- Recommend BF16 mmproj for Gemma4 E2B and E4B models.
- Note known degraded audio performance with other quantizations.
- Add reference link to the relevant llama.cpp PR/issue comment.

Signed-off-by: JamePeng <[email protected]>

- Wrapped the existing Qwen3-VL image loading example in a `<details>` block to improve README readability and save vertical space.
- Introduced a complete, production-ready "Omni MultiModal" example demonstrating simultaneous Vision and Audio processing using the `Gemma4ChatHandler`.
- Added a universal `build_media_payload` helper function to dynamically route and encode local files into OpenAI-compatible `image_url` and `input_audio` payload structures.
- Added crucial documentation clarifying multimodal capability differences across Gemma-4 variants (E2B/E4B supporting full audio/vision vs. 31B/26BA4B supporting vision only).

Signed-off-by: JamePeng <[email protected]>

Signed-off-by: JamePeng <[email protected]>

…mplate fixes

- Add `preserve_thinking` parameter to optionally retain `<think>` reasoning
  blocks across all historical conversational turns (defaults to False to save tokens).
- Improve template robustness by adding an `is defined` safety check for `enable_thinking`.
- Simplify JSON serialization logic for tool call arguments in the Jinja template.
- Update class docstring to explicitly indicate support for Qwen 3.5 and Qwen 3.6 models.
- Include `preserve_thinking` state in verbose processing logs.

Signed-off-by: JamePeng <[email protected]>

….35 (Node 24 runtime)

- Create docs/wiki/ directory with full folder structure
- Add SCHEMA.md, index.md and contributing guidelines
- Set up core/, features/, examples/, types/ and subdirectories
- Prepare for LLM-powered living documentation (Llama class, multi-modal chat handlers, vision/audio examples, etc.)
- Include .gitkeep files to preserve empty directories

This lays the foundation for a modern, maintainable wiki that will replace outdated static docs.
Future commits will populate pages with up-to-date content generated from latest source code.

Signed-off-by: JamePeng <[email protected]>