| title | User Guide |
|---|---|
| nav_order | 3 |
Everything the Bodhi Map app can do. Open app.html (the interactive map — or click Enter the map from the index.html landing) and follow along.
- Nodes are concepts, grouped into clusters (color-coded compound boxes) and laid out with a force-directed
fcoselayout for clean separation. - Edges are typed relationships. The relationship legend (in the dock) shows each verb's color and line style:
is-a— is a type ofimproves-on— improves onalternative— alternative torequires— requirescombines— combines withbuilds-on— builds on
- Navigation: drag to pan, scroll/pinch to zoom. Click empty space to deselect.
- Click a node → it spotlights, fades the rest, and opens the detail panel with: summary, deeper detail, when to use, an example snippet, its typed connections (clickable), and references.
Ranked, typo-tolerant fuzzy search over concept names and summaries (e.g. "attn" → Attention, "speculate" → Speculative Decoding). Navigate results with ↑ / ↓ and open with Enter.
Filters combine — a node shows only if it passes all active filters.
- Level filter (top bar): chips
1–5toggle learning depth, from Foundations (1) to Advanced/Alignment (5). Each chip's tooltip names the level. - Cluster filter (legend dock): the legend doubles as a filter. Click a cluster's color swatch to show/hide that cluster (it dims with a strikethrough when hidden); click the cluster name to zoom to it. Use this to pare the map down to your area of interest.
- ✦ New (top bar): spotlights concepts added in the current release together with their direct neighbors, so you can see how the latest additions wire into the existing map. The count on the chip shows how many are new. (Hidden automatically if a release adds nothing.)
- 🧭 Which technique? — a decision wizard. Answer a few questions about your data, compute, and goals; it recommends the right rung of the adaptation ladder (RAG vs prompting vs LoRA/QLoRA vs full fine-tuning vs DPO/RLVR) and opens it in the graph.
- ▶ Guided Path — an animated walkthrough of a real workflow: base model → dataset prep → SFT → LoRA → DPO → eval → quantize, with a note at each step.
Side-by-side matrices for the families people most need to choose between — e.g. PEFT methods (LoRA · QLoRA · DoRA · AdaLoRA · Adapters · Prefix/Prompt/P-tuning · IA³ · BitFit), Preference optimization (PPO · DPO · IPO · KTO · ORPO · SimPO · GRPO), adaptation strategies, quantization, and attention variants. Click any row to jump to that concept in the graph.
One click downloads the entire knowledge base as a single LLM-ready markdown file (llm-bodhi-knowledge.md): every concept with its type, summary, detail, when-to-use, example, typed relations, and references. Drop it into an LLM/RAG pipeline as grounding context — the agent-ready payoff of the OKF backing.
- Light / dark theme — toggle top-right or press
t; the preference persists across sessions. - Minimap — orientation in the corner; click/drag to pan, press
mto hide/show. - Fit / Re-layout — buttons to fit the graph to screen or recompute the layout.
The open concept or comparison is encoded in the URL hash — #concept=lora, #compare=peft — so any view can be bookmarked or shared and reopens exactly there.
| Key | Action |
|---|---|
f |
Focus search |
↑ / ↓ / Enter |
Navigate / open search results |
t |
Toggle light/dark theme |
m |
Toggle minimap |
? |
Help & welcome |
Esc |
Close panel / modal / flow |
The layout adapts to narrow screens; the legend dock collapses on mobile.
On phones the app opens a dedicated touch view (m.html) — a searchable, filterable card list with a bottom-sheet detail for each concept. It has full parity with the desktop tools: a tools row under the search bar opens the 🧭 Which technique? wizard, the ▶ Guided Path stepper, and the ⚖️ Compare tables (tabbed, horizontally scrollable), plus ★ Build map for your personalized, shareable view.