@memberjunction/ai-reranker

Two-stage retrieval reranking service for MemberJunction's AI agent framework. Provides centralized reranker management and LLM-based reranking capabilities for improving the relevance of vector search results, particularly agent memory notes. Supports configurable reranking models, observability integration, and agent-level configuration.

Architecture

graph TD
    subgraph "@memberjunction/ai-reranker"
        RS["RerankerService
Singleton Manager"]
        style RS fill:#2d8659,stroke:#1a5c3a,color:#fff

        LR["LLMReranker
LLM-Based Reranking"]
        style LR fill:#7c5295,stroke:#563a6b,color:#fff

        CT["Config Types
RerankerConfiguration"]
        style CT fill:#b8762f,stroke:#8a5722,color:#fff
    end

    subgraph "Two-Stage Retrieval"
        S1["Stage 1: Vector Search
Fast approximate retrieval"]
        style S1 fill:#2d6a9f,stroke:#1a4971,color:#fff

        S2["Stage 2: LLM Reranking
Precise relevance scoring"]
        style S2 fill:#7c5295,stroke:#563a6b,color:#fff

        S3["Final Results
Reranked by relevance"]
        style S3 fill:#2d8659,stroke:#1a5c3a,color:#fff
    end

    S1 -->|"Top-K candidates"| RS
    RS --> S2
    S2 -->|"Scored & sorted"| S3

    subgraph Dependencies
        AI["@memberjunction/ai
BaseReranker"]
        style AI fill:#2d6a9f,stroke:#1a4971,color:#fff

        AIP["@memberjunction/ai-prompts
AIPromptRunner"]
        style AIP fill:#2d6a9f,stroke:#1a4971,color:#fff

        AIE["@memberjunction/aiengine
NoteMatchResult"]
        style AIE fill:#2d6a9f,stroke:#1a4971,color:#fff
    end

    AI --> RS
    AIP --> LR
    AIE --> RS
Copy

Installation

npm install @memberjunction/ai-reranker
Copy

Key Exports

RerankerService (Singleton)

Centralized service for managing reranker instances and performing note reranking. Caches reranker instances by model for efficient reuse.

import { RerankerService, parseRerankerConfiguration } from '@memberjunction/ai-reranker';

// Parse configuration from agent entity
const config = parseRerankerConfiguration(agent.RerankerConfiguration);

if (config?.enabled) {
    const result = await RerankerService.Instance.rerankNotes(
        vectorSearchResults,   // NoteMatchResult[] from vector search
        userQuery,             // The user's query for relevance scoring
        config,                // RerankerConfiguration
        contextUser
    );

    if (result.success) {
        console.log(`Reranked ${result.notes.length} notes in ${result.durationMs}ms`);
    }
}
Copy

LLMReranker

LLM-based reranking implementation that uses AI prompts to score document relevance. Extends BaseReranker from @memberjunction/ai.

import { createLLMReranker } from '@memberjunction/ai-reranker';

const reranker = createLLMReranker(modelId, contextUser);
Copy

DecisionReranker

Scores each document with one typed-decision Likelihood question ("This note bears on the request: "), asked through AIDecisionRunner with the query as the state. A document's score is the answer's probability. An agent opts in by pointing RerankerConfiguration.rerankerModelId at the Decision Reranker model.

• Runs: its decision runs are children of the rerank's MJ: AI Prompt Runs row, whose TotalCost and token rollups include them, so the agent run counts the rerank.
• Batching: a decision call carries at most the smallest Decision.MaxQuestionsPerCall its prompt's models declare, or, when none declares one, 20 documents (decisionMaxDocumentsPerCall). Larger reranks are split across parallel calls.
• Time budget: the calls share one budget, 15 seconds unless decisionTimeoutMS sets another. When it runs out the calls are aborted and the rerank fails, so the agent falls back as fallbackOnError says.
• Threshold: its probabilities are not calibrated, so set minRelevanceThreshold low; 0.1 is the recommended start.
• Failures: a failed call, or a missing or invalid answer for any document, fails the whole rerank. No score is invented.

