search embeddings and ann

Search, Embeddings, And ANN Frontier

Search Surface

OntoIndex search is exposed through the CLI, HTTP API, and MCP facade/super-function layers.

Main code areas:

Path	Responsibility
ontoindex/src/core/search/	Query classification, retrieval composition, semantic cache, frontier search, replay gates.
ontoindex/src/core/embeddings/	Chunking, embedding pipeline, ANN neighbors, HTTP embedding client, structural extraction.
ontoindex/src/mcp/local/backend-search.ts	Local MCP search backend.
ontoindex/src/mcp/super/explore.ts	Agent-friendly concept exploration.

Retrieval Modes

The MCP search facade supports:

• semantic
• cypher
• repomap

Semantic search can opt into structured output, skeletons, Markdown context, passive related facts, and retrieval policies such as symbol, process, route, and requirement neighborhoods.

Embeddings

The current self index reports zero embeddings in the repo listing. That means graph and lexical retrieval are available, but vector-backed lanes may degrade or skip unless embeddings are generated.

Embedding-related commands and APIs are available through:

• CLI analyze options
• /api/embed
• ontoindex/src/core/embeddings/embedding-pipeline.ts
• ontoindex/src/core/embeddings/ann-neighbor-store.ts

ANN Neighbor Graph

Recent architecture work added a semantic ANN neighbor frontier. The core idea is to store bounded nearest-neighbor links for embedded code chunks and use them as a fast semantic expansion lane.

The relevant code lives under:

• ontoindex/src/core/embeddings/ann-neighbor.ts
• ontoindex/src/core/embeddings/ann-neighbor-store.ts
• ontoindex/src/core/search/semantic-frontier-search.ts
• ontoindex/src/core/search/semantic-frontier-adapter.ts

Operational State

If semantic quality looks weak, first check:

node /opt/demodb/_workfolder/OntoIndex/ontoindex/dist/cli/index.js status

Then regenerate embeddings only when explicitly needed, because embedding work is heavier than graph-only indexing.