from the wrong version is actively harmful. Version must be a first-class filter, defaulted from the reader's context where we know it. Latency budget is 300ms at p95. Our docs are a static site and we are not adding a heavyweight runtime to it.

• Build passage-level semantic retrieval across the documentation corpus
• Make product version a first-class filter, defaulted from reader context and always visible
• Meet a 300ms p95 latency budget and document the architecture that achieves it
• Keep the docs site static — the search service is external, with a lightweight client
• Instrument no-result and low-confidence queries so the docs team sees the gaps

• Semantic search service with passage-level results
• Version-aware filtering
• Latency report against the 300ms p95 target
• Query gap analytics for the documentation team