@tacoreai/knowledge-engine

Deterministic text indexing and retrieval algorithms. The package accepts plain values and records and returns plain values and records; it does not own runtime clients, authorization, persistence, storage, API registration, business data models, or business-facing error mapping.

Exports

• @tacoreai/knowledge-engine/core: text normalization, lexical term building, structure-aware chunking, extractive hierarchy generation, ranking, and reciprocal-rank fusion. It also provides plain-value helpers for stable evidence deduplication, page aggregation, lexical coverage-state calculation, and soft document-diversity reranking.
• @tacoreai/knowledge-engine/algorithm-contracts: stable algorithm and index format markers used to determine output compatibility.

Application code owns all API schemas and descriptors, access policy, business record mappings, persistence orchestration, audit behavior, and localized responses. Changing an application API therefore does not require a release of this package.

Coverage helpers only combine already-loaded plain records and calculate a receipt from an explicitly supplied scope. They do not discover that scope, fetch pages, authorize documents, persist cursors, or decide how an application continues a request. Lexical rankers preserve relevance order by default; document-diversity reranking is explicit and never drops a record merely because another record came from the same document.

Development

# Run deterministic unit and package-boundary tests.
npm test

# Inspect the publishable files without publishing.
npm pack --dry-run

The package is pure ESM and requires Node.js 22 or newer. Since version 0.2.0, it removes the former runtime, authorization, persistence-contract, and application-schema entry points. Consumers must keep those concerns in their own source tree.