MemoryVaultSearchOptions

Defined in: src/lib/memoryVault/searchTool.ts:32 

Options for the vault search tool.

Properties

bm25AdmissionDivisor?

optional bm25AdmissionDivisor: number

Defined in: src/lib/memoryVault/searchTool.ts:72 

Divisor mapping BM25 scores to the admission floor (bm25 / divisor). Default 50.

ceWeight?

optional ceWeight: number

Defined in: src/lib/memoryVault/searchTool.ts:55 

Multiplicative cross-encoder blend weight. Default 0.1. Only used when rerank is true.

decompose?

optional decompose: "off" | "llm"

Defined in: src/lib/memoryVault/searchTool.ts:81 

LLM-based query decomposition for composite/abstract queries. When set, each query is classified + (if composite) decomposed into 3–5 facet sub-queries via gpt-5-mini, then ranked via rankComposite. Requires decomposeOptions (auth) when set to “llm”.

decomposeOptions?

optional decomposeOptions: object

Defined in: src/lib/memoryVault/searchTool.ts:83 

Auth + endpoint for the decomposition LLM call. Required when decompose=“llm”.

apiKey

apiKey: string

baseUrl?

optional baseUrl: string

model?

optional model: string

entityRanking?

optional entityRanking: string[]

Defined in: src/lib/memoryVault/searchTool.ts:94 

W5 graph lane — pre-built ranking of memory IDs by entity-overlap score with the query. RRF-fused alongside cosine + BM25. Build via rankByEntityOverlap or pass-through from recall() when RecallContext.entityCtx is available.

folderId?

optional folderId: string | null

Defined in: src/lib/memoryVault/searchTool.ts:40 

When provided, only search memories in this folder (null for unfiled)

limit?

optional limit: number

Defined in: src/lib/memoryVault/searchTool.ts:34 

Maximum number of results to return (default: 5)

minSimilarity?

optional minSimilarity: number

Defined in: src/lib/memoryVault/searchTool.ts:36 

Minimum similarity threshold below which results are discarded (default: 0.1)

mmr?

optional mmr: boolean

Defined in: src/lib/memoryVault/searchTool.ts:64 

Apply Maximal Marginal Relevance after the relevance pass. Default false. Only effective on the rerank (async) pipeline.

proofCountAlpha?

optional proofCountAlpha: number

Defined in: src/lib/memoryVault/searchTool.ts:70 

Proof-count log-boost scale (Hindsight α). Default 0.1.

recency?

optional recency: RecencyOptions

Defined in: src/lib/memoryVault/searchTool.ts:59 

Recency decay curve overrides (per-year decay slope, floor, no-date multiplier).

recencyAlpha?

optional recencyAlpha: number

Defined in: src/lib/memoryVault/searchTool.ts:57 

Recency boost slope applied in the fused ranker. Default 1.0.

rerank?

optional rerank: boolean

Defined in: src/lib/memoryVault/searchTool.ts:51 

Run the cross-encoder reranker on the top-N V2 candidates. Default false. When true, switches to the async pipeline (rankFusedVaultMemoriesAsync).

rerankTopN?

optional rerankTopN: number

Defined in: src/lib/memoryVault/searchTool.ts:53 

Number of CE rerank candidates. Default 30.

rrfK?

optional rrfK: number

Defined in: src/lib/memoryVault/searchTool.ts:74 

RRF smoothing constant for lane fusion. Default 60.

scopes?

optional scopes: string[]

Defined in: src/lib/memoryVault/searchTool.ts:38 

When provided, only search memories with these scopes

supersessionBoost?

optional supersessionBoost: number

Defined in: src/lib/memoryVault/searchTool.ts:66 

Supersession score-gap transfer factor. Default 0.8.

supersessionWindow?

optional supersessionWindow: number

Defined in: src/lib/memoryVault/searchTool.ts:68 

Hard cap on the supersession candidate window. Default 50.

temporalRanking?

optional temporalRanking: string[]

Defined in: src/lib/memoryVault/searchTool.ts:102 

W6 temporal lane — pre-built ranking of memory IDs whose event-time overlaps the resolved query window, ordered by overlap score (descending). RRF-fused alongside cosine + BM25 + graph. Build via getMemoriesByEventTimeOp + scoreEventTimeOverlap, or pass-through from recall() when the query has a temporal phrase.

useFusion?

optional useFusion: boolean

Defined in: src/lib/memoryVault/searchTool.ts:46 

Use the hybrid fusion ranker (cosine + BM25 + RRF + recency) instead of cosine-only. Default true — new W1 pipeline. Pass false to fall back to the legacy cosine-only ranker (e.g. for benchmark A/B comparison).