mlx-bun library API
    Preparing search index...
    • Full stage: extract the entities a chunk is about and PERSIST them.

      Each extracted surface is RESOLVED to a canonical entity through the variant resolver (squeeze/alias/token-subset; see resolve.ts) so casing/spacing/brand variants collapse onto ONE entity instead of minting a fresh row each — the dedup fix. The resolver is seeded from the store's existing entities/aliases AND the curated gold (so notable brand-only nicknames fold), and accumulates minted canonicals across chunks when the caller threads one in via opts.

      Upserts the resolved canonical entities row (article_stem NULL — no article exists yet; ROUTE/CREATE promote it later) so the chunk_entities FK is satisfied, records the raw surface as an alias of the canonical, and writes one chunk_entities row per distinct resolved entity. Idempotent: re-running on the same chunk INSERT-OR-IGNOREs the same rows.

      Returns the extracted list (also useful to callers that route immediately).

      Parameters

      • store: MemoryStore
      • chunkId: string
      • Optionalopts: {
            call?: ExtractCall;
            maxTokens?: number;
            policy?: string;
            resolver?: EntityResolver;
            root?: string;
        }

      Returns Promise<ExtractedEntity[]>