@memberjunction/ai-vector-dupe

AI-powered duplicate record detection for MemberJunction entities -- finds, scores, tracks, and optionally auto-merges duplicate records using vector similarity, hybrid search (RRF), and optional reranking, with an optional LLM reasoning layer that judges borderline matches per-candidate before they ever reach a human.

Two-stage by design. Vector/hybrid search is the cheap, fast recall stage; the LLM reasoning layer is an opt-in, threshold-gated precision stage. Reasoning is off by default and, when enabled, only fires for matched sets whose top vector score clears a per-entity ReasoningThreshold -- so you pay for the LLM only on the matches that are actually ambiguous. See LLM Reasoning Layer.

Architecture

                         +--------------------------+
                         |   DuplicateRecordDetector |
                         |   (extends VectorBase)    |
                         +-----+----------+---------+
                               |          |
              +----------------+          +----------------+
              |                                            |
    +---------v----------+                     +-----------v---------+
    | GetDuplicateRecords|                     |  CheckSingleRecord  |
    | (list-based batch) |                     |  (single record)    |
    +--------+-----------+                     +-----------+---------+
             |                                             |
             +-------------------+-------------------------+
                                 |
                    +------------v------------+
                    |    Detection Pipeline   |
                    +-------------------------+
                    | 1. Validate Entity Doc  |
                    | 2. Vectorize records    |
                    | 3. Embed via AI model   |
                    | 4. Query vector DB      |
                    |    (hybrid if supported)|
                    | 5. Filter self-matches  |
                    | 6. Apply thresholds     |
                    | 7. LLM reasoning (opt-in,|
                    |    threshold-gated)     |
                    | 8. Persist match results|
                    | 9. Auto-merge (optional)|
                    +-------------------------+
                                 |
              +------------------+------------------+
              |                  |                   |
    +---------v------+  +-------v--------+  +-------v--------+
    | ai-vector-sync |  | ai-vectordb    |  | ai (Embeddings)|
    | (vectorizer,   |  | (VectorDBBase, |  | (BaseEmbeddings|
    |  templates)    |  |  hybrid query) |  |  GetAIAPIKey)  |
    +----------------+  +----------------+  +----------------+
Copy

Key dependencies:

The 'Agent'-mode provider (DuplicateReasoningAgentProvider) is not a dependency of this package -- it lives in @memberjunction/ai-agents and registers against the reasoning seam at runtime. See LLM Reasoning Layer.

Installation

npm install @memberjunction/ai-vector-dupe
Copy

Quick Start

List-Based Batch Detection

Detect duplicates across all records in an MJ List:

import { DuplicateRecordDetector } from '@memberjunction/ai-vector-dupe';
import { PotentialDuplicateRequest } from '@memberjunction/core';

const detector = new DuplicateRecordDetector();

const request: PotentialDuplicateRequest = {
    ListID: 'your-list-uuid',
    EntityID: 'your-entity-uuid',
    EntityDocumentID: 'your-entity-document-uuid',
    Options: {
        TopK: 10,
        OnProgress: (progress) => {
            console.log(`[${progress.Phase}] ${progress.ProcessedRecords}/${progress.TotalRecords} -- ${progress.MatchesFound} matches`);
        },
    },
};

const response = await detector.GetDuplicateRecords(request, contextUser);

if (response.Status === 'Success') {
    for (const result of response.PotentialDuplicateResult) {
        console.log(`Record: ${result.RecordCompositeKey.ToString()}`);
        for (const dupe of result.Duplicates) {
            console.log(`  Match: ${dupe.ToString()} (${(dupe.ProbabilityScore * 100).toFixed(1)}%)`);
        }
    }
}
Copy

Single-Record Check

Check one record for duplicates without creating a list -- ideal for server hooks (e.g., fire-and-forget after record save):

import { DuplicateRecordDetector } from '@memberjunction/ai-vector-dupe';
import { CompositeKey } from '@memberjunction/core';

const detector = new DuplicateRecordDetector();

const recordKey = new CompositeKey([{ FieldName: 'ID', Value: 'record-uuid' }]);

const result = await detector.CheckSingleRecord(
    'your-entity-document-uuid',
    recordKey,
    { TopK: 5 },
    contextUser
);

for (const dupe of result.Duplicates) {
    console.log(`Potential duplicate: ${dupe.ToString()} (score: ${dupe.ProbabilityScore})`);
}
Copy

DuplicateDetectionOptions Reference

Options are passed via the Options property on PotentialDuplicateRequest, or directly to CheckSingleRecord.

