Entity Resolution

The Deduplication Challenge

When extracting entities from multiple documents, the same real-world entity often appears under different names:

Mr. Edwards
Bradley Edwards
Detective Edwards
Brad Edwards
edwards

Acme Corp
Acme Corporation
ACME Corp.
Acme

Transformer architecture
Transformer
Transformers
the Transformer model

Without deduplication, your knowledge graph contains duplicate nodes for the same entity, fragmenting relationships and making analysis difficult. sift-kg solves this with a 4-layer deduplication approach that combines deterministic rules, semantic similarity, LLM reasoning, and human review.

Layer 1: Automatic Pre-Deduplication

Before entities become graph nodes, sift-kg runs automatic deduplication during sift build to catch obvious duplicates.

Phase 1: Deterministic Merging

Entities are normalized and grouped by: Unicode Normalization

"café" → "cafe"
"naïve" → "naive"

Singularization

"researchers" → "researcher"
"companies" → "company"
"analyses" → "analysis"

Title Stripping Common title prefixes are removed:

"Dr. Alice Smith" → "Alice Smith"
"Detective Joe Recarey" → "Joe Recarey"
"Mr. Edwards" → "Edwards"
"Professor Zhang" → "Zhang"

Full list of recognized titles:

[
  "detective", "det.", "officer", "sergeant", "sgt.",
  "lieutenant", "lt.", "captain", "chief", "deputy",
  "dr.", "doctor", "prof.", "professor",
  "mr.", "mrs.", "ms.", "miss",
  "judge", "justice", "senator", "representative",
  "attorney", "counsel", "reverend", "father",
  # ... and more
]

See /home/daytona/workspace/source/src/sift_kg/graph/prededup.py:30 for the complete list. Canonical Selection When multiple variants normalize to the same form, the canonical is chosen by:

Frequency: Most common variant wins
Length: Longest variant (likely more complete)
Alphabetical: First alphabetically (tiebreaker)

Example:

Variants: ["Mr. Edwards", "Bradley Edwards", "Bradley Edwards", "edwards"]
Frequencies: {"Bradley Edwards": 2, "Mr. Edwards": 1, "edwards": 1}
Canonical: "Bradley Edwards"  # Highest frequency

Implementation at /home/daytona/workspace/source/src/sift_kg/graph/prededup.py:177:

def _pick_canonical(names: list[str]) -> str:
    counts = Counter(names)
    max_count = max(counts.values())
    most_frequent = [n for n, c in counts.items() if c == max_count]
    
    if len(most_frequent) == 1:
        return most_frequent[0]
    
    # Tiebreak: longest name
    max_len = max(len(n) for n in most_frequent)
    longest = [n for n in most_frequent if len(n) == max_len]
    
    return sorted(longest)[0]  # Alphabetical tiebreaker

Phase 2: Fuzzy Semantic Matching

After deterministic grouping, remaining unique forms are compared using SemHash semantic similarity. How SemHash Works

Convert entity names to embeddings using Model2Vec (lightweight, no GPU needed)
Compute pairwise cosine similarities
Merge entities with similarity ≥ 0.95 threshold

What It Catches

# Typos and misspellings
"Recarey" ↔ "Recary" (similarity: 0.96)

# Abbreviations
"International Business Machines" ↔ "IBM" (similarity: 0.97)

# Transliteration variants
"Muammar Gaddafi" ↔ "Muammar Qaddafi" (similarity: 0.98)

Implementation From /home/daytona/workspace/source/src/sift_kg/graph/prededup.py:153:

def _semhash_cluster(
    normalized_names: list[str],
    norm_to_canonical: dict[str, str],
    threshold: float,
) -> dict[str, str]:
    """Use SemHash to find fuzzy near-duplicates.
    
