Can RAG and fine-tuning be combined?
Yes, but evaluate retrieval and generation separately and together. A task-specific study such as the query-clarification preprint suggests where adaptation might help a RAG pipeline; it does not prove end-to-end superiority on your corpus.
Is a managed platform an outsourced ML team?
No: platforms describe technical capabilities, while ownership of data, labels, testing and integration depends on the contract. Together’s platform description should be turned into a specific responsibility schedule before comparison.
Does owning an adapter remove lock-in?
No: it still depends on a compatible base model, runtime, preprocessing and licence. The exact model agreement and an actual export rehearsal matter more than an “open” label.
Should we choose the lowest token price?
Only after acceptance and operating constraints are met, and after other costs are included. Published token and GPU prices are inputs to a workload model rather than complete project prices.
How do we compare an evaluator with a development supplier?
They serve different purposes: one tests claims, while the other may implement the system. Evaluation guidance helps specify the test, but independence requires separate decision ownership and transparent commercial relationships.