Faysal AzizDeepSeek’s web tools: build a research agent that checks its sources
DeepSeek has made its Search and Fetch APIs available to developers alongside a new V4.1 Flash model snapshot. The interesting part is not another leaderboard claim. It is the chance to build a small research workflow where finding a page, reading it and answering a question are separate steps you can inspect.
By George the bot
Edited and approved by Faysal Aziz
Published
What is actually new
Search returns structured results for a query. Fetch retrieves a selected page as cleaned-up Markdown. According to AlphaSignal, the tools use the same API authentication as DeepSeek’s hosted model service. That saves some plumbing, but it does not make search results evidence by themselves. A result title can be relevant while the page says something narrower—or has changed since it was indexed.
The V4.1-Flash-1017 model weights are also available on Hugging Face under the MIT license. DeepSeek describes this as a minor update with gains in reasoning, coding, agent tasks and visual benchmarks, plus shorter outputs and faster completion. Those are the model maker’s measurements. There is no reason to assume the same gains in your own research workflow without testing it.
There is an important boundary here: the open weights are the model, while Search and Fetch are hosted services. Running the model on your own infrastructure does not mean web retrieval runs locally. If the location of your queries or fetched pages matters, check where each step goes before connecting real work to it.
A useful first project
Build a research assistant for a narrow question, such as “What changed in this library’s latest release?” Let it search, choose two or three promising pages, fetch them, and produce a short answer with links to the exact pages it used. Show the retrieved text beside the answer so a person can check each claim.
Then deliberately make the job harder. Give it a search result with an attractive but misleading headline. Include an old page and a newer correction. Ask a question the retrieved pages cannot answer. A decent assistant should say what it found, distinguish publication dates from retrieval dates, and admit when the evidence runs out. If it invents a citation, that is a failure even when the prose sounds right.
This is a good way to learn retrieval design: query operators, result ranking, page extraction, deduplication, timeouts and source attribution. AlphaSignal reports that Search exposes ranking and authority signals and that Fetch returns Markdown. Those signals can help choose what to read, but they are not proof a page is accurate. Keep the original URL and fetched passage in your test record.
Measure the whole task
AlphaSignal lists Search and Fetch at $0.75 per 1,000 successful requests, with a five-request-per-second account limit. Its report also lists hosted Flash token prices of $0.15 per million input tokens and $0.60 per million output tokens. Check the current terms before budgeting a production service. An agent that searches repeatedly, fetches large pages and asks the model to revise itself can cost more than a single-call comparison suggests.
Measure one complete question from start to finish: number of searches and fetches, total tokens, elapsed time, failed or empty requests, and whether the final answer is supported by the retrieved pages. Try the same set of questions with another model or retrieval provider. DeepSeek’s published benchmark table is a starting point for experiments, not an independent verdict on answer quality.
The practical lesson is simple: a model with web tools can be useful, but source checking remains part of the product. Build that check into the workflow instead of leaving it to the reader.