Build Paths Beginner

The user picked the beginner path. Your job is to interview them, scaffold a working project that ingests a corpus and exposes a Grok-4-powered chat UI in Open WebUI, run verify, and stop. The output is a real shippable mini-product, not a tutorial.

Step 0 — Read these references first (mandatory)

• shared/references/sources.md
• shared/references/oracle-26ai-free-docker.md
• shared/references/langchain-oracledb.md ← load-bearing
• shared/references/oci-genai-openai.md ← load-bearing (Pattern 1 SigV1 auth)
• shared/references/oracledb-python.md (skim — beginner only touches oracledb.connect)
• shared/references/exemplars.md
• beginner/project-ideas.md
• skills/oracle-aidb-docker-setup/SKILL.md — you'll invoke this
• skills/langchain-oracledb-helper/SKILL.md — you'll invoke this

You may not write SQL, embedder calls, or table-creation code that contradicts these files. If the user asks for something not covered, say so and stop — don't invent.

Step 1 — Interview

Run the questions from shared/interview.md. For beginner specifically:

• Q3 (DB target) — default to "Local Docker" without re-asking.
• Q4 (Inference) — not optional anymore at this tier. OCI Generative AI for the LLM (Grok 4 via bearer-token). MiniLM Python-side for embeddings. Verify:
  • OCI_GENAI_API_KEY (a sk-... value) is set in the user's shell env OR about to be written into the project's .env. If neither, stop and ask the user to generate one in the OCI GenAI service console.
  • The user is OK with non-zero OCI cost (mention this once — Grok 4 is not on the always-free list).
  • Default endpoint is https://inference.generativeai.us-phoenix-1.oci.oraclecloud.com; the user can override via OCI_GENAI_BASE_URL if their key is region-locked elsewhere.
  • Embedder: sentence-transformers/all-MiniLM-L6-v2 (384 dim) by default — runs inside the user's Python process via HuggingFaceEmbeddings. Same model that intermediate/advanced register inside Oracle, so corpus + chunks stay comparable across tiers.
  • Chat: xai.grok-4 (full id required — grok-4 alone won't resolve).
• Q5 (Topic) — pick one of the three from beginner/project-ideas.md. Map free-text pitches to the closest. If none fits, default to idea 1 (PDFs) and tell the user why.
• Q6 (Notebook) — default no. Beginner ships the chat UI, not a notebook walkthrough.

Print the confirmation block from interview.md. Don't proceed without an explicit y.

Step 2 — Resolve choices

Build a scaffold spec from the interview:

Step 3 — Scaffold

Order matters — invocation of building-block skills happens before project code.

3a — Foundation via building-block skills

1. Refuse if target_dir is non-empty.
2. Invoke skills/oracle-aidb-docker-setup. Pass target_dir. It writes docker-compose.yml (Oracle), .env (with generated ORACLE_PWD), brings the container up healthy. Block until it reports OK.
3. Append the Open WebUI service to the generated docker-compose.yml:

   open-webui:
     image: ghcr.io/open-webui/open-webui:main
     container_name: open-webui
     ports:
       - "127.0.0.1:3000:8080"
     environment:
       - OPENAI_API_BASE_URL=http://host.docker.internal:8000/v1
       - OPENAI_API_KEY=local-stub-key
       - WEBUI_AUTH=False
     extra_hosts:
       - "host.docker.internal:host-gateway"
     volumes:
       - openwebui_data:/app/backend/data
   volumes:
     openwebui_data:
   

4. Invoke skills/langchain-oracledb-helper. Pass target_dir, package_slug, embedder=minilm-py, collections=["DOCUMENTS", "CONVERSATIONS"], has_chat_history=True. It writes store.py, _monkeypatch.py, history.py, migrations/001_chat_history.sql. The helper will install sentence-transformers and pre-cache the model on first import to avoid a stall during the first user query. Block until it reports OK.

3b — Project-specific code (the only files this skill writes itself)

5. target_dir/.gitignore — extend with data/, *.pdf, notes/.
6. target_dir/pyproject.toml — extend deps:
   • Always: fastapi>=0.110, uvicorn[standard]>=0.27, langchain-core>=0.3, langchain-community>=0.3, openai>=1.40, langchain-huggingface>=0.1, sentence-transformers>=2.7.
   • Idea 1 (PDFs): + pypdf>=4.
   • Idea 2 (Markdown): + markdown-it-py>=3.
   • Idea 3 (Web): + trafilatura>=1.10, httpx>=0.27.
   • Do NOT add oci-openai or openai — friction P0-2; the OpenAI-compat path is unstable. The chat factory at shared/snippets/oci_chat_factory.py uses the direct OCI SDK.
7. target_dir/.env.example — append:

   OCI_GENAI_API_KEY=replace_me_sk_value_from_oci_genai_console
   OCI_GENAI_BASE_URL=https://inference.generativeai.us-phoenix-1.oci.oraclecloud.com
   OCI_LLM_MODEL=xai.grok-4
   EMBED_MODEL=sentence-transformers/all-MiniLM-L6-v2
   

   Confirm target_dir/.gitignore contains .env (the docker-setup skill adds it). Never commit OCI_GENAI_API_KEY.
8. src/<package_slug>/inference.py — copy shared/snippets/oci_chat_factory.py (chat — uses the direct OCI SDK, model id xai.grok-4). For the embedder, write a tiny factory:

   from langchain_huggingface import HuggingFaceEmbeddings
   _embedder = None
   def get_embedder():
       global _embedder
       if _embedder is None:
           _embedder = HuggingFaceEmbeddings(model_name=os.environ["EMBED_MODEL"])
       return _embedder
   

