Increase context length when your real prompt needs more room and an answer improves when the relevant evidence fits. More context can consume more memory and slow a request. Pick the smallest setting that handles your representative jobs with enough room for the response.
Build a question with a known answer
Build a synthetic document that resembles your usual inputs. Put a unique fact near the start and another near the end, such as a fictional ticket number and its approval date. Ask for both exact values and the evidence around each. Keep the same model, tools and question as you compare runs. A long conversation also carries previous messages, so use fresh sessions for each trial.
Try one small document first to confirm the model understands the question. If it gets a short, explicit example wrong, fix the prompt or baseline model before assuming a longer window will solve the problem. Then repeat with the long document. Use a few variants of the input rather than judging reliability from one completion.
Compare two context settings
Ollama documents context settings in the desktop app and server settings. If you use its API, a request can specify options.num_ctx. Use a smaller and larger setting that both fit your machine, confirm the active runner with ollama ps, and write down which client or configuration supplied the setting. Some front ends supply their own request options.
Time the same task on both settings. Track correctness, time to useful output and Activity Monitor memory pressure while the run is active. Compare warm runs separately from the first run that loads the model. If the larger window improves answers but memory pressure rises sharply while your usual apps are open, the trade-off may be too expensive for daily work.
Choose a smaller workload when needed
A long document may need only a few relevant passages. Pull those sections into a smaller prompt and include their labels so the answer can cite where it found the information. When you truly need a cross-document answer, collect short evidence summaries per section and check the final answer against the original passages. Another option is fewer simultaneous requests if the problem appears only under load.
Retest with a realistic session that includes follow-up questions. Record the model, context setting, input version and whether it passed your checks. A good result is a setting you can explain in terms of answer quality, latency and memory on your own Mac. The examples are a proposed test, not benchmark results.