Project search

Experimental developer tooling. Aksara's project search indexes descriptions of models, routes, settings, migrations, retained queries, and playbooks. It is not an application-record search API, vector database, or authorization boundary.

Start with the local engine example to search a small controlled collection without a database, provider, or network call. For project discovery, build_full_index(app=None, *, include_models=True, include_routes=True, include_migrations=True, include_queries=True, include_settings=True, include_playbooks=True) builds a fresh in-memory index from available sources. Import your model modules first and supply your application for route discovery. Results depend on what those sources expose in the current process; this is not a complete repository or database crawl.

The engine's semantic mode means TF-IDF cosine similarity over words. It does not establish understanding of synonyms, intent, or code semantics. Hybrid mode combines keyword and TF-IDF scores; no benchmark establishes it as universally better for your corpus.

SearchIndex implements its own local scoring and does not select a provider from embedding_provider, embedding_model, or embedding_dimensions settings. The separate embedding-provider registry can be used directly by applications; registering a provider does not make the index use it. Do not infer persistence, remote embeddings, or service enablement from configuration names alone.

See Studio search for its process-local cache and access boundary. Keep indexed SQL, settings, and model details private. Search kind/tag filters are retrieval filters, not tenant authorization.