VidXP already has an optional local language-model layer through Ollama. It currently turns a question into a small search plan and writes an answer from cited results, but the plan is deliberately limited.
Extend that existing path so it can understand more useful video requests without becoming required for indexing or ordinary search.
Examples include:
- Search one video, a collection, or the whole repository.
- Find what happened before, after, or during another event.
- Ask for the first match, the best matches, or every occurrence.
- Break a complex question into a few supported searches.
- Request a board, exact frames, or clips as the evidence.
The model may only use capabilities VidXP actually exposes. Unsupported plans should fall back to normal search or explain what cannot be done. Record which local model and planning rules produced a plan so behavior can be reproduced.
Generated names, biographies, or descriptions must remain suggestions backed by video evidence and require user review. Evaluate successful plans and final retrieval results through #76 rather than judging writing style alone.
VidXP already has an optional local language-model layer through Ollama. It currently turns a question into a small search plan and writes an answer from cited results, but the plan is deliberately limited.
Extend that existing path so it can understand more useful video requests without becoming required for indexing or ordinary search.
Examples include:
The model may only use capabilities VidXP actually exposes. Unsupported plans should fall back to normal search or explain what cannot be done. Record which local model and planning rules produced a plan so behavior can be reproduced.
Generated names, biographies, or descriptions must remain suggestions backed by video evidence and require user review. Evaluate successful plans and final retrieval results through #76 rather than judging writing style alone.