Evidence-Backed Web Research
Use when / do not use when
Section titled “Use when / do not use when”Use this workflow for current, factual research that benefits from linked evidence, source comparison, or verification. Do not use it for a simple answer already supported by stable local context, private data, or requests for unverified speculation.
Inputs to resolve
Section titled “Inputs to resolve”Identify the exact question, relevant geography, time window, desired depth, and whether the user needs primary sources, a recommendation, or a neutral evidence brief. Ask only for a missing constraint that would materially change the search.
Tool workflow
Section titled “Tool workflow”The upstream names below are capability labels. Select the currently advertised MCP tool whose description identifies that upstream capability.
- Turn the question into one precise
web_search_exaquery. Add a second contrasting query only when the first results lack primary sources or an important viewpoint. - Inspect titles, dates, publishers, and snippets before fetching anything. Prefer current primary sources, official documentation, research papers, or direct records; include a reputable independent source when interpretation matters.
- Use
web_fetch_exaonly for the strongest pages needed to support the answer. Normally fetch three to six sources, not every result. - Reconcile publication dates, the date an event occurred, definitions, measurement methods, and material disagreements.
Validation and fallback rules
Section titled “Validation and fallback rules”Do not treat snippets as proof when the full page is available. Separate facts supported by sources from inference. If a page cannot be fetched, replace it with another authoritative source or state the limitation. If sources conflict, describe the conflict and why one may be more applicable rather than silently choosing.
Output contract
Section titled “Output contract”Lead with the answer. Attach a direct link near every important current claim, then give the smallest useful evidence summary. Include dates and units, note meaningful uncertainty, and identify any inference explicitly.
Safety and side effects
Section titled “Safety and side effects”This workflow is read-only. Do not submit forms, bypass access controls, expose private search context, or reproduce long copyrighted passages. Keep quotations short and necessary.