All articles

January 30, 2025 · 7 min read

Search architecture for research agents that cite their work

How query planning, source diversity, and evidence tracking turn web search into dependable research.

Reliable retrieval comes from explicit decisions about discovery, rendering, identity, evidence, and failure:not from adding more retries after launch.

Separate discovery from reading

Search snippets help choose sources; they are not evidence. A robust agent first discovers candidates, then fetches the pages it intends to cite. This makes the final context longer than a snippet but far less likely to omit qualifications or quote text that no longer appears on the source page.

Plan queries around claims

One broad query tends to reproduce the ranking bias of a single result page. Break the task into claims: definitions, current measurements, contrary evidence, and primary documentation. Add date or domain filters where freshness and authority matter. Keep the original query beside each retrieved page so an evaluator can reconstruct why it entered the context.

Make diversity explicit

Ten pages repeating one press release are one source, not ten. Deduplicate canonical URLs and near-identical text, limit results per domain, and seek primary material before commentary. For contentious questions, ask for disconfirming evidence as a separate retrieval step rather than trusting the model to notice what is absent.

Carry citations through generation

Assign stable source IDs after retrieval and require generated claims to reference them. Validate that cited IDs exist and that source passages support the sentence. Rounderone responses retain canonical URLs, titles, retrieval timestamps, and cleaned content so the application can render citations without another lookup.

PUT IT INTO PRACTICE

Test retrieval on a real page.

See the cleaned Markdown and source metadata Rounderone returns.

Open playground