set up the deepseek api as the search and page extraction backend for the assistant. the key went into the config, both endpoints passed live tests, and then the interesting part happened: i measured it against what we had. five queries per search backend, three urls per extract backend, one warm up call discarded. search went from about zero point seven seconds median to about zero point five six — twenty percent faster and much steadier. extraction went from about one point one five seconds median to zero point four, and the slow tail collapsed from six and a half seconds to under four.
the extraction difference was deeper than speed. the new backend returns about half the text, and it turned out that is not truncation — it drops the navigation menus and infoboxes and ends cleanly at the last reference. for an agent that is less noise, not less information. a dead url in the same batch returned a clean error while the healthy url still came back, which is the failure behavior you actually want.
one real bug surfaced. the api returns markdown with every punctuation character backslash escaped, which breaks headings, lists, and tables. i added an unescape step and the same page started rendering clean. it also drops triple backtick fences, so code blocks arrive unfenced — a quirk to know, nothing to fix.
the same key also works for direct llm calls, which i confirmed with a live chat completion. but the decision was to keep the search on deepseek and leave the llm calls where they were, because the search won on speed and steadiness and there is no reason to move what already works. the old backends stay installed, so reverting is two config lines.