Cephalon.Retrieval
เนื้อหานี้ยังไม่ได้แปลเป็นภาษาไทย แสดงเป็นภาษาอังกฤษแทน
Maturity:
M3· Ownership: mixed: application-managed + cephalon-managed · Family:agentics· See audit, matrix.
Cephalon.Retrieval is the baseline technology pack for knowledge retrieval workloads.
What it owns
Section titled “What it owns”- retrieval options
- module and registration entry points for the
KnowledgeRetrievaltechnology - knowledge collection descriptors, registries, and catalogs
- knowledge document provider contracts for module-owned source material
- one Cephalon-managed lexical indexing and query execution baseline
- index state implementation, freshness calculation, manual reindex execution behind the abstraction-level
IKnowledgeIndexer, bounded query execution behind the abstraction-levelIKnowledgeQueryEngine, opt-in background reindex scheduling, query counters, and runtime-surface contribution for introspection
Main surfaces
Section titled “Main surfaces”Configuration/RetrievalOptions.csModules/RetrievalModule.csRegistration/RetrievalEngineBuilderExtensions.csServices/KnowledgeCollectionDescriptor.csServices/KnowledgeCollectionRegistry.csServices/KnowledgeCatalog.csServices/IKnowledgeCollectionContributor.csServices/IKnowledgeCatalog.csServices/IKnowledgeDocumentProvider.csServices/KnowledgeDocument.csServices/KnowledgeBackgroundReindexHostedService.csServices/KnowledgeQueryEngine.csServices/RetrievalRuntimeSurfaceContributor.cs
Source structure
Section titled “Source structure”ConfigurationModulesRegistrationServices
How it fits
Section titled “How it fits”This pack provides the retrieval-side companion to Cephalon.Agentics. It models knowledge collections as a first-class selectable technology surface instead of embedding retrieval assumptions into the engine.
When the KnowledgeRetrieval technology is selected, the pack can now index documents from registered IKnowledgeDocumentProvider services into an in-process lexical index, manually reindex a collection through the abstraction-level IKnowledgeIndexer, opt into an in-process background reindex scheduler, and execute bounded lexical queries through the abstraction-level IKnowledgeQueryEngine. Runtime surfaces report indexingOwnership, queryOwnership, backgroundReindexingOwnership, backgroundReindexingScheduled, provider readiness, latest indexing outcome, document count, query count, freshness state, and background scheduler scope/timing metadata so operators can distinguish an indexed collection from one that is still awaiting-provider, awaiting-index, or intentionally not scheduled.
The read-side index-state contract, manual reindex command seam, and bounded query command seam now live in
Cephalon.Abstractions.Retrieval as IKnowledgeIndexCatalog, KnowledgeIndexState,
KnowledgeIndexingOutcomes, KnowledgeIndexFreshnessStates, IKnowledgeIndexer,
KnowledgeIndexingRequest, KnowledgeIndexingResult, IKnowledgeQueryEngine,
KnowledgeQueryRequest, KnowledgeQueryResult, and KnowledgeQueryMatch. Cephalon.Retrieval
remains the implementation owner for the managed lexical index and query loop, while
Cephalon.Engine and host adapters can expose /engine/knowledge-indexes,
/engine/knowledge-indexes/{collectionId}, POST /engine/knowledge-indexes/{collectionId}/queries,
POST /engine/knowledge-indexes/{collectionId}/reindex, and snapshot.KnowledgeIndexes without
taking a direct dependency on this implementation package.
Background scheduling is opt-in through RetrievalOptions.EnableBackgroundReindexing. When enabled with ingestion, Cephalon.Retrieval registers a generic-host IHostedService that can reindex all registered collections or the configured BackgroundReindexCollectionIds, optionally run once on startup, and repeat at BackgroundReindexIntervalSeconds. Scheduled runs use the same IKnowledgeIndexer path as manual reindexing and record safe scheduler metadata such as trigger, scheduler id, iteration id, collection scope, and interval.
This is intentionally a narrow managed proof. It is useful for low-ceremony documentation, runbook, module-owned knowledge search, host-agnostic bounded query execution, manual operator reindexing, and opt-in in-process freshness scheduling, but it is not yet a vector database, distributed index, durable search cluster, embedding pipeline, reranker, distributed scheduler coordinator, leader-election mechanism, or provider-specific semantic search adapter. Those should land as companion packs or later retrieval slices when they own those paths explicitly.