oracle-ai-developer-hub

Part 2 — In-DB embeddings with OracleEmbeddings

Why in-database embeddings?

Most LangChain examples embed text by calling out to a remote API:

from langchain_openai import OpenAIEmbeddings
embeddings = OpenAIEmbeddings(model="text-embedding-3-small")  # ← HTTP round-trip per call

Every embed_query and embed_documents call is a network hop to OpenAI (or whoever). For a workshop this is fine. For a production agent that’s also running the database on the same node, it’s absurd — you’d send the text to OpenAI just to get a vector back, then write that vector into Oracle, when Oracle already has an embedding function sitting next to your data.

Oracle’s in-database ONNX embedder fixes this. The model lives inside the database, the embedding call is a SQL function, and your LangChain code never knows the difference.

ALL_MINILM_L12_V2 — what’s loaded

Property	Value
Model name in Oracle	ALL_MINILM_L12_V2
Architecture	MiniLM (sentence-transformers)
Embedding dimension	384
File size	~117 MB
Loaded by	app/scripts/onnx_setup.py
Loaded via	DBMS_VECTOR.LOAD_ONNX_MODEL(...)

After the load step, you can verify it directly in SQL:

SELECT VECTOR_EMBEDDING(ALL_MINILM_L12_V2 USING 'football cleats' AS DATA) FROM dual;
-- returns a 384-element VECTOR — no network call left the database.

How OracleEmbeddings wraps it

langchain_oracledb.OracleEmbeddings is a normal LangChain Embeddings subclass — it can be passed to OracleVS, OracleSemanticCache, or AsyncOracleStore like any other embedder. Internally it just runs the SQL above for every text it’s given.

from langchain_oracledb import OracleEmbeddings

embeddings = OracleEmbeddings(
    conn=oracle_client,                                          # sync handle
    params={"provider": "database", "model": "ALL_MINILM_L12_V2"},
)

The params dict tells OracleEmbeddings where the model lives. "provider": "database" means “the model is already loaded in this schema; don’t try to download anything.” "model": "ALL_MINILM_L12_V2" is the name we used in the LOAD_ONNX_MODEL step.

What you’ll build in TODO 1

Construct embeddings per the snippet above. The hard-stop checkpoint embeds a short string and asserts the resulting vector has the expected 384 dimensions. If it doesn’t, you’ve either pointed the embedder at the wrong model or the model wasn’t loaded.

Solution

Drop this into the TODO 1 cell, replacing the embeddings = None line:

from langchain_oracledb import OracleEmbeddings

embeddings = OracleEmbeddings(
    conn=oracle_client,
    params={"provider": "database", "model": ONNX_MODEL},
)

Next

→ Part 3 — OracleVS — use this embedder to power the vector knowledge base.