An open, interactive knowledge graph of modern AI — from tokens & attention to LoRA, DPO, reasoning models, and agents.
128 concepts · 12 clusters · fully offline · zero dependencies · MIT-licensed.
The hard part of learning modern AI isn't finding definitions — it's seeing the connections between techniques: what improves on what, what to use when, how it all fits into a pipeline. Bodhi Map lays it out as one navigable graph — every technique a node, every relationship a typed edge — so you can see the whole landscape, pick a path, and go deep where you want.
Who's it for? Anyone learning or building with modern AI — self-learners and students, engineers ramping into LLMs / fine-tuning / agents, and educators who want a shared map to point people at.
Find it useful? ⭐ Star it to help others find it, and contribute a concept — adding one is just writing a Markdown file. Created by Sanjeev · Innovate with Sanjeev.
🔍 Explore concepts and their connections — search or click any node to see its typed relationships in the graph, plus a plain-language explainer, code, and sources.
▶ More demos — decision wizard · build-your-own-map · compare · guided path · mobile
🧭 Which technique? — a decision wizard that asks a few questions and recommends the right approach, then opens it in the graph.
🗺️ Build your own learning map — pick the topics you care about; the graph filters to your selection, saved on-device and shareable as a link.
⚖️ Compare — side-by-side matrices for the families you actually have to choose between (PEFT · preference optimization · attention variants · …).
▶ Guided Path — an animated walkthrough of a real workflow: base model → SFT → LoRA → DPO → eval → quantize.
📱 Works on mobile too — a touch-native card view with the same interactive tools (wizard, guided path, compare, build-your-own-map), installable as an app.
- 🕸️ Interactive graph — 128 concepts across 12 clusters, color-coded, with typed edges (
is-a,improves-on,alternative,requires,combines,builds-on) on a clean force-directed layout. - 🔍 Rich detail on every node — plain-language summary, deeper detail, when to use, a code snippet, its connections, and paper references.
- 🧭 Which technique? — a decision wizard that recommends the right approach (RAG vs prompting vs LoRA/QLoRA vs full fine-tuning vs DPO/RLVR).
- ▶ Guided Path — an animated walkthrough of a real workflow (base model → SFT → LoRA → DPO → eval → quantize).
- ⚖️ Compare — side-by-side matrices for PEFT methods, preference optimization, attention variants, and more.
- 🗺️ Build your own learning map — pick your topics → a personalized, saved, shareable view.
- ⬇️ For agents — one-click export of the whole knowledge base as a single LLM-ready markdown file for RAG/agents.
- ⚡ Fast to use — fuzzy search (
f), level & cluster filters, a ✦ "what's new" highlight, light/dark themes, and shareable deep links (#concept=lora). - 📱 Offline & installable — zero dependencies, works over
file://, installable as a PWA on desktop or mobile, with the wizard, guided path, and compare tables available on the touch view too.
Foundations & architecture (tokens → attention → MoE) · multimodal (ViT, CLIP, VLMs, fusion, diffusion) · the full adaptation ladder · all major PEFT methods · SFT & instruction tuning · preference alignment (RLHF/PPO, DPO, IPO, KTO, ORPO, SimPO, GRPO, RLVR, RLAIF) · efficiency (quantization, GPTQ/AWQ, distillation, FSDP/ZeRO, FlashAttention, speculative decoding, model merging) · reasoning, agents, data & evaluation.
Bodhi (बोधि) is a Sanskrit/Pali word for awakening — the moment scattered facts suddenly resolve into understanding. The 🍃 is a nod to the Bodhi tree, under which that insight is said to have dawned. The name is the goal: not to pile up information, but to help understanding click.
The motivation is simple. Modern AI moves faster than anyone can read, and the knowledge that matters is real but scattered across hundreds of papers, blog posts, and docs. The hardest part isn't finding definitions — it's seeing the connections: how LoRA relates to full fine-tuning, when DPO beats PPO, where quantization fits in a pipeline. Newcomers drown in tabs; practitioners keep a messy map in their heads. Bodhi Map is an attempt to draw that map in the open.
A few principles guide it:
- Understanding over hype — every concept earns its place with a real source and a plain-language why and when to use, not buzzwords.
- Connections are the content — techniques are nodes; the typed edges
(
improves-on,alternative,combines) are where the real insight lives. - Open & offline by default — no accounts, no CDNs, no telemetry. One folder,
open
index.html, and it works on a plane. - A path, not a pile — clustered from foundations to the frontier so you can walk from tokens to agents at your own pace.
Use it now — nothing to install: just open the live app ↗.
Install it as an app (PWA) — works offline, launches full-screen, adds to your home screen / dock:
-
Desktop: open the live site → click the install icon in the address bar.
-
Mobile: open the site → Add to Home Screen (iOS) or the Install app prompt (Android) — or scan:
Run locally: clone the repo and open index.html (landing → Enter the map) or app.html directly — no server, no build step.
