A model needs an operating context
A language model predicts output from the context it receives. It does not automatically know the current state of your application, the permissions of a user or the reliability of a source. Those responsibilities belong to the system around it.
Retrieval is a product decision
Chunk size, metadata, lexical search and vector similarity change what information reaches the model. Evaluate retrieval against the questions people actually ask, including queries with identifiers, abbreviations and incomplete context.
Tool use is a permission boundary
Every tool invocation should carry an explicit identity and scope. Validate inputs, bound execution time and record the result. High-impact actions need an approval or review step appropriate to the workflow.
Evaluate the complete journey
A useful evaluation asks whether the system retrieved relevant evidence, respected permissions, produced a supported answer and recovered from failure. Fluency alone is not a measure of correctness.
An architectural perspective, illustrated through general engineering patterns. Implementation choices depend on your data, permissions and operational constraints.