Question 1

How does the dual projection model maintain semantic consistency between text and graph representations?

Accepted Answer

The key insight is that text and mind-map are not truly independent representations, they are two projections of the same semantic structure. We defined a canonical internal model that captures both the linear ordering of text and the hierarchical nesting of concepts. Each user action, whether typing text or dragging a mind-map node, is first expressed as an operation on this canonical model, then projected to both views. The projection functions are designed to be commutative with concurrent operations: if two users simultaneously restructure text and rearrange mind-map nodes, both projections converge to the same consistent state. The hardest part was handling edge cases where a text edit changes the document structure, like adding a new heading that creates a mind-map node. These require intent analysis to determine whether the user meant to create a new structural element or simply typed a particular character sequence.

Question 2

What was your approach to conflict resolution when multiple users edit different views simultaneously?

Accepted Answer

We implemented a multi-layer conflict resolution strategy. At the CRDT level, Yjs handles basic conflict resolution for concurrent character insertions and deletions. But our dual-view system introduces higher-level conflicts, for example, user A moves a section in the text view while user B rearranges the same content in the mind-map. For these structural conflicts, we built a semantic merge algorithm that analyzes the intent behind each operation. If intents are compatible (both users trying to reorganize for clarity), the system merges them with a deterministic priority rule. If intents genuinely conflict (one user promoting a section to a top-level concept while another is deleting it), the system preserves both changes and surfaces a visual conflict marker that users can resolve interactively. In practice, fewer than 2% of real-world concurrent edits result in true conflicts requiring user intervention.

Question 3

How did you implement undo/redo that works correctly across both views and multiple users?

Accepted Answer

Multi-user undo in a CRDT system is notoriously difficult because undoing one user's operation should not undo another user's concurrent changes. We implemented a selective undo model based on operation inversion: each operation records enough information to compute its inverse, and undoing applies this inverse operation through the same CRDT merge mechanism. The added complexity with dual views is that an undo might have different visual manifestations depending on which view you are looking at, undoing a text edit might appear as a node deletion in the mind-map. We maintain per-user operation stacks that track operations in canonical model terms, so undo/redo is view-independent. The user always sees a consistent and intuitive result regardless of which view they or their collaborators were editing in.

Question 4

What performance optimizations were critical for maintaining real-time responsiveness?

Accepted Answer

Three optimizations were essential. First, incremental projection: instead of recomputing the full text-to-graph or graph-to-text projection on every change, we compute only the delta projection for the affected region. This reduced projection latency from O(n) to O(log n) for most operations. Second, lazy materialization: the mind-map view only fully computes visible nodes and their immediate neighbors, deferring deep subtree computation until the user expands or navigates there. Third, operation batching with a 16 ms coalescing window, rapid keystrokes are batched into a single CRDT operation, dramatically reducing both network traffic and merge computation. We also implemented an awareness protocol extension that shows real-time cursor positions, selections, and activity indicators across both views with sub-50 ms latency.

Operational Transformation Algorithms

The Core Problem

Dual Projection Model

Offline-First Architecture

Follow Up Questions

How does the dual projection model maintain semantic consistency between text and graph representations?

What was your approach to conflict resolution when multiple users edit different views simultaneously?

How did you implement undo/redo that works correctly across both views and multiple users?

What performance optimizations were critical for maintaining real-time responsiveness?

Related Blog Posts

A2A and the Real Problems of Enterprise Agentic Systems

Next
Challenge

Agent Framework

The Core Problem

Dual Projection Model

Offline-First Architecture

Follow Up Questions

How does the dual projection model maintain semantic consistency between text and graph representations?

What was your approach to conflict resolution when multiple users edit different views simultaneously?

How did you implement undo/redo that works correctly across both views and multiple users?

What performance optimizations were critical for maintaining real-time responsiveness?

Related Blog Posts

A2A and the Real Problems of Enterprise Agentic Systems

NextChallenge

Next
Challenge