DeepSeek lets developers add web search, document Q&A and voice to chatbots

DeepSeek opened the web search, document upload and link-reading features of its consumer chat app to developers on November 13. Chatbots that use its V4.1 Flash model can turn on those features with the same model name, deepseek-flash. The company also released separate interfaces for transcribing speech and reading answers aloud.

DeepSeek had already released its standalone Search and Fetch APIs on October 17. With those endpoints, the developer sends a query or URL and gets back search results or a page’s text. The new tools can run during a conversation. Once a developer enables them, the model decides when to use them and DeepSeek’s servers handle the lookup and bring the retrieved material into the conversation. That works across the Messages, Responses and Chat Completions interfaces.

In standard mode, a search can produce an answer with citations. If a user includes an image, the model can use it to search for images. With reasoning enabled, the model can run several searches, open results and look for details inside pages as it works through a question. A separate link-reading tool handles URLs that the user supplies. A chat interface can display the answer and its citations, while a program that needs the search results or page text as data can call the standalone endpoints directly.

Choosing whether a document’s figures reach the model

A PDF, Word document, PowerPoint file or image goes to the Files API first. DeepSeek processes the upload and returns an identifier. A chat app passes that identifier along with the user’s question about the document. A reference in an earlier message carries into later turns. If the conversation and its attachments exceed the model’s context limit, the API returns an error.

The upload method determines what the model receives. One mode passes text. The other detects and crops figures, photos and diagrams, then places them alongside the text in their original positions. If a user asks about a line on a chart, the relevant figure has to reach the model. DeepSeek says the automatic process can miss some figures. The crops it does extract count toward the request’s image limit and input tokens.

A single chat request can refer to up to 20 documents with a combined 8 MiB of extracted text. Repeating a reference counts again. DeepSeek also offers standalone optical character recognition for software that needs the recognition results directly. That endpoint can return positions for the text it extracts from a page.

Voice input and playback have separate interfaces

The speech recognition API accepts a completed recording up to 120 seconds long and returns the full transcript. It detects the language from the audio, with Korean among the supported languages. An app can send that transcript as a chat message, then use the separate text-to-speech API to play back the answer.

The speech synthesis service accepts a complete answer or text sent in chunks as the model generates it. In the latter case, the service can return audio while more text arrives. DeepSeek also lets developers enable basic content screening in chat requests. It examines inputs and outputs and returns a refusal indicator when it flags content.

The published rates separate model tokens from tool usage. Search and page retrieval each cost 5 yuan per 1,000 billable calls, whether the developer calls the endpoint directly or the model calls the tool. Document parsing and standalone OCR cost 1.5 yuan per 1,000 pages. The price table lists temporary discounts for speech recognition at 0.39 yuan per hour of audio and speech synthesis at 1.5 yuan per 10,000 characters.

DeepSeek published the source code for a web chat demo built with the APIs. It can import conversations exported from DeepSeek’s website, replacing the conversation history already in the demo. Its interface shows how to present search citations, attach files, accept voice input and play back answers. DeepSeek’s hosted APIs still handle the replies and document processing.