2nd/10_Wiki/Topics/AI_and_ML/Semantic Search.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-semantic-search	Semantic Search	10_Wiki/Topics	verified	self	Vector Search Dense Retrieval Neural Search Semantic Search with AI	none	A	0.93	applied	search retrieval embeddings vector-db rag		2026-05-10	pending	language framework python faiss

Semantic Search

매 한 줄

"매 query → embedding → ANN nearest neighbors in vector space". 매 BM25 매 lexical 한계를 dense retrieval (DPR, ColBERT) 매 극복. 매 2026 production: hybrid (BM25 + dense + reranker), 매 모범: OpenAI text-embedding-3-large, Cohere v3, Voyage-3, BGE-M3, Jina-v3.

매 핵심

매 Pipeline

1. Index time: doc → chunk → embed → vector DB.
2. Query time: query → embed → ANN search → (rerank) → results.
3. Hybrid: BM25 score + dense score → RRF or weighted.
4. Rerank: cross-encoder on top-100 → top-10 (Cohere Rerank, BGE-Reranker).

매 Embedding models (2026)

• OpenAI text-embedding-3-large (3072d, MRL truncatable).
• Cohere embed-v3 (multilingual, dot-product).
• Voyage-3 (state-of-art retrieval).
• BGE-M3 (open, multi-vector, sparse+dense).
• Jina-v3 (8k context, MRL).
• NV-Embed-v2 (NVIDIA, MTEB top).

매 ANN algorithms

• HNSW (graph): 매 default, fast, high recall.
• IVF-PQ (Faiss): 매 huge scale, compressed.
• DiskANN: 매 on-disk billion-scale.
• ScaNN (Google): 매 best at fixed memory.

매 Vector DBs

• Pinecone (managed).
• Weaviate (open + hybrid built-in).
• Qdrant (Rust, fast).
• Milvus (large-scale).
• pgvector (Postgres).
• LanceDB (embedded, columnar).
• Turbopuffer (serverless 2024+).

매 응용

1. RAG knowledge retrieval.
2. Code search (Cursor, Sourcegraph).
3. E-commerce / product search.
4. Multimodal (CLIP image+text).

💻 패턴

Basic dense retrieval

from openai import OpenAI
import numpy as np
import faiss

client = OpenAI()

def embed(texts):
    r = client.embeddings.create(model="text-embedding-3-large", input=texts)
    return np.array([d.embedding for d in r.data], dtype="float32")

docs = ["Doc 1 text...", "Doc 2 text...", "..."]
doc_vecs = embed(docs)
index = faiss.IndexHNSWFlat(3072, 32)
faiss.normalize_L2(doc_vecs)
index.add(doc_vecs)

q_vec = embed(["What is X?"])
faiss.normalize_L2(q_vec)
D, I = index.search(q_vec, 10)
print([docs[i] for i in I[0]])

Hybrid (BM25 + dense) with RRF

from rank_bm25 import BM25Okapi

bm25 = BM25Okapi([d.split() for d in docs])

def rrf(rankings, k=60):
    scores = {}
    for ranking in rankings:
        for rank, doc_id in enumerate(ranking):
            scores[doc_id] = scores.get(doc_id, 0) + 1 / (k + rank)
    return sorted(scores.items(), key=lambda x: -x[1])

def hybrid_search(query, k=10):
    bm25_top = np.argsort(-bm25.get_scores(query.split()))[:50]
    q_vec = embed([query]); faiss.normalize_L2(q_vec)
    _, dense_top = index.search(q_vec, 50)
    fused = rrf([bm25_top.tolist(), dense_top[0].tolist()])
    return [docs[i] for i, _ in fused[:k]]

Cross-encoder reranking

import cohere
co = cohere.Client()

def rerank(query, candidates, top_n=10):
    r = co.rerank(query=query, documents=candidates,
                  model="rerank-english-v3.0", top_n=top_n)
    return [candidates[res.index] for res in r.results]

Chunking with overlap

def chunk_text(text, size=500, overlap=50):
    words = text.split()
    chunks = []
    for i in range(0, len(words), size - overlap):
        chunk = " ".join(words[i:i+size])
        chunks.append(chunk)
    return chunks

# 매 better: 매 semantic chunker (매 paragraph + heading aware)
from langchain.text_splitter import RecursiveCharacterTextSplitter
splitter = RecursiveCharacterTextSplitter(chunk_size=500, chunk_overlap=50,
                                          separators=["\n\n", "\n", ". ", " "])

MRL truncation (Matryoshka)

# text-embedding-3-large: 3072d, truncatable to 256/512/1024
def embed_mrl(text, dim=512):
    full = embed([text])[0]
    truncated = full[:dim]
    return truncated / np.linalg.norm(truncated)
# 매 6× memory savings, 매 ~95% recall.

