Summary

Comprehensive fixes for VLM and RAG documentation discrepancies identified in detailed documentation review.

Changes (6 files)

1. docs/spec/vlm-client.mdx

• ✅ Update default VLM model: Qwen2.5-VL-7B → Qwen3-VL-4B-Instruct-GGUF (8 occurrences)
• ✅ Fix timeout: 300s for extraction (was 60s), clarify 60s for loading
• ✅ Document MIME type auto-detection (PNG, JPEG, GIF, WebP, BMP support)

2. src/gaia/rag/sdk.py (CODE)

• ✅ Normalize VLM default: Qwen2.5-VL-7B → Qwen3-VL-4B-Instruct-GGUF
• Ensures consistency with VLMClient and other GAIA components

3. docs/spec/rag-sdk.mdx

• ✅ Update VLM model reference: Qwen2.5-VL-7B → Qwen3-VL-4B

4. docs/reference/cli.mdx

• ✅ Fix VLM model in init profiles table: Qwen2.5-VL-7B → Qwen3-VL-4B

5. docs/guides/chat.mdx

• ✅ Fix ChatConfig assistant_name default: "assistant" → "gaia"
• ✅ Fix VLM model in PDF indexing note: Qwen2.5-VL-7B → Qwen3-VL-4B

6. src/gaia/vlm/mixin.py (CODE)

• ✅ Remove invalid Qwen3-VL-8B-Instruct-GGUF example (model doesn't exist)

Verification

All changes verified against implementation:

• src/gaia/llm/vlm_client.py:73 - Qwen3-VL-4B default ✓
• src/gaia/llm/vlm_client.py:253 - 300s timeout ✓
• src/gaia/llm/vlm_client.py:27-58 - MIME detection ✓
• src/gaia/chat/sdk.py:39 - "gaia" assistant_name ✓

Impact

• ✅ Consistent VLM model across all GAIA components (Qwen3-VL-4B)
• ✅ Accurate timeout expectations (5 min for complex forms)
• ✅ Documented MIME type auto-detection feature
• ✅ Correct ChatConfig defaults
• ✅ Removed invalid code examples

Testing

• Documentation and minor code changes only
• No breaking changes
• All updates align with existing implementation