Split View: GraphRAG 완전 가이드: 지식 그래프가 RAG의 한계를 어떻게 넘는가

Standard RAG:
Documents -> Chunk -> Embed -> Vector DB -> Similarity Search -> Relevant Chunks

GraphRAG:
Documents -> Entity Extraction -> Knowledge Graph -> Community Detection ->
          -> Hierarchical Summaries -> Multi-level Retrieval (Local + Global)

Input text:
"Samsung launched the Galaxy S25 in 2024, competing directly with Apple's
iPhone 16. The Galaxy S25 is powered by the Qualcomm Snapdragon 8 Elite chip."

Extracted entities:
- Samsung (organization)
- Galaxy S25 (product)
- Apple (organization)
- iPhone 16 (product)
- Qualcomm (organization)
- Snapdragon 8 Elite (product/technology)

Extracted relationships:
- Samsung --[launched]--> Galaxy S25
- Galaxy S25 --[competes with]--> iPhone 16
- Galaxy S25 --[powered by]--> Snapdragon 8 Elite
- Qualcomm --[manufactures]--> Snapdragon 8 Elite

Level 0 (most detailed): individual entity/relationship summaries
Level 1: small community summaries
Level 2: mid-scale community summaries (most commonly used)
Level 3 (most comprehensive): global topic summaries

# Install
pip install graphrag

# Initialize project
mkdir my-graphrag-project
graphrag init --root ./my-graphrag-project

# Key settings in settings.yaml
llm:
  api_key: ${GRAPHRAG_API_KEY}
  type: openai_chat
  model: gpt-4o-mini  # for indexing (mini to reduce cost)
  model_supports_json: true

embeddings:
  llm:
    model: text-embedding-3-small

input:
  type: file
  file_type: text
  base_dir: "input"  # put your documents here

chunks:
  size: 1200
  overlap: 100

# Indexing (building the knowledge graph)
# graphrag index --root ./my-graphrag-project
# What this does internally:
# 1. Chunk documents
# 2. Extract entities and relationships (LLM calls)
# 3. Build the graph
# 4. Run community detection
# 5. Generate summaries for each community (LLM calls)

# Querying via Python API
import asyncio
from graphrag.query.api import local_search, global_search

# Local Search: entity/relationship-focused questions
async def search_local(query: str):
    result = await local_search(
        config_dir="./my-graphrag-project",
        data_dir="./my-graphrag-project/output",
        root_dir="./my-graphrag-project",
        community_level=2,
        response_type="multiple paragraphs",
        query=query,
    )
    return result.response

# Global Search: questions requiring synthesis across the full knowledge base
async def search_global(query: str):
    result = await global_search(
        config_dir="./my-graphrag-project",
        data_dir="./my-graphrag-project/output",
        root_dir="./my-graphrag-project",
        community_level=2,
        response_type="multiple paragraphs",
        query=query,
    )
    return result.response

# Usage
local_result = asyncio.run(search_local(
    "What is the relationship between Samsung and TSMC?"
))

global_result = asyncio.run(search_global(
    "What are the main trends in the semiconductor industry across these documents?"
))

import pandas as pd
import networkx as nx

# Load entities and relationships from GraphRAG output
entities_df = pd.read_parquet("./output/entities.parquet")
relationships_df = pd.read_parquet("./output/relationships.parquet")

print(f"Total entities: {len(entities_df)}")
print(f"Total relationships: {len(relationships_df)}")

# Build NetworkX graph
G = nx.DiGraph()

for _, entity in entities_df.iterrows():
    G.add_node(entity["title"], type=entity["type"])

for _, rel in relationships_df.iterrows():
    G.add_edge(
        rel["source"],
        rel["target"],
        weight=rel["weight"],
        description=rel["description"]
    )

# Find hub entities (most connected)
top_hubs = sorted(G.degree(), key=lambda x: x[1], reverse=True)[:10]
print("Top 10 most connected entities:")
for entity, degree in top_hubs:
    print(f"  {entity}: {degree} connections")

Global queries (requiring whole-document synthesis):
- Standard RAG:  comprehensiveness 40%, diversity 57%
- GraphRAG:      comprehensiveness 72%, diversity 62%
  -> 80% improvement in comprehensiveness!

Local queries (specific fact retrieval):
- Standard RAG:  accuracy ~65%
- GraphRAG:      accuracy ~70%
  -> Modest improvement

Latency:
- Standard RAG:  0.5-2 seconds
- GraphRAG:      3-10 seconds (due to community summary aggregation)

# GraphRAG indexing cost estimate
# Assumptions: 1,000 documents, 1,000 tokens each

num_documents = 1000
tokens_per_doc = 1000
total_tokens = num_documents * tokens_per_doc  # 1M tokens

# Estimated LLM calls during indexing
entity_extraction_calls = total_tokens / 1200  # 1 call per chunk
# Each call: ~800 input tokens + ~400 output tokens

community_summary_calls = 200  # number of communities x levels
# Each call: ~2000 input tokens + ~500 output tokens

# Cost using GPT-4o-mini ($0.15/1M input, $0.60/1M output)
entity_input_cost = (entity_extraction_calls * 800 / 1_000_000) * 0.15
entity_output_cost = (entity_extraction_calls * 400 / 1_000_000) * 0.60
community_cost = (community_summary_calls * 2500 / 1_000_000) * 0.40

total_indexing_cost = entity_input_cost + entity_output_cost + community_cost
print(f"Estimated indexing cost: ${total_indexing_cost:.2f}")
# 1,000 documents: approx $1-5 (using GPT-4o-mini)
# 10,000 documents: approx $10-50
# 100,000 documents: approx $100-500

# Query cost (Global Search)
# Global search processes all community summaries for each query
# 200 communities x 500 tokens each = ~100K tokens processed per query
global_query_cost = (100_000 / 1_000_000) * 2.50  # GPT-4o pricing
print(f"Global search cost per query: ${global_query_cost:.3f}")  # $0.25/query

# pip install lightrag-hku
from lightrag import LightRAG, QueryParam
from lightrag.llm import gpt_4o_mini_complete

rag = LightRAG(
    working_dir="./lightrag-storage",
    llm_model_func=gpt_4o_mini_complete,
)

# Insert documents
with open("document.txt", "r") as f:
    rag.insert(f.read())

# Query (4 modes: naive, local, global, hybrid)
result = rag.query(
    "What are the main trends?",
    param=QueryParam(mode="global")  # hybrid mode is also effective
)