ColBERT (multi-vector late interaction)

from colbert.modeling.colbert import ColBERT

# 매 token-level vectors per query+doc; 매 max-sim per query token then sum.
def colbert_score(query_vecs, doc_vecs):
    # query_vecs: [Q, d], doc_vecs: [D, d]
    sim = query_vecs @ doc_vecs.T  # [Q, D]
    return sim.max(axis=1).sum()   # 매 sum of per-token max

pgvector hybrid (production)

CREATE TABLE docs (id bigserial, content text, embedding vector(1536),
                   tsv tsvector GENERATED ALWAYS AS (to_tsvector('english', content)) STORED);
CREATE INDEX ON docs USING hnsw (embedding vector_cosine_ops);
CREATE INDEX ON docs USING gin (tsv);

-- Hybrid query
WITH dense AS (
  SELECT id, 1 - (embedding <=> $1) AS score FROM docs ORDER BY embedding <=> $1 LIMIT 50
), sparse AS (
  SELECT id, ts_rank_cd(tsv, websearch_to_tsquery($2)) AS score
  FROM docs WHERE tsv @@ websearch_to_tsquery($2) LIMIT 50
)
SELECT id, COALESCE(d.score, 0) * 0.7 + COALESCE(s.score, 0) * 0.3 AS score
FROM dense d FULL OUTER JOIN sparse s USING (id)
ORDER BY score DESC LIMIT 10;

Multimodal CLIP search

import torch
from transformers import CLIPModel, CLIPProcessor

model = CLIPModel.from_pretrained("openai/clip-vit-large-patch14")
proc = CLIPProcessor.from_pretrained("openai/clip-vit-large-patch14")

def embed_image(img):
    with torch.no_grad():
        return model.get_image_features(**proc(images=img, return_tensors="pt"))

def embed_text(t):
    with torch.no_grad():
        return model.get_text_features(**proc(text=t, return_tensors="pt"))
# 매 same vector space → cross-modal search.

매 결정 기준

상황	Approach
Quick prototype	매 OpenAI embeddings + Faiss/LanceDB
Production RAG	매 hybrid (BM25 + dense) + Cohere rerank
Self-host open	매 BGE-M3 + Qdrant + BGE-reranker
Multilingual	매 BGE-M3, Cohere multilingual, embed-v4
Code search	매 Voyage-code-3 또는 jina-code-v2
Multimodal	매 CLIP / SigLIP / Jina-CLIP

기본값: 매 production RAG → hybrid (BM25 + dense) + cross-encoder rerank.

🔗 Graph

• 부모: Information Retrieval · Embeddings
• 변형: Dense Retrieval · Sparse Retrieval · Information-Retrieval-IR · ColBERT
• 응용: RAG · Recommender Systems
• Adjacent: BM25 · Cross-Encoder Reranking · CLIP

🤖 LLM 활용

언제: 매 RAG retrieval, 매 semantic deduplication, 매 cross-lingual search, 매 recommendation. 언제 X: 매 exact-match (use BM25), 매 small corpus (<1k docs — 매 LLM-direct 가 simpler), 매 high-precision regex needs.

❌ 안티패턴

• Dense-only: 매 BM25 매 still wins on rare terms / proper nouns — 매 hybrid.
• No reranker: 매 top-10 quality 매 leaves 30% on table.
• Bad chunking: 매 fixed-size mid-sentence — 매 use semantic / heading-aware.
• No metadata filter: 매 hybrid filter (date/source) before vector search.
• Cosine without normalize: 매 silent bug — 매 always normalize L2.

🧪 검증 / 중복

• Verified (Karpukhin DPR 2020, Khattab ColBERT 2020, MTEB benchmark, Cohere Rerank docs).
• 신뢰도 A.

🕓 Changelog

날짜	변경
2026-05-08	Phase 1
2026-05-10	Manual cleanup — hybrid, MRL, ColBERT, pgvector, multimodal