Interface RetrievalPlan

Structured retrieval plan produced by the query classifier.

Replaces the simple 'none'|'simple'|'moderate'|'complex' strategy string with a granular specification of what retrieval sources to use, how to combine them, and what memory types to consult.

The plan is data — it describes WHAT to do, not HOW. The UnifiedRetriever interprets the plan and executes it.

Example

const plan: RetrievalPlan = {
  strategy: 'moderate',
  sources: { vector: true, bm25: true, graph: true, raptor: true, memory: true, multimodal: false },
  hyde: { enabled: true, hypothesisCount: 1 },
  memoryTypes: ['episodic', 'semantic'],
  modalities: ['text'],
  temporal: { preferRecent: false, recencyBoost: 1.0, maxAgeMs: null },
  graphConfig: { maxDepth: 2, minEdgeWeight: 0.3 },
  raptorLayers: [0, 1],
  deepResearch: false,
  confidence: 0.85,
  reasoning: 'Moderate complexity question about auth architecture.',
};

See

buildDefaultPlan for creating sensible defaults per strategy level

interface RetrievalPlan {
    strategy: RetrievalStrategy;
    sources: RetrievalPlanSources;
    hyde: {
        enabled: boolean;
        hypothesisCount: number;
    };
    memoryTypes: MemoryTypeFilter[];
    modalities: ModalityFilter[];
    temporal: TemporalConfig;
    graphConfig: GraphTraversalConfig;
    raptorLayers: number[];
    deepResearch: boolean;
    confidence: number;
    reasoning: string;
}

Properties

strategy sources hyde memoryTypes modalities temporal graphConfig raptorLayers deepResearch confidence reasoning

Properties

strategy

strategy: RetrievalStrategy

Base retrieval strategy (determines overall pipeline depth).

Maps to the existing strategy tier system but carries richer config.

  • 'none': Skip retrieval entirely.
  • 'simple': Vector + BM25 only. Fast path.
  • 'moderate': All sources, HyDE enabled, single pass.
  • 'complex': All sources, HyDE with multi-hypothesis, deep research, decomposition.

sources

sources: RetrievalPlanSources

Which retrieval sources to query.

Each flag independently enables or disables a retrieval source. The UnifiedRetriever skips sources whose dependencies are not available at runtime regardless of these flags.

hyde

hyde: {
    enabled: boolean;
    hypothesisCount: number;
}

HyDE (Hypothetical Document Embedding) configuration.

When enabled, a hypothetical answer is generated via LLM before embedding, bridging vocabulary gaps between questions and documents.

Type declaration

  • enabled: boolean

    Whether to use HyDE for this retrieval.

  • hypothesisCount: number

    Number of diverse hypotheses to generate. More hypotheses improve recall at the cost of additional LLM calls.

    Default

    1 for moderate, 3 for complex

See

HydeRetriever

memoryTypes

memoryTypes: MemoryTypeFilter[]

Which cognitive memory types to consult.

  • 'episodic': Past interactions, events, conversations.
  • 'semantic': Facts, knowledge, learned concepts.
  • 'procedural': Workflows, how-to knowledge, skills.
  • 'prospective': Upcoming intentions, reminders, planned actions.

modalities

modalities: ModalityFilter[]

Which content modalities to search in multimodal index.

Text search is always performed via vector/BM25 sources; non-text modalities are handled by the multimodal indexer.

temporal

temporal: TemporalConfig

Temporal preferences for result ordering and filtering.

Controls whether recent results are boosted and how aggressively older content is penalized.

graphConfig

graphConfig: GraphTraversalConfig

Graph traversal configuration for GraphRAG source.

Controls the depth and selectivity of entity relationship traversal starting from seed chunks discovered by vector/BM25 search.

raptorLayers

raptorLayers: number[]

Which RAPTOR tree layers to search.

  • 0: Leaf chunks (original documents) — detail queries.
  • 1: First-level cluster summaries — theme queries.
  • 2+: Higher-level summaries — "big picture" queries.

An empty array searches all layers (default RAPTOR behaviour).

deepResearch

deepResearch: boolean

Whether deep research mode is enabled.

Deep research decomposes the query into sub-queries and recurses with moderate-level plans per sub-query. Only used with complex strategy.

confidence

confidence: number

Confidence score from the classifier (0 to 1).

Indicates how certain the classifier is about this plan. Low confidence may trigger plan escalation in the router.

reasoning

reasoning: string

Human-readable reasoning from the classifier explaining why this plan was selected.

Settings

Member Visibility

Theme

On This Page

strategysourceshydememoryTypesmodalitiestemporalgraphConfigraptorLayersdeepResearchconfidencereasoning