Newshot DeepSeek Splits Search, File Reading, and Voice into APIs
DeepSeek has newly opened up capabilities including web search, file input, link reading, safety verification, and voice: developers only need to enable parameters in the existing API to call ready-made tools or standalone interfaces, without having to build search orchestration, web scraping, file parsing, and context assembly themselves.
- Search, files, and link reading opened up as APIs
- Tool calling and context assembly handled server-side
- Supports search, fetching, OCR, ASR, and TTS
- Standalone interfaces can be used independently of the model
Content is based on the official DeepSeek API changelog; model benchmark scores are vendor baselines, and no third-party retest results have been provided by the official source.
To add web-based lookups, PDF reading, and voice interaction to a chatbot, developers previously had to separately integrate search, document parsing, web scraping, and voice services. DeepSeek has brought these capabilities into the existing API: enable the parameters, and the model can use search, files, and links, while Search, Fetch, OCR, ASR, and TTS can also be called separately. Tool calling has been pushed down to the server side, but whether search is accurate and whether the benchmark scores count still need to be verified by yourself.
Five categories of tool interfaces opened up
The focus of this release is not a new model, but splitting the external capabilities in the chat product into callable interfaces. There are five categories of standalone interfaces, and the capabilities available alongside conversation number more than five.
These capabilities are carried by deepseek-flash (the name of the chat model currently used by DeepSeek), which now supports three types of interfaces: Messages, Responses, and Chat Completions. When you need to look up information, you can add search; when processing materials, you can accept files or links; for voice interaction, you can connect voice capabilities.
Search, Fetch, OCR, ASR, and TTS — these five interfaces can be used alone or together with conversation.
Safety verification is also within the scope of this release. After developers enable the safety verification parameter, the platform reviews the input received by the model and the output it generates; when content hits a risk, the request returns a refusal indicator and no longer continues to provide a normal answer. This mechanism sits in the model interaction chain, and can be used with conversation or enabled separately by developers.
The intermediate layer that was saved — who manages it now
Search, file reading, and link reading can be used by enabling a parameter in the existing API.
For chatbot and knowledge base Q&A developers, the change lands in request configuration. After web search is enabled, the model in non-thinking mode automatically initiates web or image searches based on the question and attaches citations in the answer; file input and link reading also go through the same server-side connection, so the client no longer needs to manually organize scrape results and stuff them back into messages.
The call chain in deep thinking mode is slightly different. After the search tool is enabled, the model can autonomously search the web and read links during multi-step reasoning, without waiting for developers to prepare external materials in advance; when to continue searching and how to bring results into subsequent reasoning are decided by the model itself, while developers still call it through the existing API.
What is saved is "assembling the toolchain yourself." Developers still need to choose when to enable the corresponding capabilities and receive the results returned by the server; how search is triggered and how results enter the conversation context no longer require building an intermediate layer yourself. The DeepSeek changelog is very clear about where this work belongs: on the server side.
How are files and links handed to the model?
DeepSeek has taken over the two processing stages from file to text and from webpage to context. What developers hand over is not just an attachment or an address; the content is first extracted and scraped, then used by the model to answer.
File input supports formats such as PDF, Office documents, and images. The system first performs plain text extraction and can also do mixed text-image parsing (processing text and image information together); the results are then assembled into the conversation context (the range of information the model can reference when generating an answer) for the model to answer.
Links go through another channel. After enabling deepseek_read_url (the tool switch that lets the model read webpage links), the model automatically recognizes links in user messages, scrapes the webpage content, and then answers in combination with the context. Developers need to hand the link to the model, without first scraping it themselves and converting it into a large block of text.
How large is the search foundation, and how reliable are the results?
The DeepSeek web client and API share the same self-developed search foundation.
This foundation handles hundreds of millions of search requests from around the world every day, with an index covering hundreds of billions of webpages globally. API search connects to this index that is already used at high frequency and covers hundreds of billions of webpages, not a separate experimental tool.
Search supports multiple languages and returns relevant, authoritative, and timely results in real time. Multilingual capability determines whether it can handle materials in different languages; real-time return affects whether newly published pages can enter the results. Relevant, authoritative, and timely together constitute the quality goals stated by the official source, but they describe the retrieval mechanism, not a guarantee of the correctness of every answer.
The official source says it will filter SEO cheating, GEO manipulation, and low-quality AIGC through strict identification mechanisms, reducing manipulated or low-quality content seen by users and LLMs (large language models).
The toolchain has entered the API, but compliance responsibility remains with developers
After DeepSeek brings search, files, links, and voice into the API, what developers build less of is the toolchain; the responsibility is not reduced.
Safety verification is a step that must be handled during integration. The official documentation states that after the safety verification parameter is enabled, the server reviews model input and output, and returns a refusal indicator when risk is hit. Developers still need to handle interface feedback, logging, and user appeals accordingly, and cannot treat a single refusal as the service having completed all safety handling.
When providing services externally, the official source explicitly requires enabling the safety verification parameter. The basic capability is responsible for giving risk judgments and refusal indicators, while developers still need to undertake content safety review and complete large model registration and AI labeling as required. Search results come from webpages, and files are also uploaded by users; checks before and after content enters the model cannot rely only on the final answer.
DeepSeek has also built and open-sourced a demo client based on these APIs. It can import DeepSeek conversation history and is used to verify integration results. The demo client is only a reference implementation and cannot replace the integrator in handling real user content, refusal responses, and compliance displays.
How these open interfaces perform in real applications in terms of latency, stability, and unit task cost still depends on actual calls.
Sources: DeepSeek API changelog; the framing is official release information, in which model results are vendor baselines with no third-party retest seen, and pricing and specific calling methods are subject to the official documentation.
Glossary · Plain-language annotations appearing in this article
The full text of in-depth pieces, research interpretations, and tool tutorials, as well as all podcasts, are in the members' area.