Query-time web search

When an agent needs current information, it sends the question to Tavily and receives ranked web results scoped to the topic. The agent reads those results, keeps only sources that meet the relevance rule your team approved, and writes a short grounded summary back into the workflow that requested it. Freshness and domain filters follow the settings you configure. When results conflict, look thin, or come from sources outside your allowed list, the agent holds the answer and passes the query to a person rather than guessing at the truth.

Cited source retrieval

Every answer Tavily returns carries the links it drew from, so claims stay checkable. A Beam agent attaches those citations to whatever it produces, whether a research note, a support reply, or a draft, following the citation format your team set. It reads each source title and snippet, keeps the ones that actually support the claim, and records them alongside the text. If a statement has no supporting source, or the only source is low quality, the agent marks that line and hands it to a reviewer before anything ships.

RAG context assembly

Retrieval augmented generation needs clean context, and Tavily supplies passages matched to a query. An agent gathers the top passages, trims them to the length your prompt allows, and feeds that context into the model call it makes next. It applies your approved rules on which domains count and how many sources to include. When Tavily returns little usable material, or the passages contradict one another, the agent stops and asks a person to confirm the direction instead of building an answer on weak ground.