    Returns mapping from variant → canonical normalized form.
    """
    records = [{"text": name} for name in normalized_names]
    sh = SemHash.from_records(records=records, columns=["text"])
    result = sh.self_deduplicate(threshold=threshold)
    
    merges: dict[str, str] = {}
    for item in result.selected_with_duplicates:
        kept_name = item.record["text"]
        for dup_record, _score in item.duplicates:
            dup_name = dup_record["text"]
            if dup_name != kept_name:
                merges[dup_name] = kept_name
    
    return merges

Results

Pre-dedup typically reduces entity counts by 10-30%:

Pre-dedup: 1,247 entities → 912 unique (335 merged)

Saved entities don’t waste LLM tokens in the next resolution stage.

Layer 2: LLM-Based Entity Resolution

After pre-dedup, sift resolve finds duplicates that require semantic understanding.

Why LLM Resolution?

Pre-dedup catches mechanical duplicates. LLM resolution catches:

Partial names: “Joseph Recarey” vs “Joe Recarey”
Nicknames: “William” vs “Bill”
Context-dependent equivalence: “the Transformer architecture” vs “Transformers”
Professional vs personal names: “Dr. Alice Smith” vs “Alice”

Batching Strategy

Entities are processed in type-specific batches to improve accuracy. Sorting Entities are sorted before batching:

PERSON entities: Sorted by surname so name variants cluster

"Detective Joe Recarey" → sort key: "recarey detective joe recarey"
"Joseph Recarey" → sort key: "recarey joseph recarey"
# These appear adjacent in sorted list

Other types: Alphabetically by name

Implementation at /home/daytona/workspace/source/src/sift_kg/resolve/resolver.py:40:

def _person_sort_key(name: str) -> str:
    """Sort PERSON entities by surname.
    
    'Mr. Edwards', 'Bradley Edwards', 'Edwards' all sort under 'edwards'.
    """
    normalized = unidecode(name).lower().strip()
    
    # Strip title prefixes
    changed = True
    while changed:
        changed = False
        for prefix in _TITLE_PREFIXES:
            if normalized.startswith(prefix + " "):
                normalized = normalized[len(prefix) + 1:].strip()
                changed = True
                break
    
    # Sort by last word (surname), then full name
    parts = normalized.split()
    surname = parts[-1] if parts else normalized
    return f"{surname} {normalized}"

Overlapping Windows Large entity sets are split into batches with overlap:

Batch size: 100 entities max
Overlap: 20 entities between consecutive batches

This prevents duplicates from being missed at batch boundaries.

Batch 1: entities[0:100]
Batch 2: entities[80:180]  # 20-entity overlap with batch 1
Batch 3: entities[160:260] # 20-entity overlap with batch 2

LLM Resolution Prompt

For each batch, the LLM receives entity data:

[
  {
    "id": "person:joe_recarey",
    "name": "Detective Joe Recarey",
    "aliases": ["Joe Recarey", "Joseph Recarey"]
  },
  {
    "id": "person:joseph_recarey",
    "name": "Joseph Recarey",
    "aliases": []
  },
  {
    "id": "person:bradley_edwards",
    "name": "Bradley Edwards",
    "aliases": ["Brad Edwards"]
  }
]

The prompt asks the LLM to identify:

Duplicates: Same entity with different names → merge proposals
Variants: Parent-child relationships → EXTENDS relations

From /home/daytona/workspace/source/src/sift_kg/resolve/resolver.py:393:

Analyze these PERSON entities and identify:
1. Duplicates — entities that refer to the exact same thing (merge them)
2. Variants — entities that are a subtype, version, or specific implementation
   of a parent entity (link them with EXTENDS)

Look for:
- Name variations (abbreviations, nicknames, full vs common names, misspellings)
- Title/honorific prefixes that don't change identity (Dr., Mr., Detective, etc.)
- First name vs nickname variants
- Aliases — if an entity's aliases list contains a name matching another entity,
  they are very likely the same
- Same person referenced differently across documents
- DO NOT merge genuinely different people (father and son, unrelated people
  sharing a surname)

IMPORTANT: If entity B is a variant/subtype/version of entity A (not the same
thing, but derived from it), put it in "variants" NOT "groups". Only true
duplicates go in "groups".