RerankerConfiguration

Configuration type stored as JSON in the agent's RerankerConfiguration column:

interface RerankerConfiguration {
    /** Master switch for reranking */
    enabled: boolean;
    /** ID of the AI Model (type 'Reranker') to rerank with */
    rerankerModelId: string;
    /** Fetch N * retrievalMultiplier candidates, then rerank to the top N. Default: 3 */
    retrievalMultiplier: number;
    /**
     * Minimum relevance score (0-1) to keep a result. When none reaches it, the agent keeps the
     * vector search results instead. Default: 0.5; for a DecisionReranker, 0.1 is recommended.
     */
    minRelevanceThreshold: number;
    /** The prompt a prompt-backed reranker runs: LLMReranker's chat prompt, or DecisionReranker's decision prompt */
    rerankPromptID?: string;
    /** Extra note fields to include in each document's text */
    contextFields?: string[];
    /** On a failed rerank, use the vector search results (true) or throw (false). Default: true */
    fallbackOnError: boolean;
    /** Rerank examples too, with the same reranker and threshold. Default: false */
    rerankExamples?: boolean;
    /** DecisionReranker's time budget, in milliseconds. Default: 15000 */
    decisionTimeoutMS?: number;
    /** DecisionReranker's documents per call when no model declares a limit. Default: 20 */
    decisionMaxDocumentsPerCall?: number;
}
Copy

RerankServiceResult

interface RerankServiceResult {
    notes: NoteMatchResult[];     // Reranked notes sorted by relevance
    success: boolean;             // Whether reranking succeeded
    durationMs: number;           // Time taken in milliseconds
    usage?: {                     // Token usage for cost tracking
        promptTokens?: number;
        completionTokens?: number;
        cost?: number;
    };
    runStepID?: string;           // Created run step ID (if observability enabled)
}
Copy

RerankObservabilityOptions

Integration with agent run step tracking:

const result = await RerankerService.Instance.rerankNotes(
    notes,
    query,
    config,
    contextUser,
    {
        agentRunID: runId,        // Link to agent run
        parentStepID: stepId,     // Parent step for hierarchy
        stepNumber: 3,            // Step sequence number
        OnStepCreated: step => run.Steps.push(step) // Count the rerank in the agent run
    }
);
Copy

RerankExamples takes the same options and records a Rerank Examples step. The step's PromptRun is the rerank's run, so an agent that adds the step to its run's steps counts the rerank's cost and tokens toward MaxCostPerRun and MaxTokensPerRun; BaseAgent does this.

How Two-Stage Retrieval Works

1. Stage 1 -- Vector Search: Fast approximate nearest-neighbor search using embeddings to retrieve a broad set of candidate notes (e.g., top 50)
2. Stage 2 -- LLM Reranking: An LLM evaluates each candidate note against the query and assigns a precise relevance score
3. Result: The top-K most relevant notes are returned, sorted by LLM-determined relevance

This approach combines the speed of vector search with the accuracy of LLM-based understanding, significantly improving retrieval quality for agent memory.

Interaction with Consolidation

The candidate corpus the reranker scores is the in-process vector store (AIEngine._noteVectorService). Since v5.30.x, that store is kept in sync with persisted Status='Active' (see @memberjunction/aiengine README), which has two consequences for reranking:

• Consolidated source notes drop out immediately. When MemoryManagerAgent's consolidation pipeline merges a cluster, source notes are revoked in the same operation that removes their vector entries. They stop appearing as rerank candidates without an MJAPI restart.
• Protection tier is metadata, not a rerank gate. ProtectionTier (Immutable / Protected / Standard / Ephemeral) lives on the note entity and flows through to the reranker as part of each candidate's data. The reranker itself does not filter or weight by tier; downstream consumers (MemoryManagerAgent's decay phase, retrieval policies layered on top of RerankServiceResult) are responsible for any tier-aware logic.

For the consolidation pipeline mechanics and protection-tier semantics, see @memberjunction/ai-agents.

Dependencies

• @memberjunction/ai -- BaseReranker, RerankParams, RerankResult
• @memberjunction/ai-prompts -- AIPromptRunner for LLM-based reranking
• @memberjunction/ai-core-plus -- Extended entity classes
• @memberjunction/aiengine -- AIEngine, NoteMatchResult
• @memberjunction/core -- MJ framework core
• @memberjunction/core-entities -- Generated entity classes
• @memberjunction/global -- Class factory