On November 13, DeepSeek announced that it had integrated tools for file input, web search, link reading, and safety checks into the V4.1 Flash API. Suppose you want to build an app that lets users upload a financial report PDF and then ask, “Did the company later revise this figure?” From reading the file to checking announcements, which steps can be handled by the API, and which still need to be handled by the app itself?
【From wiring up tools yourself to having the server assemble the information】
If a team builds this workflow itself, the app typically first uses a parsing tool to extract text and tables from the PDF, then organizes a search based on the company name, reporting period, and metric, opens relevant announcements, and finally sends document excerpts and webpage content to the model together. The output of each step becomes the input for the next.
After integrating the App tools, developers can provide files and questions through Messages, Responses, or Chat Completions, and explicitly enable web search according to the interface they use, as described by DeepSeek. The server can handle file parsing, tool calls, and providing the model with context from retrieved materials for its answers. The file types listed officially also include Office documents and images.
To find an announcement with a known title, you can use ordinary web search and retrieve citations. To track changes across several announcements, the deep-thinking search described officially can retrieve and read web pages through multi-step reasoning. Application teams can choose a search method based on the task and compare the two methods’ latency and costs using actual requests.
【Citations are not enough; consider what they can prove】
If a search finds a newer announcement, the application should still make clear to readers what each of the two documents says. If a later announcement publishes figures for the next reporting period, that only proves that data for a new period is available. To confirm that the original figures were revised, check whether the announcement explicitly refers to the same reporting period and metric, and provides the revised figures. When presenting an answer, you can separately cite the original figures from the financial report and the conclusions supported by the later announcement.
Before launch, you can test with sets of ordinary PDFs, scanned documents, and files with complex tables to check whether the model misreads figures, matches data to the wrong rows or columns, or cites web pages that do not support its conclusions. Also record the time from file submission to receipt of a complete answer, as well as the cost of each completed task. According to the announcement, tool usage fees are added on top of standard token fees. Actual latency, table-parsing accuracy, and per-task cost all need to be measured using samples.
File permissions are still a decision for the application side—for example, which materials may be uploaded and which need to be de-identified first. According to DeepSeek’s integration instructions, public-facing applications must enable safety checks, while developers remain responsible for content moderation, any applicable model registration, and AI-content labeling.
For research and educational purposes only; this does not constitute investment advice.