You write the Oracle data-layer modules. You do not write app code, chains, or UI.

Step 0 — References (mandatory)

• shared/references/langchain-oracledb.md — load-bearing.
• shared/snippets/metadata_monkeypatch.py — copy verbatim into _monkeypatch.py.
• shared/snippets/oracle_chat_history.py — copy verbatim into history.py (includes the init_table() helper running an idempotent PL/SQL DDL block — Oracle does NOT support CREATE TABLE IF NOT EXISTS).
• shared/snippets/in_db_embeddings.py — copy verbatim into store.py when embedder == "in-db-onnx" (this is THE InDBEmbeddings subclass referenced everywhere; it lives here so the helper actually ships it instead of asking the user to invent it).
• shared/references/onnx-in-db-embeddings.md — only if embedder == "in-db-onnx".

Step 1 — Validate inputs

• target_dir/.env exists and has DB_DSN, DB_USER, DB_PASSWORD. The DB_USER MUST be the app user that oracle-aidb-docker-setup Step 6 created (NOT SYSTEM — SYSTEM's tablespace can't hold JSON columns; you'll get ORA-43853). If not, stop — tell the user to run oracle-aidb-docker-setup first.
• package_slug matches [a-z][a-z0-9_]*. Reject otherwise.
• collections non-empty. Naming: <PROJECT_PREFIX>_<KIND> enforced — e.g. for slug pdf_chat and kind DOCUMENTS, the actual table name is PDF_CHAT_DOCUMENTS. Document this in the file's docstring.
• For embedder == "in-db-onnx", confirm the user has a registered ONNX model name. If not, stop — point them at shared/references/onnx-in-db-embeddings.md step 3 (the onnx2oracle recipe).

Step 2 — Pick embedding dim

Hard-code EXPECTED_DIM in store.py. verify.py (written by the tier skill) asserts len(embedder.embed_query("dim check")) == EXPECTED_DIM — runtime check, not import-time.

The whole point of defaulting to minilm-py in beginner and in-db-onnx in intermediate/advanced: same model, same dim, same chunk-size sweet spot across tiers. A corpus ingested at tier 1 can be re-ingested at tier 2 against the same embedding space — only the inference location changes.

Step 3 — Write _monkeypatch.py

Copy shared/snippets/metadata_monkeypatch.py verbatim. Add a docstring at the top:

"""
langchain-oracledb stores metadata as JSON strings, but its similarity_search
return path doesn't always parse them back. This monkeypatch makes the parsing
consistent. MUST be imported before any OracleVS instantiation in your app.

Source: shared/references/langchain-oracledb.md § "Metadata-as-string fix"
"""

Tier skills then add from <package_slug>._monkeypatch import * at the top of store.py, app.py, and any other entry point. Don't skip this — categorical metadata filtering breaks silently without it.

Step 4 — Write store.py

Skeleton:

"""
Oracle vector store layer for <package_slug>.

Owns:
- Multi-collection OracleVS wrapper (one logical collection per kind)
- Embedder factory (selected at scaffold time: <embedder>)
- Connection lifecycle (one shared connection, lazy)

Cites:
- shared/references/langchain-oracledb.md
- shared/snippets/in_db_embeddings.py (when embedder=in-db-onnx)
"""
from . import _monkeypatch  # noqa: F401  -- must be first

import os
import oracledb
from langchain_oracledb.vectorstores.oraclevs import OracleVS

# DistanceStrategy lives in langchain_community — `langchain-oracledb` does
# NOT re-export it. Friction P1-1.
from langchain_community.vectorstores.utils import DistanceStrategy

# ... embedder import per choice ...

PROJECT_PREFIX = "<UPPER_PACKAGE_SLUG>"
EXPECTED_DIM = <384 (MiniLM-L6-v2) | 1024 (cohere)>

_conn = None
_embedder = None

def get_connection() -> oracledb.Connection:
    global _conn
    if _conn is None or not _conn.ping():
        _conn = oracledb.connect(
            user=os.environ["DB_USER"],
            password=os.environ["DB_PASSWORD"],
            dsn=os.environ["DB_DSN"],
        )
    return _conn

def get_embedder():
    global _embedder
    if _embedder is None:
        _embedder = <embedder factory call>  # see "Step 2 — Pick embedding dim"
    return _embedder

def get_store(kind: str) -> OracleVS:
    table = f"{PROJECT_PREFIX}_{kind.upper()}"
    return OracleVS(
        client=get_connection(),
        embedding_function=get_embedder(),
        table_name=table,
        distance_strategy=DistanceStrategy.COSINE,
    )

def bootstrap() -> None:
    """Idempotent: ensure each collection table exists by inserting a
    no-op then immediately deleting it. Cheaper than checking USER_TABLES."""
    for kind in <collections list>:
        store = get_store(kind)
        ids = store.add_texts(["__bootstrap__"], metadatas=[{"_skip": True}])
        store.delete(ids)

