Knowledge

Overview

Multi-modal document ingest, object-store originals, hybrid retrieval, and authoritative knowledge on the unified Spectron substrate.

Knowledge workflows are how authoritative material enters Spectron: manuals, policies, product data, repositories, and warehouse exports. Bytes land in object storage where appropriate; structured index state (chunks, entities, embeddings, keyword graph, document links, traces) lives in SurrealDB under the same ACID write path as conversational memory.

This is the same unified substrate described in Unified substrate and authority: one graph and retrieval stack for documents and conversation, distinguished by source.kind = "document" (and operator upsert) plus higher default source.trust.

  • Ingest profiles trade completeness for cost (TextOnlyMultimodalFull).

  • Content addressing (for example Blake3 hashes) makes re-upload and rechunk idempotent.

  • Chunks are first-class records with embeddings, spans into originals, and links to extracted entities.

  • Extraction runs the same reconciler as conversational turns – document facts can contradict turn facts and surface uncertainty instead of silent merges.

  • Keyword graph (RAKE + PMI edges) and section embeddings / document links precompute structure the hybrid ranker reads cheaply at query time.

See Surface, models, and security and the retrieval pages below.

When you are ready to wire agents, continue to Memory for sessions and operations, or to Principles and goals for the full conceptual spine.

Was this page helpful?