oracle-ai-developer-hub

Workshop TODO checklist

notebook_student.ipynb has eight TODOs: TODO 1 is an inline prompt, TODOs 2–8 are stubs. Each has one checkpoint cell that prints a single ✅/❌ line and fails on the stub, on trivial returns and on the obvious bug. Each stub’s comment block holds its signature, steps and checkpoint. Run All is supposed to halt at the first unfinished TODO. Answers: notebook_complete.ipynb.

TODO	Name	Section	Guide
1	QUESTION: ask the bare model	§1.3	part-1-setup.md
2	recall_memories	§2.2	part-2-oamp-memory.md
3	_scan_tables	§2.3	part-2-oamp-memory.md
4	retrieve_knowledge	§3.2	part-3-retrieval.md
5	retrieve_tools	§4.1	part-4-tools-and-skills.md
6	tool_run_sql	§4.2	part-4-tools-and-skills.md
7	tool_list_skills	§4.4	part-4-tools-and-skills.md
8	agent_turn	§5.3	part-5-agent-loop.md

§3.3 (hybrid_search_knowledge) is provided: run it, no TODO.

Part 6 (guide) has no TODO: it runs the harness you built.

Before you start

• Oracle is reachable and the FINANCE schema is seeded (app/scripts/seed.py).
• ALL_MINILM_L12_V2 is loaded (and RERANKER_ONNX if the cross-encoder is provisioned).
• The identity rules are installed (app/scripts/setup_deep_security.py); the Codespace does this during setup.
• The kernel is the Python 3.11 workshop kernel with the packages from requirements.txt.
• §1.2 preflight is green. Each ❌ row prints its own fix; only a missing FINANCE or embedder stops the notebook.

After the eight TODOs

• Run the two-turn demo in §5.4 (discovery, live SQL).
• Open the app at http://localhost:3000.
• Ask “Which branch regions have the most FLAGGED or BLOCKED transactions?”
• Switch the header persona to Analyst — Europe & Middle East and re-ask; the rows change.
• As Compliance Officer, ask for Suspicious Activity Reports: all rows. As Analyst (default), ask again: no rows. The database answered differently; nothing in the UI changed.

Part 6 (capstone)

• §6.1: read the ledger DDL and build the queue. The cell prints the alerts in the window and the ones it will work (TRIAGE_LIMIT).
• §6.2: read one evidence pack, then the policy and the validator. Unknown decisions become REVIEW_REQUIRED.
• §6.3: run the triage; read the tool trace and the ledger rows.
• §6.4: read the impact board (the only assumption is MANUAL_MINUTES_PER_ALERT), then ask the agent about its own decisions.
• Set IGNORE_WATERMARK = False and re-run §6.1: a scheduled job reports no new alerts instead of re-triaging the book.

Reference material

Not in the notebook; read as needed: DBFS · Oracle MLE · Deep Data Security · duality views · tool-output offload (the notebook shows offload in §5.5).

To deploy against an Oracle that is not the workshop Codespace, run the same provisioning the Codespace runs: cd app && python scripts/bootstrap.py && python scripts/seed.py && python scripts/setup_advanced.py && python scripts/setup_deep_security.py, in that order (seed.py drops and recreates FINANCE, so setup_deep_security.py must follow). provision.sh sequences all four and is safe to re-run. See part-1-setup.md.