Sometimes an agent needs to look beyond your file. Rixoffice builds web search into the AI stack itself, so the composer can check facts, pull in current figures and find images without you opening a browser, copying snippets or losing your place in the document. The search becomes a tool the agent uses, not a detour you take.
What the search layer does
Search is handled by a dedicated ai-search package that every app shares. It provides two capabilities:
- Web search through Serper, which serves Google results. If Serper is ever unavailable, the package falls back to DuckDuckGo automatically, so a search keeps working instead of failing.
- Image search with the same fallback behavior, so agents can find pictures as easily as text.
Because the search layer sits under the skills, it is available everywhere: the docx skill carries web_search and image_search, the slides skill has web and image search built in, and the markdown skill includes web_search alongside its core tools. One package, one behavior, every app — there is nothing per-app to learn.
Where search shows up in practice
- RixDocs: ask for a paragraph with current figures and the agent searches, reads the results and writes the text with the facts it found.
- RixSlides: deck generation searches the web for supporting facts and images as part of the pipeline, so slides arrive illustrated, not empty.
- RixSheets: ask about context for your numbers — industry benchmarks, unit definitions — and the agent can ground its answer in search results.
- PDF Editor and Markdown Studio: look up terminology, check a claim or find a reference image without leaving the file.
The common thread is that you stay in your file. The research happens inside the request, and the output arrives as an edit or an answer in place. That is a bigger change than it sounds: the traditional loop of switch to browser, search, skim, copy, switch back, paste and reformat disappears, and with it the context loss that comes every time you leave a document mid-thought.
How agents use search results
The search tools return structured results, and the agent reasons over them the same way it reasons over your document:
- You ask a question that needs outside knowledge.
- The agent formulates a focused query and calls web_search.
- It reads the returned results and, if needed, refines the query and searches again.
- It writes the answer or edit, drawing on both your file and the search findings.
Step three is the difference between a search box and an agent: the model can iterate — narrow the query, try synonyms, follow a different angle — before it answers you. A search box returns what you typed; an agent keeps searching until the material actually supports the request.
Images, without the detour
Image search earns its place in the toolset. Ask the agent to add a suitable illustration for a section or a slide and it searches, picks a fitting result and places it with the right sizing. In documents this replaces the usual export-save-insert dance; in decks it is built directly into generation, which is why AI-generated decks come out illustrated rather than text-only.
Combined with the attachment reader, you get the full sourcing loop: your own reference files through AI attachments, the live web through search, and both woven into the same edit.
Privacy and the search path
One thing is worth knowing precisely: what reaches the search provider is the query, nothing more. Your document content is not uploaded to Google or DuckDuckGo. The agent formulates a query locally, the query is sent, results come back and the reasoning happens in your AI provider of choice. The complete picture of what leaves your machine is covered in AI privacy.
The fallback design also says something about reliability. A single search dependency is a single point of failure, so the ai-search package treats DuckDuckGo as a standing second path rather than an emergency hack. You notice it only on the rare day when the primary is down — a search takes a moment longer and still returns results.
Getting good results
Search quality follows request quality:
- Name the kind of source you want — official statistics, recent news, technical documentation.
- Let the agent iterate; if the first answer is thin, ask it to dig deeper or try a different angle.
- For images, describe subject and mood, for example: a minimal illustration of a supply chain, preferably photographic.
- Ask for sources in the output when the claim will be scrutinized — the agent can surface where its facts came from.
A good rhythm for research-heavy writing: first ask the agent what it can establish from the web about your topic, read the summary, then direct the follow-up — the questions get sharper once you know what the first pass found, and the agent keeps the earlier results in the same conversation.
One boundary to keep in mind: search extends what the agent knows, not what it can verify exhaustively. For claims that carry legal or financial weight, the agent's sources are a starting point — ask it to list them, then treat the primary source as the authority. That is how you would use a human researcher, and it is the right contract here too.
What to explore next
- Agent skills — where the search tools sit in each app's toolset.
- AI deck generation — search-driven illustration in action.
- AI attachments — combine web research with your own reference files.