Stop duplicating entities in your Neo4j Knowledge Graph. AutoGraft intercepts entities extracted by LangChain or LlamaIndex, uses a 3-layer hybrid approach (Deterministic -> Vector -> LLM) to merge duplicates, and generates clean Cypher queries.

Installation

pip install autograft

(To use integrations, install with pip install autograft[langchain] or pip install autograft[llamaindex])

Why AutoGraft?

• LLM-Agnostic: Works with OpenAI, Groq, Ollama, OpenRouter via litellm.
• Massive Cost Savings: Reduces Entity Resolution token costs by up to 100% by resolving locally.
• Plug & Play: Drop-in replacement before your Neo4j database.
• Blazing Fast: C/C++ (RapidFuzz) and NumPy local matching.

Performance Benchmark (660 Documents / 10 Industries)

Evaluated across a massive suite of 660 real-world enterprise documents spanning 10 key sectors: Legal, Tech, Insurance, Finance, Healthcare, Manufacturing, Retail, Energy, Education, and Real Estate (with complex acronyms like GDPR, K8s, AWS, D&O, EBITDA, KYC/AML, SOFR, SCOTUS).

LLM Engine Configured: groq/llama-3.1-8b-instant for extraction & arbitration, and groq/llama-3.3-70b-versatile for precision auditing.

Figure 1.1: Enterprise RAG Entity Resolution Performance Metrics (660 Docs / 10 Industries)

Detailed Metric Breakdown:

• Top-Left (Total Tokens Consumed): LangChain Naive and AutoGraft consume 0 resolution tokens, while Full LLM ER consumes 685,608 tokens.
• Top-Right (LLM ER API Calls): LangChain Naive and AutoGraft make 0 API calls, while Full LLM ER makes 2,448 external API calls.
• Bottom-Left (Neo4j Duplicates Avoided): LangChain Naive creates 620 duplicates (0 avoided), while Full LLM ER and AutoGraft resolve all 620 duplicates.
• Bottom-Right (Estimated LLM Cost): Compares the Entity Resolution financial cost. LangChain Naive costs $0 (but fails to deduplicate), Full LLM ER costs $0.13712, and AutoGraft costs $0 (while perfectly deduplicating the graph).

Figure 1.2: Enterprise Knowledge Graph Cost Scaling (Up to 1,000,000 Documents)

For 1,000,000 documents, AutoGraft maintains $0.00 LLM Entity Resolution API costs while guaranteeing a 100% clean, deduplicated Knowledge Graph.

Figure 1.3: Entity Resolution Precision by Industry Sector (100.0% Overall)

For complete evaluation methodology and dataset documentation, see BENCHMARK.md.

Configuration & Credentials

AutoGraft can be configured via environment variables (.env) or programmatically via the AutoGraftConfig class.

from autograft import AutoGraftConfig
from autograft.integrations import AutoGraftNeo4jMiddleware

config = AutoGraftConfig(
    model="openai/gpt-4o",
    api_key="sk-...",
    api_base="https://custom.endpoint/v1",  # Optional: for proxies, Azure, or local LLMs
    match_threshold=0.85,
    id_attr="id",
    aliases_attr="aliases",
    matching_algorithm="token_sort_ratio"
)
autograft_graph = AutoGraftNeo4jMiddleware(graph, config=config)

Architecture (3-Layer Short-Circuit)

1. Layer 1 (Deterministic): Exact string match & alias matching via rapidfuzz (0 tokens, 0.1ms).
2. Layer 2 (Semantic): Vector cosine similarity via numpy (0 tokens, 0.5ms).
3. Layer 3 (LLM Arbiter): LiteLLM call ONLY for residual ambiguous cases (e.g., "J. Dupont" vs "Jean Dupont").

Plug & Play Integrations

AutoGraft provides native 1-line integrations for both LangChain and LlamaIndex. By wrapping your Graph store, AutoGraft intercepts entities, deduplicates them locally at 0 cost, and safely forwards them to your database.

LangChain

from langchain_community.graphs import Neo4jGraph
from autograft.integrations import AutoGraftNeo4jMiddleware

# 1. Connect to your Neo4j Database
graph = Neo4jGraph(url="bolt://localhost:7687", username="neo4j", password="password")

# 2. Wrap it with AutoGraft (1 line of code!)
autograft_graph = AutoGraftNeo4jMiddleware(graph)

# 3. Add your extracted documents. AutoGraft will silently deduplicate everything locally!
autograft_graph.add_graph_documents(extracted_graph_documents)

LlamaIndex

from llama_index.graph_stores.neo4j import Neo4jPropertyGraphStore
from autograft.integrations import AutoGraftLlamaIndexMiddleware

# 1. Connect to Neo4j Graph Store
store = Neo4jPropertyGraphStore(username="neo4j", password="password", url="bolt://localhost:7687")

# 2. Wrap it with AutoGraft
autograft_store = AutoGraftLlamaIndexMiddleware(store)

# 3. Use in your LlamaIndex pipeline (upsert_nodes will be automatically deduplicated)
# e.g., index = PropertyGraphIndex.from_documents(documents, property_graph_store=autograft_store)

License

MIT