Architecture Diagram Generator

[ms-shashank]

Current file: app/src/lib/tools/architecture-diagram.ts
Current model: deepseek-r1-0528
Current approach: Single prompt asking for Mermaid diagram. No Mermaid syntax validation, no rendering check, no iterative correction.

Problems with current approach:

• Generated Mermaid code frequently has syntax errors (unquoted labels with special characters, invalid arrow types, broken subgraph nesting).
• No verification that the diagram actually renders.
• Single-pass cannot self-correct syntax issues.
• Complex systems produce overly flat diagrams without meaningful grouping.

Upgrade plan:

Step	Agent	Action
1	System Decomposer	Analyze the system description. Identify components, services, databases, external systems, and their relationships. Output a structured component graph.
2	Diagram Generator	Convert the component graph into Mermaid syntax. Apply grouping (subgraphs), proper arrow types, and descriptive labels.
3	Syntax Validator	Programmatic: Parse the Mermaid code using a Mermaid parser or CLI (mmdc from @mermaid-js/mermaid-cli). Catch syntax errors before returning to user.
4	Repair Agent	If syntax validation fails, feed the specific errors back and regenerate only the broken sections. Max 3 retries.

• You are free to enhance the agents stacks in the above plan layout, the above one is just for reference. You can enhance more if needed.

Model suggestions to start with:

• Step 1: Try deepseek-r1-0528 for system analysis (reasoning task). Also try glm-5 for complex systems architecture.
• Step 2: Try qwen-3-coder-30b or kimi-k2.6 for Mermaid code generation (structured code output).
• Step 4: Try deepseek-v3.2 for quick syntax repair (lightweight).

Model Selection Guidance

• You are free to pick any model from the Oxlo catalog based on your own testing and evaluation.
• The Models suggestions above, not mandates. Try them first, and if they do not meet the accuracy target, experiment with alternatives.

Compare against: GPT 5.3 Thinking and Claude Sonnet 4.6 Thinking.

Acceptance criteria:

• Generated Mermaid code must pass syntax validation in 95%+ of cases (after repair loop).
• Diagrams must render correctly when processed by Mermaid CLI.
• Complex systems (5+ components) use subgraph grouping.
• Overall quality matches or exceeds GPT 5.3 Thinking & Claude Sonnet 4.6 on test cases.
• Overall accuracy at 80%+.