Embedder factory expressions:

• embedder == "minilm-py":

  from langchain_huggingface import HuggingFaceEmbeddings
  _embedder = HuggingFaceEmbeddings(model_name=os.environ.get(
      "EMBED_MODEL", "sentence-transformers/all-MiniLM-L6-v2"
  ))

• embedder == "in-db-onnx": copy shared/snippets/in_db_embeddings.py verbatim into the project as src/<package_slug>/in_db_embeddings.py, then:

  from .in_db_embeddings import InDBEmbeddings
  _embedder = InDBEmbeddings(get_connection(), model_db_name=os.environ.get(
      "ONNX_MODEL_NAME", "MY_MINILM_V1"
  ))

• embedder == "oci-cohere": copy shared/snippets/oci_cohere_embeddings.py per its docstring.

Replace placeholders with concrete values from inputs. The bootstrap dance is the load-bearing trick — OracleVS.from_texts creates the table on first call; subsequent calls are no-ops.

Step 5 — Write history.py (if has_chat_history)

Copy shared/snippets/oracle_chat_history.py verbatim into target_dir/src/<package_slug>/history.py. Add the LangChain glue at the bottom:

def get_history_factory(conn):
    """Return a callable suitable for RunnableWithMessageHistory."""
    def _factory(session_id: str):
        return OracleChatHistory(conn, session_id)
    return _factory

Then migrations/001_chat_history.sql (Oracle does NOT support CREATE TABLE IF NOT EXISTS — wrap DDL in a PL/SQL anonymous block that swallows ORA-00955; matches the snippet's INIT_DDL):

BEGIN
    EXECUTE IMMEDIATE q'[
        CREATE TABLE chat_history (
            session_id VARCHAR2(120) NOT NULL,
            seq        NUMBER GENERATED ALWAYS AS IDENTITY,
            payload    CLOB CHECK (payload IS JSON),
            created_at TIMESTAMP DEFAULT SYSTIMESTAMP NOT NULL,
            PRIMARY KEY (session_id, seq)
        )
    ]';
EXCEPTION WHEN OTHERS THEN
    IF SQLCODE != -955 THEN RAISE; END IF;
END;
/

Schema matches the snippet's contract (single CLOB payload validated as JSON). The tier skill is responsible for running this migration during bootstrap (or calling history.init_table(get_connection())). Document that requirement in the docstring at the top of history.py.

Step 6 — Smoke

After writing, run from target_dir:

from <package_slug>.store import bootstrap, get_store, get_embedder, EXPECTED_DIM
bootstrap()
v = get_embedder().embed_query("dim check")
assert len(v) == EXPECTED_DIM, f"dim mismatch: got {len(v)} expected {EXPECTED_DIM}"
print("langchain-oracledb-helper: OK")

If dim mismatches: drop the offending tables (DROP TABLE <PREFIX>_<KIND> CASCADE CONSTRAINTS), fix the embedder, re-bootstrap. Mismatches happen most when the user changes embedder mid-project.

Stop conditions

• langchain-oracledb not in pyproject.toml. Tell the user to add langchain-oracledb>=0.1 and stop.
• Embedder choice doesn't match what the project already uses (existing tables have a different dim). Surface mismatch, don't silently re-bootstrap.
• has_chat_history=True but no chat_history table after migration. Stop — the migration didn't run.

What you must NOT do

• Don't skip _monkeypatch.py. Filtered retrievals break silently.
• Don't from langchain_oracledb.chat_message_histories import ... — it doesn't exist.
• Don't write SQL DDL for vector tables manually. OracleVS.from_texts (via the bootstrap dance) handles it.
• Don't pin a different distance strategy unless the user asks. COSINE is the default and matches the rest of the skill set.

Final report

langchain-oracledb-helper: OK
  store:        target_dir/src/<package_slug>/store.py
  monkeypatch:  target_dir/src/<package_slug>/_monkeypatch.py
  history:      target_dir/src/<package_slug>/history.py        (if scaffolded)
  migrations:   target_dir/migrations/001_chat_history.sql      (if scaffolded)
  embedder:     <minilm-py|in-db-onnx|oci-cohere> (dim=<384|384|1024>)
  collections:  <list>
  next:         hand off to the tier skill — it writes app code that imports `store`.