Return valid JSON only:
{
  "groups": [
    {
      "canonical_id": "id of the best/most complete entity",
      "canonical_name": "the preferred name",
      "member_ids": ["id1", "id2"],
      "confidence": 0.0-1.0,
      "reason": "brief explanation"
    }
  ],
  "variants": [
    {
      "parent_id": "id of the parent/base entity",
      "child_id": "id of the variant/subtype",
      "confidence": 0.0-1.0,
      "reason": "brief explanation"
    }
  ]
}

Domain Context

When a domain provides system_context, it’s prepended to the prompt:

system_context: |
  You are analyzing academic papers to map the intellectual landscape.
  Distinguish carefully:
  - "Transformer" as an architecture is a THEORY
  - "GPT-2" the trained model is a SYSTEM
  These should NOT be merged — GPT-2 EXTENDS Transformer

This helps the LLM make domain-appropriate decisions.

Cross-Type Deduplication

After per-type resolution, sift-kg finds entities with identical names but different types:

CONCEPT: "reading comprehension"
PHENOMENON: "reading comprehension"

These are merged automatically (no LLM call needed):

Canonical type: The one with more connections (higher degree)
Reason: “Same name across types (CONCEPT vs PHENOMENON). Relations will be combined.”

From /home/daytona/workspace/source/src/sift_kg/resolve/resolver.py:190:

def _find_cross_type_duplicates(kg: KnowledgeGraph) -> list[MergeProposal]:
    # Group by normalized name
    name_groups: dict[str, list[tuple[str, str, int]]] = defaultdict(list)
    for nid, data in kg.graph.nodes(data=True):
        entity_type = data.get("entity_type", "")
        if entity_type in SKIP_TYPES:
            continue
        name = data.get("name", "").strip().lower()
        degree = kg.graph.degree(nid)
        name_groups[name].append((nid, entity_type, degree))
    
    proposals = []
    for _name, group in name_groups.items():
        types = {t for _, t, _ in group}
        if len(types) < 2:  # Only one type, no cross-type dup
            continue
        
        # Canonical = highest degree
        group.sort(key=lambda x: x[2], reverse=True)
        canonical_id, canonical_type, _ = group[0]
        # Create merge proposal...

Output

sift resolve produces: merge_proposals.yaml

proposals:
  - canonical_id: person:bradley_edwards
    canonical_name: Bradley Edwards
    entity_type: PERSON
    status: DRAFT
    members:
      - id: person:mr_edwards
        name: Mr. Edwards
        confidence: 0.92
      - id: person:detective_edwards
        name: Detective Edwards
        confidence: 0.88
    reason: "Same person with different titles"
  
  - canonical_id: theory:transformer
    canonical_name: Transformer
    entity_type: THEORY
    status: DRAFT
    members:
      - id: theory:transformer_architecture
        name: Transformer architecture
        confidence: 0.95
      - id: theory:transformers
        name: Transformers
        confidence: 0.90
    reason: "Same architecture with naming variations"

relation_review.yaml (updated with variants)

relations:
  - source_id: system:gpt_2
    source_name: GPT-2
    target_id: theory:transformer
    target_name: Transformer
    relation_type: EXTENDS
    confidence: 0.95
    status: DRAFT
    flag_reason: "Variant relationship discovered during entity resolution"
    evidence: "GPT-2 implements the Transformer architecture"

Layer 3: Human Review

LLMs make mistakes. Human review validates proposals before applying changes.

