Generative AI Google Brain

The entire GoogleCloudPlatform/generative-ai repository compiled into a fully interlinked Obsidian second brain: concept maps, one page per notebook, an entity page for every model and API, and distilled best-practice guides, all cross-linked.

318 Google Cloud GenAI notebooks become 686 navigable, source-cited pages you can browse by concept, trace by relationship, and learn best practices from, without opening a single .ipynb.

Why this brain

• Browse by concept, not by folder. 20 thematic maps (Gemini, Agents and ADK, RAG and Grounding, Vertex AI Search, Evaluation, and more) sit on top of 317 per-notebook pages.
• Every claim is grounded. Each page links back to its source notebook on GitHub with a content hash. Adversarially verified at 95/100, with effectively zero hallucinations.
• Best practices, distilled. One source-cited guide per theme, plus the per-notebook practices that back it.

🌐 Open release. This public mirror lives at AgriciDaniel/google-genai-brain and is built for the AI Marketing Hub Pro community. It bundles Google's Apache-2.0 sample content (see THIRD_PARTY_NOTICES); our pipeline and synthesis are MIT.

Coverage

Quality gate, enforced by lint_vault.py and CI: 0 dead links, 0 dead embeds, 0 orphan pages.

Architecture

Four interlinked grammars in one vault. Open the graph view and they connect.

• Concepts (wiki/concepts/): thematic Maps of Content. Start here to learn an area.
• Source pages (wiki/source/): one page per notebook, mirroring the repo tree, with distilled code patterns, models and APIs used, gotchas, and best practices.
• Entities (wiki/entities/): models, APIs, and SDKs. Each links to every notebook that uses it.
• Best practices (wiki/best-practices/): distilled, source-cited guidance per theme.

Here is the real Obsidian graph view of this brain: 686 pages, color-coded by area, fully connected with no orphans.

Pipeline

Division of labor. codex (gpt-5.5, xhigh) does the judgment: reading each notebook and distilling summaries, gotchas, and best practices. Deterministic Python owns the structure: provenance, wikilinks, the graph, and a lint gate that guarantees zero dead links and zero orphans. A Claude Workflow adversarially verifies. The model is never trusted for file paths, URLs, hashes, or links.

Methodology

The build follows a read, write, verify kernel:

1. Perceive. Clone the repo immutably, inventory 318 notebooks, and strip each to lean text.
2. Analyze. codex reads one notebook per call and returns a strict JSON extraction (models, APIs, code patterns, gotchas, best practices) tagged from a controlled vocabulary.
3. Validate. Deterministic rendering plus a hard lint gate (zero dead links, zero orphans), then an independent 66-agent verification pass against the source notebooks.
4. Act. Assemble concept maps, entity pages, and best-practice guides, all cross-linked, and ship only when the gate is green.

The full extraction contract is in references/EXTRACTION-PROMPT.md; verification results are in references/VERIFICATION.md.

Sample output

A source page (excerpt): gemini/use-cases/entity-extraction/evaluate.ipynb

Evaluates Gemini-based document classification on image samples and visualizes confusion matrices.

Summary. Loads Gemini project settings from a .env file, sets gemini-2.5-flash as the evaluation model, and runs document classification evaluation through local modules. It evaluates a stratified 120-sample set, then a focused 60-sample subset, prints summary metrics, and plots ordered confusion matrices.

Models and APIs. gemini-2.5-flash (linked to its model page) and Vertex AI.

Gotchas. GEMINI_PROJECT_ID must be set or the notebook raises ValueError. Sample sizes are fixed at 120 and 60.

Best practices. Use environment variables for configuration. Set a random state for reproducible sampling. Use stratify=True for class-balanced samples. Inspect summary metrics together with confusion matrices.

Related. Concepts: Evaluation, Applied Use Cases. Entities: Gemini. Plus embedded architecture diagrams from the notebook's directory.

Browse it on the web

No Obsidian required. The entire brain is published as a fast, fully searchable static site, built with Quartz from the same wiki/ vault:

agricidaniel.github.io/google-genai-brain

Full-text search, the interactive graph, backlinks, popovers, and dark mode all work in the browser. All 686 pages publish; the site rebuilds and redeploys on every push to main via .github/workflows/deploy.yml.

Build it yourself: cd site && npm ci && npx quartz build (add --serve for a local preview at localhost:8080).

Use it

1. Open this folder as a vault in Obsidian.
2. Start at wiki/meta/Start Here.md.
3. Open the graph view. Use backlinks on any entity page to find every notebook that uses it.

For agent use and rebuild commands, see SKILL.md.

Rebuild

The committed .raw/extracted/ lets you regenerate the deterministic layer offline:

python3 scripts/render_source_pages.py && python3 scripts/build_brain.py
python3 scripts/lint_vault.py     # expect LINT CLEAN

A full re-scrape (new Google notebooks) needs the codex CLI with gpt-5.5:

bash scripts/build_all.sh

Compared to

Use cases

• GCP solutions architect. Find the right pattern (RAG, grounding, agents, evaluation) in minutes instead of opening 318 notebooks.
• Developer advocate or educator. Pull grounded, per-theme best practices and the matching architecture diagrams for a talk or workshop.
• AI engineer onboarding. Learn the Gemini and Vertex AI surface through concept maps that link straight to runnable examples.
• Team shipping a GenAI feature. Compare embedding and vector-store options and see exactly which notebooks use what before you commit.

FAQ

Is the content accurate? Source pages were adversarially verified at 95/100 with no hallucinations. The concept guides are synthesized across many notebooks and cite their sources. See references/VERIFICATION.md.

Does it need API keys or cloud access? No. It builds and reads with zero credentials and runs no cloud workloads.

How current is it? Snapshot of 2026-06-09. Re-run scripts/build_all.sh to refresh when Google pushes new notebooks.

Can I rebuild without paying for codex? Yes. The committed extractions regenerate the full vault deterministically. Only a fresh re-scrape needs codex.

What about Google's license? The upstream content is Apache-2.0 and remains Google's. This repo links back to every notebook and preserves attribution in THIRD_PARTY_NOTICES.md. Our original pipeline and synthesis are MIT.

Why Obsidian? The value is the graph: backlinks, the connection view, and local search over a fully interlinked knowledge base.

Built by

• Blog agricidaniel.com/blog
• AI Marketing Hub (free) skool.com/ai-marketing-hub
• AI Marketing Hub Pro skool.com/ai-marketing-hub-pro
• YouTube @AgriciDaniel
• GitHub @AgriciDaniel

Project info

License (MIT) · Contributing · Security · Support · Code of Conduct · Third-party notices · Changelog

Reference material only. Always verify against the linked upstream notebook and the official Vertex AI documentation before shipping production code.