LLM wiki systems and codebase management are so similar that everyone builds the same thing over time.
If you can, understand what you need and split up tasks to various specialized agents. Qwen 3.5 has stronger image capabilities compared to Gemma 4, but Gemma 4 has better linguistic skills. STT models for audio, too.
Something like a Ralph loop to parse your data hoard and the various models work together under a larger orchestrator model (like Kimi K2.6/2.5) that will compress the raw knowledge base down to make a persistent wiki.
If you limit yourself to one model, then just stick with the large generalist model. Just know that it will be more expensive than having smaller models take up specific since they work faster and more efficiently.