Interactive Review Process

The sift review command presents each proposal:

┌─ Merge 1/23 ─────────────────────────────────────┐
│ Merge into: Bradley Edwards (person:bradley_ed…) │
│ Type: PERSON                                      │
│                                                   │
│ Members to merge                                  │
│   Member              ID                 Confid…  │
│   Mr. Edwards         person:mr_edwards     92%   │
│   Detective Edwards   person:detective_e…   88%   │
│                                                   │
│ Reason: Same person with different titles         │
└───────────────────────────────────────────────────┘
  [a]pprove  [r]eject  [s]kip  [q]uit →

User decisions:

Approve: Status changes to CONFIRMED, will be applied
Reject: Status changes to REJECTED, ignored
Skip: Stays DRAFT, can review later
Quit: Saves progress, remaining proposals stay DRAFT

Implementation at /home/daytona/workspace/source/src/sift_kg/resolve/reviewer.py:39.

Auto-Approval

High-confidence proposals can be auto-confirmed:

sift review --auto-approve 0.85

All proposals where every member has confidence ≥ 0.85 are automatically confirmed without interactive review.

Auto-approved 15 proposals (all members ≥ 85% confidence)

Remaining proposals are reviewed interactively.

Relation Review

Flagged relations are also reviewed:

┌─ Relation 1/12 ──────────────────────────────────┐
│ Alice Smith  —[BENEFICIAL_OWNER_OF]→  Acme Corp │
│                                                   │
│ Low confidence │ confidence: 62% │ from: doc.pdf │
└───────────────────────────────────────────────────┘
  Evidence: Alice listed as beneficial owner in filing
  [a]pprove  [r]eject  [s]kip  [q]uit →

Auto-reject low-confidence relations:

sift review --auto-reject 0.5

Relations with confidence < 0.5 are automatically rejected.

Review Files After Changes

After review, files are updated in place: merge_proposals.yaml

proposals:
  - canonical_id: person:bradley_edwards
    status: CONFIRMED  # Changed from DRAFT
    # ... rest unchanged
  
  - canonical_id: theory:transformer
    status: REJECTED  # User rejected this merge
    # ...

You can also manually edit these files before running sift apply-merges.

Layer 4: Apply Merges

The final stage executes confirmed changes to the graph.

Merge Operation

For each CONFIRMED proposal, sift-kg: 1. Merge Node Data Canonical entity accumulates data from merged members:

# Combine source documents
canonical_docs = canonical.get("source_documents", [])
member_docs = member.get("source_documents", [])
for doc in member_docs:
    if doc not in canonical_docs:
        canonical_docs.append(doc)

# Keep higher confidence
if member.get("confidence") > canonical.get("confidence"):
    canonical["confidence"] = member["confidence"]

# Merge attributes (canonical takes precedence)
for key, value in member.get("attributes", {}).items():
    if key not in canonical_attrs:
        canonical_attrs[key] = value

# Track member names as aliases
member_name = member.get("name")
if member_name not in aliases:
    aliases.append(member_name)

From /home/daytona/workspace/source/src/sift_kg/resolve/engine.py:95. 2. Rewrite Edges All relations pointing to/from merged members are redirected to the canonical:

for source, target, key, data in kg.graph.edges(data=True, keys=True):
    new_source = merge_map.get(source, source)
    new_target = merge_map.get(target, target)
    
    if new_source != source or new_target != target:
        kg.graph.remove_edge(source, target, key=key)
        
        # Skip self-loops
        if new_source == new_target:
            stats["self_loops_removed"] += 1
            continue
        
        kg.graph.add_edge(new_source, new_target, key=key, **data)

3. Remove Merged Nodes

for member_id in valid_map:
    if kg.graph.has_node(member_id):
        kg.graph.remove_node(member_id)
        stats["nodes_removed"] += 1

4. Remove Rejected Relations Relations marked REJECTED in relation_review.yaml are deleted:

rejection_keys: set[tuple[str, str, str]] = set()
for entry in review_file.rejected:
    rejection_keys.add((entry.source_id, entry.target_id, entry.relation_type))
    # Also handle symmetric
    rejection_keys.add((entry.target_id, entry.source_id, entry.relation_type))

for source, target, key, data in kg.graph.edges(data=True, keys=True):
    rel_type = data.get("relation_type", "")
    if (source, target, rel_type) in rejection_keys:
        kg.graph.remove_edge(source, target, key=key)
        removed += 1

From /home/daytona/workspace/source/src/sift_kg/resolve/engine.py:140.

