The embedder decides what "similar" even means.

Same chunks as Lesson 2. Pick the wrong embedding model and retrieval fails silently, no error, just wrong answers.

Recall: KB course drew the line between exact-match (BM25/keyword) and meaning-match (embeddings). Today you build the meaning-match side, and hit the exact case where it loses on purpose. Refresher: KB Lesson 2.

The one idea: An embedding model turns text into a list of numbers such that similar meaning sits close together in that number-space. Choosing that model is a design decision, not a formality. An English-only embedder does not error on Korean text, it just places it somewhere meaningless, and every downstream stage inherits that silently.

Prove the wrong choice fails silently

from langchain_openai import OpenAIEmbeddings

# English-tuned embedder, applied to the Korean rider clause from Lesson 2
english_only = OpenAIEmbeddings(model="text-embedding-3-small")
korean_text = "회사는 다음 중 어느 한 가지로 보험금 지급사유가 발생한 때에는 보험금을 지급하지 아니합니다."
english_text = "The company will not pay the benefit under these excluded causes."

vec_kr = english_only.embed_query(korean_text)
vec_en = english_only.embed_query(english_text)
# These two sentences mean nearly the same thing. A good multilingual
# embedder places them close together. An English-tuned one often won't.

Nothing crashes. That is the danger: a bad embedding choice looks identical to a working system until a customer asks a question the retriever cannot place correctly.

The fix: a multilingual embedder

from langchain_huggingface import HuggingFaceEmbeddings

embeddings = HuggingFaceEmbeddings(
    model_name="sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2",
    encode_kwargs={"normalize_embeddings": True},
)

This model is trained on 50+ languages including Korean, runs free and local (no API key, no per-call cost), and is the same category of fix as the Thai embedder swap from your DuckyCutz work. Same lesson, second language.

Store it: Chroma

from langchain_chroma import Chroma

vector_store = Chroma(
    collection_name="samsung_life_demo",
    embedding_function=embeddings,
    persist_directory="./chroma_demo_db",   # local file, survives restarts
)

vector_store.add_documents(documents=good_chunks)  # from Lesson 2

Chroma is a vector database that stores each chunk's embedding alongside the chunk's own text and metadata (source file, policy ID). "Local file" matters for the pitch later: nothing here needs to leave the client's infrastructure.

Retrieve, and hit the identifier gap

results = vector_store.similarity_search("does this policy cover flood damage?", k=3)
for r in results:
    print(r.metadata["source"], "→", r.page_content[:80])

# Now the exact-identifier case
results = vector_store.similarity_search("policy S-240-113", k=3)
for r in results:
    print(r.metadata["source"], "→", r.page_content[:80])

The first query works well, "flood damage" and "coverage" sit close in meaning-space. The second often doesn't: an embedder maps S-240-113 as a near-arbitrary string, not a special exact token. This is the same gap KB Lesson 2 named: embeddings are bad at exact identifiers, keyword/BM25 search is built for exactly this case. Production systems run both and merge results (hybrid retrieval), covered in KB Lesson 3. For this demo, note the gap and move on, don't build the hybrid layer until a real customer question needs it.

Every result carries its receipt

Notice r.metadata["source"] above. Every chunk retrieved carries the file (and policy ID) it came from. This is not incidental, it is the raw material Lesson 5's citation check runs on: an answer can only cite a source that was actually retrieved.

Prove it stuck

You embed Korean text with an English-only model. What actually happens?

Why does "policy S-240-113" retrieve badly with pure embedding search?

Why does every retrieved chunk carry its source metadata?

Read this next (primary source)

★ LangChain's RAG tutorial, Part 1, the embed/store/retrieve sections, plus the Chroma docs "Usage" page for the store API you just used.

Ask your teacher. Try: "show me a real hybrid search result for the policy ID case", "what other multilingual embedding models exist besides the sentence-transformers one?", or "how do I check retrieval quality before showing a demo to a prospect?"

← Lesson 2: Corpus + chunking Next → Lesson 4: LangGraph