Tutorial Advanced Features in Practice

Tutorial: Advanced Features in Practice

This tutorial demonstrates how to combine all of Mosaic's advanced features into a single, powerful workflow. You'll start with a complex question, explore it through multiple models and branches, prune irrelevant paths, and distill the results into a synthesis.

Estimated time: 25 minutes

---

Scenario

You're evaluating database technologies for a new application. You want to compare PostgreSQL, MongoDB, and SQLite across several dimensions. This is a perfect use case for Mosaic's advanced features.

---

Step 1: Set Up RAG Context

First, upload any relevant documents you have:

1. Open the DocumentPanel (📄)
2. Upload any notes, requirements documents, or comparison articles you have
3. Enable RAG (toggle ON)

Even without documents, the RAG system is useful — this step shows how to combine RAG with the other features.

---

Step 2: Start with Parallel Debate

Let's see how different models approach the comparison:

1. Open Settings → Debate Models
2. Select 2-3 models (e.g., Mistral Large, GPT-4o, Claude Sonnet 4)
3. Click the root node
4. Press Shift+Enter (not just Enter)
5. Type:

Compare PostgreSQL, MongoDB, and SQLite for a new web application.
Consider: performance, scalability, data modeling flexibility, ecosystem, and use cases.

6. Press Shift+Enter again (the text is already in the input)

You'll see multiple response nodes fan out, one for each model:

                    ┌── [Mistral Large: comprehensive comparison]
[You: comparison] ──┼── [GPT-4o: structured comparison with tables]
                    └── [Claude: nuanced analysis with tradeoffs]

---

Step 3: Review Confidence Scores

After each response completes, look for the confidence badge (a number in the corner of each response node):

• Mistral Large: 92
• GPT-4o: 88
• Claude Sonnet: 95

Claude's response has the highest confidence. This might be the one to explore first, but let's branch from all of them.

---

Step 4: Branch from Each Model

Explore each model's response by branching:

1. Click Mistral Large's response and ask about scalability:

Can you elaborate on PostgreSQL's horizontal scaling options?

2. Click GPT-4o's response and ask about schema flexibility:

How does MongoDB's schema-less design help in rapid prototyping?

3. Click Claude's response and ask about performance:

Under what conditions does SQLite outperform the other two?

Your canvas now has three branches:

[root → comparison]
  ├── [Mistral: comparison] → [You: scalability] → [AI: response]
  ├── [GPT-4o: comparison] → [You: schema] → [AI: response]
  └── [Claude: comparison] → [You: performance] → [AI: response]

---

Step 5: Use Tendril Suggestions

As each branch response completes, look for suggestion nodes (dimmer nodes connected with dotted lines). These are AI-generated follow-up questions.

• From the scalability branch, a suggestion might be: "How does read replication work in PostgreSQL?"
• From the schema branch: "What are the trade-offs of MongoDB's document model?"
• From the performance branch: "When should you avoid SQLite?"

Click a suggestion to materialize it into a real branch. This saves you from typing follow-up questions — the AI predicts what you might want to ask next.

---

Step 6: Add a Deep Dive

Let's go deeper on one branch:

1. Click the latest node in the scalability branch
2. Ask:

Show me a real-world architecture that uses PostgreSQL with horizontal scaling

3. After the response, ask:

Write a Python script that demonstrates connection pooling with PostgreSQL
using asyncpg or psycopg2

4. Run the code using the inline code execution button (▶️)

Your scalability branch now has practical code you can test immediately.

---

Step 7: Prune Irrelevant Branches

By now, your canvas might have 8-10 nodes. Some branches may be less relevant to your goal. Let's prune:

1. Click the Prune button (🧹) in the TopBar
2. When asked for a goal, enter:

I'm trying to decide which database to use for a high-traffic web application
with flexible data requirements

3. Mosaic scores each leaf branch against this goal
4. Branches scoring below 40 are dimmed

For example, if the "when to avoid SQLite" branch scored 72 (relevant), but a "JSONB in PostgreSQL" tangent scored 15 (less relevant), the tangent would be dimmed.

The Prune Banner appears at the bottom. You can Restore All if you disagree with the pruning.