Output Statistics

Applied 23 merges: 23 nodes removed, 3 self-loops dropped
Removed 8 rejected relations

Graph updated!
  Entities: 892 (was 915)
  Relations: 1,247 (was 1,258)

The updated graph is saved to graph_data.json.

Real-World Example

Let’s trace a person entity through all 4 layers:

Initial Extractions

Three documents extract variations of the same person:

[
  {"name": "Mr. Edwards", "type": "PERSON", "source": "doc1.pdf"},
  {"name": "Bradley Edwards", "type": "PERSON", "source": "doc2.pdf"},
  {"name": "Bradley Edwards", "type": "PERSON", "source": "doc3.pdf"},
  {"name": "Detective Edwards", "type": "PERSON", "source": "doc4.pdf"},
  {"name": "brad edwards", "type": "PERSON", "source": "doc5.pdf"}
]

Layer 1: Pre-Dedup

Normalization:

"Mr. Edwards" → "edwards" (title stripped, lowercased)
"Bradley Edwards" → "bradley edwards"
"Detective Edwards" → "edwards" (title stripped)
"brad edwards" → "brad edwards"

Grouping by normalized form:

{
  "bradley edwards": ["Bradley Edwards", "Bradley Edwards"],
  "edwards": ["Mr. Edwards", "Detective Edwards"],
  "brad edwards": ["brad edwards"]
}

Canonical selection:

Group "bradley edwards": "Bradley Edwards" (frequency 2)
Group "edwards": "Mr. Edwards" (arbitrary, both frequency 1)
Group "brad edwards": "brad edwards" (only variant)

SemHash clustering:

"bradley edwards" ↔ "brad edwards" (similarity: 0.97) → MERGE
"edwards" ↔ "bradley edwards" (similarity: 0.89) → NO MERGE (below 0.95)

Result after pre-dedup:

3 unique entities:
  - person:bradley_edwards (from "Bradley Edwards", "brad edwards")
  - person:mr_edwards (from "Mr. Edwards", "Detective Edwards")

Note: “edwards” and “bradley edwards” weren’t merged because semantic similarity was below the 0.95 threshold. This is where LLM resolution helps.

Layer 2: LLM Resolution

Entities sent to LLM:

[
  {
    "id": "person:bradley_edwards",
    "name": "Bradley Edwards",
    "aliases": ["brad edwards"]
  },
  {
    "id": "person:mr_edwards",
    "name": "Mr. Edwards",
    "aliases": ["Detective Edwards"]
  }
]

LLM response:

{
  "groups": [
    {
      "canonical_id": "person:bradley_edwards",
      "canonical_name": "Bradley Edwards",
      "member_ids": ["person:bradley_edwards", "person:mr_edwards"],
      "confidence": 0.92,
      "reason": "Same person. 'Mr. Edwards' and 'Detective Edwards' are title variations of Bradley Edwards"
    }
  ],
  "variants": []
}

Merge proposal created:

canonical_id: person:bradley_edwards
canonical_name: Bradley Edwards
entity_type: PERSON
status: DRAFT
members:
  - id: person:mr_edwards
    name: Mr. Edwards
    confidence: 0.92
reason: "Same person. 'Mr. Edwards' and 'Detective Edwards' are title variations"

Layer 3: Human Review

┌─ Merge 1/1 ──────────────────────────────────────┐
│ Merge into: Bradley Edwards                      │
│ Type: PERSON                                      │
│                                                   │
│ Members to merge                                  │
│   Member         ID                    Confidence │
│   Mr. Edwards    person:mr_edwards          92%   │
│                                                   │
│ Reason: Same person with title variations         │
└───────────────────────────────────────────────────┘
  [a]pprove  [r]eject  [s]kip  [q]uit → a

User approves. Status changes to CONFIRMED.

