Skip to main content

AI Search (Phase 3)

:::caution Phase 3, Not yet implemented This page describes the planned AI Search architecture. No search infrastructure is deployed yet. Implementation begins in Phase 3. :::

How AI-powered documentation search will work: hybrid retrieval over code-derived content, two-tier query routing, and a chat widget in the docs UI.

Overview​

Phase 3 replaces the default Docusaurus search with a full hybrid search stack on Azure AI Search. Users will be able to type natural language questions and get direct answers grounded in code-derived content, not a ranked list of pages.

Hybrid search architecture​

Azure AI Search provides the retrieval layer. Each documentation page is indexed as a set of overlapping chunks (512 tokens, 10% overlap) stored with three signals: a BM25-indexed plain text field for keyword recall, a dense vector embedding (BGE-large-en-v1.5, 1024 dimensions) for semantic recall, and a semantic ranker score for reranking at query time. This outperforms pure keyword or pure vector search on technical documentation: product terms like "WebRunner" and "tRFC" benefit from exact keyword matching while intent-level queries benefit from vector similarity.

Two-tier query routing​

TierModelTriggerPurpose
GatekeeperClaude Haiku 4.5All queriesClassify query, extract key terms, reject off-topic
SpecialistClaude Sonnet 4.6Complex / multi-stepSynthesize multi-chunk answer, generate code examples

The gatekeeper runs in under 200ms and handles simple factual queries directly. The specialist handles multi-step questions, cross-page comparisons, and any query requiring code generation.

Chat widget integration​

The search UI will use the Algolia AskAI widget embedded in the Docusaurus navbar. The widget sends queries to an Azure Function that runs the two-tier routing and returns a streamed answer with source citations. The Algolia AskAI slot is already reserved in docusaurus.config.ts, Phase 3 wires the backing service into that slot without structural site changes.

Coming in Phase 3​

  • Azure AI Search index provisioning (Terraform module)
  • Indexing pipeline: MDX → chunks → embeddings → index
  • Azure Function for two-tier query routing
  • Algolia AskAI widget configuration
  • Cost estimate: per-query inference cost, index storage, embedding refresh on release