Large Model Visual QA Analysis

Problem: Context Gaps in Open-Ended Image QA

Engineers often need more than image labels. They may need to read chart trends, extract key information from documents, or infer the context of a scene. Traditional CV outputs fragmented labels, while an LLM alone lacks pixel-level visual evidence. visual-qa-analysis addresses this cross-modal question answering problem by placing image understanding and natural-language reasoning in the same pipeline.

How It Works: CV Parsing + LLM Reasoning + Cloud Records

Core workflow:
- Image input: supports jpg, png, jpeg, webp, up to 20MB, via local path or public URL.
- Visual parsing: uses CV to identify objects, scenes, text, and chart data as structured cues for reasoning.
- Open-ended QA: combines an LLM for semantic understanding, logical inference, and knowledge association, without preset answer templates.
- History lookup: uses --list to fetch records from a cloud API and renders a Markdown table with reportImageUrl links.

Boundaries: Analysis, Not Final Verification

The skill relies on LLM-generated responses, so outputs are reference-only and important claims should be verified. History queries are constrained to the cloud API rather than local memory files. The script must run in the skill root directory, and open-id must be provided or validly obtained, not self-assumed.