Layer 4: Apply Merge

Before merge:

Nodes:
  - person:bradley_edwards
    name: "Bradley Edwards"
    source_documents: ["doc2.pdf", "doc3.pdf", "doc5.pdf"]
    aliases: ["brad edwards"]
  
  - person:mr_edwards
    name: "Mr. Edwards"
    source_documents: ["doc1.pdf", "doc4.pdf"]
    aliases: ["Detective Edwards"]

Edges:
  - person:bradley_edwards → org:acme (EMPLOYED_BY)
  - person:mr_edwards → org:acme (EMPLOYED_BY)

After merge:

Nodes:
  - person:bradley_edwards
    name: "Bradley Edwards"
    source_documents: ["doc1.pdf", "doc2.pdf", "doc3.pdf", "doc4.pdf", "doc5.pdf"]
    aliases: ["brad edwards", "Mr. Edwards", "Detective Edwards"]

Edges:
  - person:bradley_edwards → org:acme (EMPLOYED_BY) [self-loop removed]

The duplicate EMPLOYED_BY relation becomes a self-loop and is dropped. Final result: One consolidated entity with complete provenance and all name variations tracked.

Advanced Features

Semantic Clustering

For very large entity sets (1000+), use embedding-based clustering instead of alphabetical batching:

sift resolve --embeddings

Requires installing the embeddings extra:

pip install sift-kg[embeddings]

How it works:

Generate embeddings for all entity names
Cluster by cosine similarity
Each cluster becomes a batch sent to the LLM

Benefits:

Semantically similar entities are batched together
Reduces false negatives (missed duplicates)
Better for multilingual entities

Tradeoff: Slower and requires more memory.

Domain-Specific Resolution Context

Provide domain context to help the LLM make better decisions:

name: "Academic Research"
system_context: |
  When resolving entities:
  
  - "Transformer" (architecture) vs "BERT" (model): DO NOT MERGE
    BERT EXTENDS Transformer, they are related but distinct
  
  - "ResNet" vs "Residual Networks": MERGE (same thing)
  
  - "John Smith" (researcher) vs "J. Smith" (author): LIKELY SAME
    unless there's evidence of different affiliations

Reviewing Previous Decisions

If you want to re-review previously confirmed/rejected proposals:

Edit merge_proposals.yaml and change status back to DRAFT:

proposals:
  - canonical_id: person:alice
    status: DRAFT  # Was CONFIRMED, changing to re-review
    # ...

Run sift review again:
sift review

Manual Merge Proposals

You can manually add merge proposals to merge_proposals.yaml:

proposals:
  - canonical_id: org:acme_corp
    canonical_name: Acme Corporation
    entity_type: ORGANIZATION
    status: CONFIRMED  # Or DRAFT to review first
    members:
      - id: org:acme
        name: Acme
        confidence: 1.0
      - id: org:acme_inc
        name: Acme Inc.
        confidence: 1.0
    reason: "Manual merge — same company"

Then run sift apply-merges to execute.

Performance Considerations

Cost Estimation

Entity resolution costs scale with entity count:

# Rough estimate
entities_per_type = 500
batch_size = 100
batches_per_type = math.ceil(entities_per_type / (batch_size - 20))  # Account for overlap
total_batches = batches_per_type * num_entity_types

# Cost per batch: ~$0.01 with gpt-4o-mini
total_cost = total_batches * 0.01

For 1,000 entities across 10 types with gpt-4o-mini:

entities/type ÷ 80 effective batch size = 7 batches/type
batches/type × 10 types = 70 batches
batches × $0.01/batch = $0.70 total

Speed Optimization

Use concurrency to parallelize LLM calls:

sift resolve --concurrency 8  # Process 8 batches simultaneously

Default is 4. Higher values speed up resolution but may hit rate limits.

