fireworks-tech-graph Biggest update yet! 🎉This was originally a small open source project and has accumulated 8.5k stars after several months of development. Thank you all for your affirmation and likes. This year I will sprint for 1w star🚀. If you like it, you can give it three times in a row. Open source address: Special thanks to several friends who submitted PRs. Your high-quality PRs ensure the high level of this open source project. This version of the update is driven by the GPT 5.6 sol model. Institutions and companies interested in sponsoring are also welcome to contact me. The following is the content of this update, let’s start playing happily! This upgrade focuses on one thing: getting technical diagrams from "generated complete" to "verifiable delivery". Based on loop engineering to ensure the quality of the output, especially the previous frame-through lines, the problem of confusing layout has been greatly improved.
- Structured SVG Validator - Use XML AST to replace attribute regularization, ordinary text such as `retrieve(top_k=5)` will no longer trigger misjudgments - Completely check `marker-start`, `marker-mid`, `marker-end` references - Support absolute and relative `M/L/H/V/Q/C/S/T` path - Support group transform, curve path sampling, dash node and container distinction - Collision report directly gives the path The coordinates of the fragment and the node being traversed
- Routing & Style Contract - The order of rules is clearly: user requirements → selected style tokens → diagram layout → universal defaults - Arrow labels adopt offset-first by default, and the background color will be added only when there is still a collision - Bidirectional flow with the same semantics allows shared corridors - read/write, request/response, sync/async retain independent directions and visual encoding - The wrong waypoint of Style 6 has been fixed, and the label rules of Style 7 have been unified
- Visual Review Gate - After exporting SVG to PNG, the check must be read back - Check for frame penetration, cropping, label extrusion, legend occlusion and arrow corridor - Rerun validator and render check after each round of correction - When the running environment cannot read the image, clearly report `visual_review: skipped`
- Regression & Distribution - Validator single test: 10/10 - Python
- 9 and 3.14 both passed - 8 styles of regression: 8/8 - Style 8 adds static SVG regression fixture - Chinese and English README and 8 latest sample pictures are updated simultaneously - npm package incorporates validator helper, tests and Style 8 fixture - `__pycache__`, `.pyc` and other running caches will not enter the release package The complete process now is: Prompt → Diagram Contract → Semantic IR → Style Spec → Route Planner → SVG Build → XML / Marker Gat e → Path Collision Gate → PNG Review → Verified Assets The model is responsible for generation, and the validation harness is responsible for converging the output to a deliverable state. The first two pictures show the comparison before and after this upgrade, and the third picture is a complex Deep Research Agent Runtime architecture diagram.