---

Step 8: Distill the Canvas

Now distill all valuable branches into a synthesis:

1. Click the Distill button (🌿) in the TopBar
2. Mosaic collects all leaf nodes and sends them to the AI
3. A new distillation node appears with a gold border
4. Gold dashed edges connect it to all source leaves

The distillation might produce:

## Database Comparison Synthesis

### Winner by Use Case
- **PostgreSQL**: Best for high-traffic web apps with complex queries
  (supported by Mistral scalability deep-dive and Claude performance analysis)
  
- **MongoDB**: Best for rapid prototyping with evolving schemas
  (supported by GPT-4o schema flexibility analysis)
  
- **SQLite**: Best for embedded/single-user scenarios
  (performance branch confirms limitations at scale)

### Recommendation
For your high-traffic web application with flexible data requirements,
PostgreSQL is the strongest candidate...

---

Step 9: Export the Result

Your work is now complete. Export the entire canvas:

1. Open Settings (⚙️)
2. Click Export
3. A JSON file downloads: mosaic-database-comparison.json

This file contains all nodes, edges, positions, bookmarks, and analytics. Share it with colleagues or archive it for later reference.

---

Step 10: Import and Continue Later

To revisit this analysis:

1. Open Settings (⚙️)
2. Click Import
3. Select your exported JSON file
4. The entire canvas restores with all nodes, branches, and positions intact

---

Pro Tips for Power Users

Tip 1: Combine RAG with Distillation

Before distilling, upload relevant documents and enable RAG. The distillation will include context from your documents, making the synthesis more informed than one based purely on the AI's training data.

Tip 2: Use Pruning Before Distillation

Pruning removes low-value branches before distillation, which means:

• Less context for the AI to process (lower token cost)
• Higher quality synthesis (only relevant content included)
• Faster distillation (fewer leaves to summarize)

Tip 3: Parallel Debate Then Compare Then Distill

The most powerful workflow:

1. Run parallel debate (3-4 models)
2. Branch from each model's response to explore specific aspects
3. Compare confidence scores to identify the most reliable answers
4. Prune branches that don't align with your goal
5. Distill the remaining branches into a final synthesis

Tip 4: Save Intermediate States

Before major operations like pruning or clearing, export your canvas:

• Settings → Export → mosaic-before-pruning.json
• This gives you a safety net to restore if pruning is too aggressive

Tip 5: Use Bookmarks as Milestones

Bookmark nodes that represent key insights or decisions. Later:

• Filter to bookmarks-only to see the "table of contents" of your canvas
• Distill only bookmarked branches for a curated synthesis
• Share exported canvas with bookmarks as a summary document

---

Complete Workflow Summary

1. Upload RAG documents (optional)
2. Parallel debate with multiple models
3. Review confidence scores
4. Branch from each model's response
5. Follow suggestion tendrils
6. Deep dive with code execution
7. Prune irrelevant branches
8. Distill into synthesis
9. Export for sharing/archiving

This workflow took ~25 minutes and produced:

• Multi-model analysis
• 3+ conversation branches
• Practical code examples
• AI-generated synthesis
• Exportable archive

In a linear chat interface, this would require multiple separate conversations, manual copy-pasting, and significant context-switching. Mosaic's canvas handles it all in one spatial view.

---

Time-Saving Shortcuts for This Workflow

Goal	Shortcut / Action
Start parallel debate	Shift+Enter (instead of Enter)
Open Settings	Click gear icon (⚙️)
Fit all nodes	Press F
Delete a node	Select it → Delete
Undo a mistake	Ctrl+Z
Bookmark important node	Right-click → Bookmark
Open search	Ctrl+F
New canvas	Ctrl+N
Collapse a branch	Right-click → Collapse
Export canvas	Settings → Export

Practice this workflow a few times. The combination of parallel debate → branching → pruning → distillation is where Mosaic's spatial canvas truly shines compared to traditional linear chat interfaces.

---

Next Steps

• Tutorial - Customizing Mosaic — Fine-tune system prompts, temperature, and theme
• Advanced AI Features — Deep dive into each feature's technical details
• Overall Architecture and Data Flow — How features interact under the hood