2nd/10_Wiki/Topics/AI_and_ML/LlamaIndex.md

id, title, category, status, canonical_id, aliases, duplicate_of, source_trust_level, confidence_score, verification_status, tags, raw_sources, last_reinforced, github_commit, tech_stack

id	title	category	status	canonical_id	aliases	duplicate_of	source_trust_level	confidence_score	verification_status	tags	raw_sources	last_reinforced	github_commit	tech_stack
wiki-2026-0508-llamaindex	LlamaIndex	10_Wiki/Topics	verified	self	GPT Index LlamaIndex Framework	none	A	0.9	applied	rag llamaindex retrieval indexing agents		2026-05-10	pending	language framework python/ts llama-index

LlamaIndex

매 한 줄

"매 LangChain 이 chain, LlamaIndex 는 index". Data → Index → Query Engine, RAG 에 특화된 framework.

매 핵심

매 핵심 추상

• Documents / Nodes: 원본 → chunked nodes
• Index: VectorStoreIndex, SummaryIndex, KnowledgeGraphIndex, TreeIndex
• Query Engine: retriever + synthesizer + (optional) postprocessor
• Agents: ReAct, OpenAI tool, function calling
• Workflows: event-driven multi-step (LangGraph 대응)

매 vs LangChain

	LlamaIndex	LangChain
강점	RAG 데이터 indexing	Agent / chain orchestration
추상	Index 중심	Chain/Runnable 중심
Eval	LlamaIndex Eval (faithfulness, relevancy)	LangSmith
권장	RAG 헤비	다양한 tool/agent

매 응용

1. 사내 docs Q&A
2. Code RAG (repo 전체 indexing)
3. Multi-doc summarization
4. Knowledge graph + RAG hybrid
5. Agentic RAG (자가 query 재작성)

💻 패턴

Pattern 1: Vector index basic

from llama_index.core import VectorStoreIndex, SimpleDirectoryReader
from llama_index.llms.anthropic import Anthropic

docs = SimpleDirectoryReader("./data").load_data()
index = VectorStoreIndex.from_documents(docs)
qe = index.as_query_engine(llm=Anthropic(model="claude-opus-4-7"), similarity_top_k=5)
print(qe.query("회사 휴가 정책 요약"))

Pattern 2: Persistent vector store (Chroma)

import chromadb
from llama_index.vector_stores.chroma import ChromaVectorStore
from llama_index.core import StorageContext

client = chromadb.PersistentClient(path="./chroma")
collection = client.get_or_create_collection("docs")
vs = ChromaVectorStore(chroma_collection=collection)
storage = StorageContext.from_defaults(vector_store=vs)
index = VectorStoreIndex.from_documents(docs, storage_context=storage)

Pattern 3: Hybrid retrieval (vector + BM25)

from llama_index.retrievers.bm25 import BM25Retriever
from llama_index.core.retrievers import QueryFusionRetriever

vec_r = index.as_retriever(similarity_top_k=5)
bm25_r = BM25Retriever.from_defaults(nodes=index.docstore.docs.values(), similarity_top_k=5)
fusion = QueryFusionRetriever([vec_r, bm25_r], num_queries=1, mode="reciprocal_rerank")

Pattern 4: Re-ranker (Cohere)

from llama_index.postprocessor.cohere_rerank import CohereRerank

reranker = CohereRerank(top_n=3)
qe = index.as_query_engine(node_postprocessors=[reranker], similarity_top_k=20)
# 20 후보 → rerank → top 3

Pattern 5: Sub-question for multi-doc

from llama_index.core.query_engine import SubQuestionQueryEngine
from llama_index.core.tools import QueryEngineTool

tools = [
    QueryEngineTool.from_defaults(query_engine=qe_a, name="finance"),
    QueryEngineTool.from_defaults(query_engine=qe_b, name="hr"),
]
sub = SubQuestionQueryEngine.from_defaults(query_engine_tools=tools)
sub.query("작년 인건비 대비 헤드카운트 변화는?")

Pattern 6: Eval (faithfulness)

from llama_index.core.evaluation import FaithfulnessEvaluator
ev = FaithfulnessEvaluator(llm=Anthropic(model="claude-opus-4-7"))
resp = qe.query("X 가 무엇?")
result = ev.evaluate_response(response=resp)
assert result.passing  # answer grounded in retrieved context?

Pattern 7: Agent with tools

from llama_index.core.agent import ReActAgent
agent = ReActAgent.from_tools([search_tool, calc_tool], llm=Anthropic(...))
agent.chat("작년 매출 대비 올해 성장률?")

Pattern 8: Workflow (event-driven)

from llama_index.core.workflow import Workflow, step, Event

class RetrieveEvent(Event): query: str
class GenEvent(Event): nodes: list

class RagFlow(Workflow):
    @step
    async def retrieve(self, ev: RetrieveEvent) -> GenEvent:
        return GenEvent(nodes=retriever.retrieve(ev.query))
    @step
    async def generate(self, ev: GenEvent):
        return synthesize(ev.nodes)

매 결정 기준

상황	Tool
RAG 가 메인	LlamaIndex
복잡한 agent / tool 오케스트레이션	LangChain / LangGraph
Production simple RAG	LlamaIndex + Chroma/Qdrant
Multi-doc 합성	SubQuestionQueryEngine
정확도 push	hybrid + reranker

기본값: VectorStoreIndex + Chroma + Cohere rerank + faithfulness eval.

🔗 Graph

• 부모: RAG
• 변형: LangChain
• 응용: Embedding
• Adjacent: LLM_Ops_and_Tuning, Prompt_Engineering

🤖 LLM 활용

언제: doc Q&A, code RAG, multi-doc summarization. 언제 X: 단일 prompt 로 충분 (RAG overkill), real-time chat 만 필요 (index 비용).

❌ 안티패턴

• Chunk size 무조건 default → recall 저하
• Re-rank 안 함 → 상위 k 노이즈
• Eval 없이 prod → silent quality drop
• 모든 doc 한 index → namespace 분리 안하면 권한/품질 혼탁
• VectorStoreIndex 만 사용, BM25 안 섞음 → keyword query 약함

🧪 검증 / 중복

• Verified (LlamaIndex docs, ChromaDB, Cohere rerank). 신뢰도 A.

🕓 Changelog

날짜	변경
2026-05-08	Phase 1
2026-05-10	Manual cleanup — vs LangChain, hybrid+rerank+eval patterns