Thresholds

Thresholds can be configured at two levels -- on the EntityDocument record (default) or overridden per-run via DuplicateDetectionOptions. When threshold overrides are provided in the options, they take precedence over the EntityDocument values.

A server hook normalizes 1.0 thresholds to sensible defaults (0.70 for potential, 0.95 for absolute) to prevent degenerate behavior when thresholds are left at the maximum.

Hybrid Search and Reciprocal Rank Fusion (RRF)

When the configured vector database supports hybrid search (VectorDBBase.SupportsHybridSearch === true), the detector automatically combines vector similarity and keyword search for higher-quality results.

How It Works

1. The record's template text is sent as both a vector embedding and a keyword query.
2. The vector DB returns results from both retrieval methods.
3. Results are fused using Reciprocal Rank Fusion (RRF), a rank-based algorithm that is score-scale independent.

RRF Formula

FusedScore(d) = SUM_i [ 1 / (k + rank_i(d)) ]
Copy

Where rank_i(d) is the 1-based rank of document d in list i, and k is a smoothing constant (default: 60).

Using ComputeRRF Directly

The ComputeRRF utility is exported for use in custom pipelines:

import { ComputeRRF, ScoredCandidate } from '@memberjunction/ai-vector-dupe';

const vectorResults: ScoredCandidate[] = [
    { ID: 'rec-1', Score: 0.95 },
    { ID: 'rec-2', Score: 0.87 },
    { ID: 'rec-3', Score: 0.82 },
];

const keywordResults: ScoredCandidate[] = [
    { ID: 'rec-2', Score: 12.5 },  // Different scale -- RRF handles this
    { ID: 'rec-4', Score: 10.1 },
    { ID: 'rec-1', Score: 8.3 },
];

const fused = ComputeRRF([vectorResults, keywordResults], 60);
// Results sorted by fused RRF score, score-scale independent
Copy

Tuning Hybrid Search

• KeywordSearchWeight = 0.0: Pure vector similarity (semantic matching).
• KeywordSearchWeight = 0.3 (default): Slight keyword boost. Good for entities with distinctive names or codes.
• KeywordSearchWeight = 0.5: Equal weight. Useful when both semantic and lexical matches matter.
• KeywordSearchWeight = 1.0: Pure keyword search (not recommended for duplicate detection).

Reranking

When MJ's BaseReranker / RerankerService is configured, the detector can apply a second-stage reranking pass after initial retrieval. Reranking uses a cross-encoder model to re-score candidates with higher precision than embedding-based similarity alone.

Reranking is especially effective when:

• Initial retrieval returns many borderline candidates
• Entity records have complex, multi-field structures
• You need to maximize precision at the cost of slightly higher latency

See the Duplicate Detection Guide for configuration details.

LLM Reasoning Layer

Vector and hybrid search are good at recall (surfacing candidates) but not great at the final is-this-actually-the-same-entity? judgment -- two records can be semantically near without being duplicates (a parent company vs. its subsidiary), and two true duplicates can look lexically different (typos, abbreviations, stale addresses). The reasoning layer lets a (possibly small/cheap) LLM make that judgment per candidate, on top of the vector scores.

It is additive and off by default -- when EnableLLMReasoning is false on the Entity Document, the pipeline behaves exactly as the vector-only path documented above.

Cost control: the reasoning gate

The whole point is not to run an LLM on every record. Reasoning runs once per source record's matched set, and only when the gate is open. The gate (DuplicateRecordDetector.IsReasoningGateOpen) opens when:

1. EntityDocument.EnableLLMReasoning is true, and
2. the set has at least one candidate, and
3. the set's top vector MatchProbability is >= EntityDocument.ReasoningThreshold (a null threshold means "reason over any non-empty set").

So configuring ReasoningThreshold = 0.85 means "only spend an LLM call when vector search is already fairly confident" -- you tune the recall/precision/cost tradeoff per entity.

Configuration (on the Entity Document)

All reasoning configuration lives on the MJ: Entity Documents record, per entity:

The pluggable provider seam

Reasoning is delegated through an abstract DuplicateReasoningProvider, resolved at runtime via the MJ class factory by ReasoningMode. Two providers ship; both emit the identical DuplicateReasoningOutput, so promoting an entity from Prompt to Agent is a config change, not a rewrite.