The content is stored as a conformant Open Knowledge Format v0.1 bundle (Google Cloud, Apache 2.0) under knowledge/ — a directory of markdown files, one per concept, with YAML frontmatter and bundle-relative .md links forming the knowledge graph. This makes the content portable, git-shippable, and agent-ready: the same bundle that powers this visualizer can be fed to an LLM/RAG pipeline or rendered by any other OKF tool.
knowledge/ is the source of truth. The browser's js/data.js is a generated artifact — a tiny build compiles the bundle into it so the app stays single-folder and works offline (browsers can't fetch() local markdown over file://).
knowledge/ ← OKF bundle (SOURCE OF TRUTH)
index.md ← bundle root (reserved, no frontmatter)
peft/
index.md ← cluster listing (reserved)
lora.md ← one concept = one markdown file
...
okf.config.js ← controlled vocab (clusters, relations, levels) + flows (wizard, path)
scripts/okf.js ← export (data.js→bundle) + build (bundle→data.js)
js/data.js ← GENERATED — do not edit by hand
Add or edit a concept by writing a markdown file in the right cluster folder, then rebuild:
# knowledge/peft/my-technique.md
---
type: PEFT Method
title: My Technique
description: One-line essence (the node summary).
cluster: peft
level: 4
when_to_use: When to reach for it.
relations:
- improves-on:lora # creates a typed edge in the graph
references:
- My Paper|https://arxiv.org/abs/...
---
# My Technique
Deeper explanation (becomes the detail panel body).
## Example
```python
example()
```bash
npm run build # or: node scripts/okf.js build → regenerates js/data.js
npm start # or: open index.html
New clusters / relation types / wizard steps live in okf.config.js. The bundle remains valid OKF, so knowledge/ also opens in Google's reference OKF visualizer.
| File | Purpose |
|---|---|
index.html |
Landing page — the site front door |
app.html |
The interactive map — app shell + UI |
css/style.css |
Light/dark themes, typography, panel & graph layout |
fonts/ |
Manrope (woff2, 400–800) — vendored locally, no font CDN |
knowledge/ |
OKF bundle — the source of truth (concepts + typed relations) |
okf.config.js |
Controlled vocabulary + interactive flows |
scripts/okf.js |
OKF export / build tooling (zero dependencies) |
js/data.js |
Generated runtime data (do not hand-edit) |
js/graph.js |
Cytoscape rendering & interactions |
js/cytoscape.min.js + fcose deps |
Vendored graph libraries (offline) |
Full docs live in docs/ — a complete guide to using, authoring, and
extending Bodhi Map:
- Getting Started · User Guide · FAQ
- Concept Authoring · Configuration Reference · Architecture
- Contributing · Release Process · Roadmap · Changelog
Contributions — new concepts, corrections, better explanations — are welcome, and the project is set up to be self-serve:
- 🌱 Pick up an issue — start with a good first issue or a
help wantedtask; comment to claim it. - 💡 Suggest a concept or report an error — via an issue template.
- 💬 Ask or discuss in Discussions.
- 🗺️ See the BACKLOG for the roadmap.
Adding a concept is just writing a Markdown file in knowledge/ and running npm run build (which validates). See CONTRIBUTING.md and the Concept Authoring guide for the format and quality bar — the only hard rules: cite a real source, keep it accurate, and stay offline.
Bodhi Map stands on publicly published research — it does not originate the
techniques it describes, and credit for the underlying ideas belongs to the
respective authors and research teams. Every concept is grounded in its primary
source: the original paper, official blog post, or documentation. Those citations
live with the content, not buried in a footnote — each concept node shows its
References in the detail panel, and the same links are in the OKF frontmatter
under knowledge/ (200+ source links in total).
What Bodhi Map adds on top is its own contribution: the curation, hands-on experience, and practical judgment — how the pieces fit together, plain-language explanations, when to use guidance, the typed relationships between techniques, and the interactive experience (decision wizard, guided path, comparisons). The value is in the synthesis and the experience of navigating it, not in any single paper.
If you spot a missing, wrong, or better source for any concept, please open an issue or PR — accurate attribution is part of the quality bar.
Third-party software & assets (all vendored locally so the app stays offline):
| Component | Used for | License |
|---|---|---|
| Open Knowledge Format (Google Cloud) | Knowledge bundle format | Apache-2.0 |
| Cytoscape.js + fcose layout | Graph rendering & layout | MIT |
| Manrope | UI typeface (fonts/) |
SIL OFL 1.1 |
If Bodhi Map is useful to you, a ⭐ helps others discover it — thank you!
MIT © 2026 Sanjeev Azad — Innovate with Sanjeev. Use, fork, and build on it freely.
The MIT license covers this project's own code and the curation/explanations. The referenced papers and sources remain the property of their respective authors.