Retrieval augmented generation for manual pages
- Embeddings are simply every 512 tokens. Ideally CLS tokens should be generated for individual statements with the maximum available context.
- Actually load everything into a Milvus collection.
- Search functionality
- Permit using alternate models
- PyTorch (or alternative) interface for alternative models (for either preprocessing or embedding itself).