Skipping Resolution

If you’re confident in your extractions or have very clean data:

sift extract ./docs
sift build
# Skip resolve entirely
sift narrate

You can always run sift resolve later if you discover duplicates.

Best Practices

1. Review High-Impact Merges First

Sort proposals by degree (connection count) to focus on central entities:

# In merge_proposals.yaml, manually reorder or
# load into a script and sort
proposals.sort(key=lambda p: degree[p.canonical_id], reverse=True)

Merging highly-connected entities has bigger impact on graph structure.

2. Use Auto-Approval Conservatively

Start with a high threshold:

sift review --auto-approve 0.90  # Very conservative

Review the auto-approved proposals in merge_proposals.yaml. If quality is good, lower the threshold:

sift review --auto-approve 0.85

3. Iterate on Domain Context

If the LLM makes systematic mistakes, add guidance to system_context:

system_context: |
  Common mistakes to avoid:
  - "University of X" and "X University" are the SAME (merge them)
  - "John Smith" and "John Smith Jr." are DIFFERENT (father and son)
  - "ACL" (conference) and "ACL" (association) are DIFFERENT entities

4. Combine with Pre-Dedup Tuning

If pre-dedup misses obvious duplicates, you can adjust the SemHash threshold in code:

from sift_kg.graph.prededup import prededup_entities

# Lower threshold = more aggressive merging
canonical_map = prededup_entities(extractions, similarity_threshold=0.93)

Default is 0.95 (conservative).

Next Steps

How It Works

Understand the full pipeline from extraction to visualization

Domains

Learn about bundled domains and creating custom schemas

​The Deduplication Challenge

​Layer 1: Automatic Pre-Deduplication

​Phase 1: Deterministic Merging

​Phase 2: Fuzzy Semantic Matching

​Results

​Layer 2: LLM-Based Entity Resolution

​Why LLM Resolution?

​Batching Strategy

​LLM Resolution Prompt

​Domain Context

​Cross-Type Deduplication

​Output

​Layer 3: Human Review

​Interactive Review Process

​Auto-Approval

​Relation Review

​Review Files After Changes

​Layer 4: Apply Merges

​Merge Operation

​Output Statistics

​Real-World Example

​Initial Extractions

​Layer 1: Pre-Dedup

​Layer 2: LLM Resolution

​Layer 3: Human Review

​Layer 4: Apply Merge

​Advanced Features

​Semantic Clustering

​Domain-Specific Resolution Context

​Reviewing Previous Decisions

​Manual Merge Proposals

​Performance Considerations

​Cost Estimation

​Speed Optimization

​Skipping Resolution

​Best Practices

​1. Review High-Impact Merges First

​2. Use Auto-Approval Conservatively

​3. Iterate on Domain Context

​4. Combine with Pre-Dedup Tuning

​Next Steps

How It Works

Domains

The Deduplication Challenge

Layer 1: Automatic Pre-Deduplication

Phase 1: Deterministic Merging

Phase 2: Fuzzy Semantic Matching

Results

Layer 2: LLM-Based Entity Resolution

Why LLM Resolution?

Batching Strategy

LLM Resolution Prompt

Domain Context

Cross-Type Deduplication

Output

Layer 3: Human Review

Interactive Review Process

Auto-Approval

Relation Review

Review Files After Changes

Layer 4: Apply Merges

Merge Operation

Output Statistics

Real-World Example

Initial Extractions

Layer 1: Pre-Dedup

Layer 2: LLM Resolution

Layer 3: Human Review

Layer 4: Apply Merge

Advanced Features

Semantic Clustering

Domain-Specific Resolution Context

Reviewing Previous Decisions

Manual Merge Proposals

Performance Considerations

Cost Estimation

Speed Optimization

Skipping Resolution

Best Practices

1. Review High-Impact Merges First

2. Use Auto-Approval Conservatively

3. Iterate on Domain Context

4. Combine with Pre-Dedup Tuning

Next Steps