The Agent provider lives in @memberjunction/ai-agents, not this package, because ai-agents depends on ai-vector-dupe -- importing AgentRunner here would create a build cycle. It registers against the seam under the 'Agent' key, so the detector resolves it via the class factory with no static import back into the pipeline. Registration is handled by the class-registration manifest (no Load*() helper needed).

To add a custom reasoning strategy, subclass DuplicateReasoningProvider, implement Reason(input, context), and register it under a new mode key:

import { RegisterClass } from '@memberjunction/global';
import {
    DuplicateReasoningProvider,
    DuplicateReasoningInput,
    DuplicateReasoningOutput,
    DuplicateReasoningContext,
} from '@memberjunction/ai-vector-dupe';

@RegisterClass(DuplicateReasoningProvider, 'MyMode')
export class MyReasoningProvider extends DuplicateReasoningProvider {
    public async Reason(
        input: DuplicateReasoningInput,
        context: DuplicateReasoningContext,
    ): Promise<DuplicateReasoningOutput> {
        // ...inspect input.SourceRecord, input.Candidates, input.FieldDeltas...
        // return a structured verdict with per-candidate CandidateVerdicts
    }
}
Copy

Per-candidate verdicts

A matched set is the top-K neighbors of one source record, so it routinely mixes true duplicates with false positives. The reasoner therefore returns a verdict per candidate (DuplicateReasoningOutput.CandidateVerdicts), each judged independently against the source -- a false-positive candidate reads NotDuplicate even when another candidate in the same set is a confident Merge. The detector stamps each candidate's own verdict onto its match row; the set-level Recommendation/Confidence are derived values used only for the group's dominant display and the auto-merge gate.

The output also carries a proposed SurvivorRecordID and per-field survivor choices (FieldChoices) that feed Metadata.MergeRecords at merge time. A null Confidence means the model returned no usable confidence and must be rendered/stored as "unknown" -- never conflated with a real 0 (which reads as "confidently NOT a duplicate"). See DuplicateReasoningTypes for the full contract.

Progress Reporting

The OnProgress callback fires at each phase of the pipeline:

const request: PotentialDuplicateRequest = {
    // ...
    Options: {
        OnProgress: (progress) => {
            const { Phase, TotalRecords, ProcessedRecords, MatchesFound, ElapsedMs } = progress;
            const pct = TotalRecords > 0 ? ((ProcessedRecords / TotalRecords) * 100).toFixed(0) : '0';
            console.log(`[${Phase}] ${pct}% -- ${MatchesFound} matches (${ElapsedMs}ms)`);
        },
    },
};
Copy

Progress Phases

DuplicateDetectionProgress Shape

interface DuplicateDetectionProgress {
    Phase: 'Vectorizing' | 'Embedding' | 'Querying' | 'Matching' | 'Merging';
    TotalRecords: number;
    ProcessedRecords: number;
    MatchesFound: number;
    CurrentRecordID?: string;
    ElapsedMs: number;
}
Copy

API Reference Summary

DuplicateRecordDetector

ComputeRRF

function ComputeRRF(rankedLists: ScoredCandidate[][], k?: number): ScoredCandidate[]
Copy

Compute Reciprocal Rank Fusion across multiple ranked result lists. Returns candidates sorted by descending fused score.

ScoredCandidate

interface ScoredCandidate {
    ID: string;
    Score: number;
    Metadata?: Record<string, unknown>;
}
Copy

Inverse Match Deduplication

The detector maintains a _seenPairs set across the entire run to suppress inverse duplicates. If record A is identified as a duplicate of record B (A->B), the reverse match (B->A) is automatically suppressed. Pair keys use canonical ordering (smallerID::largerID) for consistent deduplication regardless of query direction.

RecordID Format and Metadata

• RecordID and MatchRecordID are stored in MJ URL segment format (e.g., ID|uuid), making them compatible with CompositeKey for entities with composite primary keys.
• RecordMetadata is stored on both DuplicateRunDetail and DuplicateRunDetailMatch entities, capturing the vector database metadata snapshot at detection time. This preserves the context used for matching even if the source record changes later.

Database Entities

The package reads from and writes to these MJ entities:

Further Reading

• Duplicate Detection Guide -- comprehensive developer guide covering end-to-end workflow, threshold tuning, hybrid search deep dive, performance optimization, and troubleshooting
• MemberJunction AI Vectors -- base vector infrastructure
• AI Vector Sync -- entity vectorization and template parsing

Development

# Build
npm run build

# Run tests
npm run test

# Watch mode
npm run test:watch
Copy

License

ISC