Deduplication Filter

Part of the prismAId Screening Tool.

Overview

The deduplication filter identifies and removes duplicate manuscripts from your dataset using intelligent field comparison and optional AI assistance.

Configuration

Basic Configuration

[filters.deduplication]
enabled = true
use_ai = false
compare_fields = ["title", "authors", "abstract", "doi"]

Configuration Parameters

Parameter	Type	Default	Description
`enabled`	boolean	false	Enable/disable the filter
`use_ai`	boolean	false	Use AI for semantic duplicate detection
`compare_fields`	array	[“title”, “abstract”]	Fields to compare for duplication

How It Works

Simple Matching (Non-AI)

When use_ai = false, the filter uses intelligent field comparison:

Priority Matching Rules:

DOI Match: If DOI fields exist and match exactly, records are considered duplicates
Combined Fields: Checks for author + year + (title OR abstract) combinations
Single Character Tolerance: Allows for minor variations (single character differences) in field comparisons
Text Normalization: Automatically handles case differences, extra whitespace, and punctuation variations

Best for: Fast processing when records have consistent metadata or minor variations

AI-Assisted Matching

When use_ai = true and LLM is configured, the filter uses semantic understanding:

AI Capabilities:

Recognizes author name variations (initials vs full names, middle names)
Handles character encoding issues (é→e, ü→u, Müller→Mueller)
Understands minor title/abstract rephrasing
Identifies duplicates despite formatting differences

AI Prompt Used: The AI compares manuscripts considering:

Author name formats (initials vs full names, middle names, order variations)
Character encoding issues (é→e, ü→u, ñ→n, ø→o, incorrect UTF-8 representation)
Non-standard character replacements (Müller→Mueller, Gómez→Gomez, Søren→Soren)
Technical simplifications in database entries
Minor transcription differences
Abbreviated vs full journal names
Different citation styles or formats
Minor typos or punctuation differences

Output Fields

The filter adds these fields to each manuscript record:

Field	Type	Description
`tag_is_duplicate`	boolean	`true` for duplicates, `false` for originals
`tag_duplicate_of`	string	ID of the original record (empty for non-duplicates)
`include`	boolean	Set to `false` for duplicates
`exclusion_reason`	string	“Duplicate of [ID]” for duplicates

Example Configurations

Basic Deduplication

[filters.deduplication]
enabled = true
use_ai = false
compare_fields = ["doi", "title"]

Comprehensive Deduplication

[filters.deduplication]
enabled = true
use_ai = false
compare_fields = ["title", "authors", "abstract", "doi", "year"]

AI-Enhanced Deduplication

[filters.deduplication]
enabled = true
use_ai = true
compare_fields = ["title", "authors", "abstract"]

[[filters.llm]]
provider = "OpenAI"
api_key = ""  # Uses environment variable
model = "gpt-4o-mini"
temperature = 0.01

Best Practices

Field Selection: Include multiple fields for better accuracy
DOI Priority: Always include DOI if available for exact matching
Author Fields: Include author names to catch same-title different-author papers
AI Usage: Use AI mode when dealing with:
- Multiple database sources with different formatting
- International datasets with character encoding variations
- Historical data with inconsistent metadata

Performance Considerations

Non-AI Mode: Very fast, processes thousands of records per second
AI Mode: Limited by API rate limits, typically 10-100 records per second
Memory Usage: Minimal, uses streaming processing

Filter Order

Deduplication is applied first in the screening pipeline to maximize efficiency by removing duplicates before other processing.