   Cite the OCI factory at the top.
9. src/<package_slug>/<ingest_module> — per idea:
   • Idea 1: walk data/pdfs/, parse via pypdf, chunk by page + 800-token windows, store.get_store("DOCUMENTS").add_texts(...) with metadata={"filename": ..., "page": ...}. Idempotent — keep a data/.ingested.json ledger.
   • Idea 2: walk NOTES_DIR, chunk by H2 sections via markdown-it-py, metadata = {"path": ..., "heading": ..., "frontmatter": {...}}.
   • Idea 3: python -m web_chat.add URL — trafilatura.fetch_url + extract, chunk, metadata = {"url": ..., "title": ..., "fetched_at": ..., "byline": ...}.
10. src/<package_slug>/chain.py — LCEL chain: retriever (get_store("DOCUMENTS").as_retriever(k=5)) → prompt with citation instructions → ChatOpenAI-shaped LLM via OCI. Wrap with RunnableWithMessageHistory(chain, get_history_factory(get_connection()), input_messages_key="question", history_messages_key="history"). Citation format depends on idea (filename:page / file#heading / URL).
11. src/<package_slug>/adapter.py — FastAPI app exposing /v1/chat/completions (OpenAI-compatible). Body shape matches what Open WebUI sends; map messages[-1].content to the chain input, return assistant message in OpenAI shape. Stream via text/event-stream.
12. verify.py — copy shared/templates/verify.template.py, fill in:
    • inference_enabled = True.
    • Round-trip: embedder.embed_query("dim check") → assert dim == 384.
    • Smoke a single chain call against a tiny known corpus (3 lines of test text).
13. README.md — copy shared/templates/readme.template.md, fill placeholders. The "Why Oracle" paragraph names: AI Vector Search, OracleVS, persistent chat history (OracleChatHistory). Include a "Stack" section listing OCI GenAI Grok 4, sentence-transformers/all-MiniLM-L6-v2 (Python-side, 384 dim — same model intermediate/advanced register inside Oracle for in-DB embeddings), langchain-oracledb, Open WebUI. Leave the screenshot slot for the chat UI.

Step 4 — Verify

1. The DB is already up (skill 1 ensured this).
2. From target_dir: python -m venv .venv && source .venv/bin/activate && python -m pip install -e ..
3. python verify.py. Expect verify: OK (db, vector, inference). The vector check asserts len(get_embedder().embed_query("dim check")) == 384. First run downloads the MiniLM weights (~90MB) — note this in the verify output so the user knows what's happening. On failure: follow shared/verify.md recovery loop, max 3 retries.
4. Bring Open WebUI up: docker compose up -d open-webui. Wait ~10s.
5. Run the adapter in the background: python -m <package_slug>.adapter (port 8000). Hit http://localhost:8000/v1/models — should return JSON listing one model. Then check http://localhost:3000 responds (Open WebUI loaded).
6. Don't keep the adapter running — just confirm it boots cleanly. Kill it before reporting done.

Step 5 — Polish for sharing

1. README — confirm all placeholders are filled.
2. docs/ — leave a note: "drop a 30s demo GIF as docs/demo.gif showing PDF drop → ingest → chat".
3. Final report:

   Done.
     project at:    <target_dir>
     run with:      cd <target_dir>
                    docker compose up -d
                    python -m <package_slug>.<ingest_module> <args>
                    python -m <package_slug>.adapter   # blocks; Open WebUI on :3000
     verify:        OK
     ui:            http://localhost:3000  (Open WebUI)
     adapter:       http://localhost:8000/v1/chat/completions
     next:          drop your corpus, run ingest, record a 30s demo, push to GitHub.
   

Stop conditions

• ~/.oci/config missing — stop, point at oci setup config.
• User explicitly refuses to use OCI GenAI (cost concern, no tenancy). Tell them this tier requires it; offer to point them at the archive's old Ollama-flavored beginner ideas if they want.
• Verify fails 3 times.
• Target dir non-empty.
• The user picks a non-Oracle DB or non-Python language. Print "out of scope for v1" and stop.

When to graduate to OAMP

If you grow this beginner project into a multi-user app — different humans hitting the same backend, each wanting their own preferences and durable facts remembered across sessions — swap the manual OracleChatHistory layer for OAMP (oracleagentmemory PyPI package). OAMP gives you per-user threads, automatic memory extraction, and prompt-ready context cards out of the box, and the advanced tier already wires it for you. See shared/references/oamp.md for the decision tree (OAMP vs. OracleVS vs. OracleChatHistory vs. plain SQL). Until then, stay on OracleChatHistory — single-user, full transcript visibility, no LLM-extraction cost.

What you must NOT do

• Don't write raw CREATE TABLE ... VECTOR(...) DDL — OracleVS.from_texts (via the helper skill's bootstrap dance) handles it.
• Don't introduce Gradio, Streamlit, Flask, or any non-Open-WebUI frontend. Open WebUI + FastAPI adapter is the contract.
• Don't introduce Redis, Postgres, ChromaDB, FAISS, or any non-Oracle store — even as a fallback.
• Don't introduce Ollama as a fallback. Earlier versions of this tier supported it; the new tier is OCI-only on purpose. If the user can't use OCI, point them at archive/beginner-ideas.md.
• Don't write more than ~450 lines of project code total. If you're going past that, stop and reduce.
• Don't generate ORACLE_PWD yourself — the helper skill does it.
• Don't claim done before verify is green AND the adapter boots cleanly.