DeepSeek's Full-Suite API: What It Handles vs What You Must Build for AI-Assisted Tech Selection
The article explores DeepSeek's new full-suite API capabilities—file parsing, web search, and link reading—by designing a message queue selection assistant, detailing which tasks the API handles (multi-step search, document parsing) and which require custom application logic (versioned conditions, material archiving, comparison records, citation tracking, and cost management).
TL;DR
A minimal first version needs only a project page, a backend, and storage for conditions, materials, and comparison records.
The model reads materials and supplements research; the application persists project conditions, material versions, and comparison results so follow-up questions have traceable context.
Ad-hoc research can use built-in search-and-read; to retain a specific document version for audit, the application fetches and saves the original via independent APIs.
What DeepSeek Opened
Previously, wiring a model to external tools required the application to call search services, fetch pages, parse documents, and feed results back. DeepSeek now embeds several capabilities directly into the conversation API (supported in Messages, Responses, and Chat Completions protocols for deepseek-flash):
File input : Upload candidate PDFs/Word docs; parsed content is referenced in the conversation for model-based analysis.
Web search : Look up a specific fact (e.g., which deployment modes a version supports); answers include source citations.
Deep thinking search : Multi-step reasoning with interleaved search, page opening, and in-page find (e.g., around "can historical messages be re-consumed?").
Read link : Given a known official doc URL, have the model fetch and read the page to answer the question.
Plain search and deep thinking search share the same deepseek_web_search tool; the difference is whether reasoning is enabled. Declaring deepseek_web_search lets the model decide whether to search; enabling reasoning allows multi-step search-and-read. When only a link needs reading, declare deepseek_read_url separately. These are available capabilities—not guaranteed to run every step in every request.
Example: asking "Does this solution support re-consuming the last 7 days of messages for the reconciliation program?" The model may first search retention policies, then open the consumer-offset documentation. If the retention window is found but the re-consumption mechanism is unclear, it continues searching. Search, fetch, and injecting results into the model context are all handled server-side by DeepSeek. A single request can contain multiple such steps and return the final analysis with citations.
File input requires upload first: successful parsing returns a file_id that the application then references in the conversation. Parsing mode affects what the model sees. Text-only parsing provides only text; mixed text-image parsing includes recognized images. If critical topology lives only in diagrams, text-only parsing loses it. Even mixed parsing may miss images; key diagrams still need manual cross-check against the original page.
Besides letting the model use these tools inside a conversation, developers can also call the underlying APIs independently : Search for research, Fetch for page content, OCR for scanned pages, ASR for speech-to-text, TTS for text-to-speech. Independent Search and Fetch were opened on 2024-10-17; the 2024-11-13 update integrated the search-and-read flow into the conversation API. They can be mixed into the application as needed.
For the selection assistant, both integration styles apply: use built-in search-and-read for ad-hoc fact-finding; use independent APIs to fetch and archive a specific document version as review evidence.
Addy Osmani's April 2024 "Agent Harness Engineering" discusses a similar division: reuse ready-made tool loops and context management, focus effort on the task-specific logic. Here, search-and-read is already provided; the next step is wiring project conditions, materials, and comparison records together.
Build a Minimal Application
Define a concrete selection scenario: Order service in Java; order state changes emit events consumed by Warehouse, Notification, and Reconciliation. Reconciliation must re-consume the last 7 days of messages after a fix. Two-person ops team; cloud-managed service not yet decided.
Two candidate packs: one self-hosted (topology + deployment steps), one cloud-managed (retention period + SLA). Self-hosted lacks disaster-recovery drill records; cloud docs don't answer how existing consumers re-read history.
UI centers on a single selection project: editable conditions at top, candidate materials in the middle, comparison results and open questions at bottom, chat box for follow-ups. Confirmed conditions and current conclusions always have a fixed, visible place.
When entering conditions, "retain 7 days" must also note the purpose: reconciliation needs re-consumption. A single historical archive may not satisfy that. Just as the earlier article "From RAG, MCP, Skills to Ontology First" clarified business metrics (contract amount, invoice amount, recognized revenue), here the purpose must be explicit so later comparisons share a common basis.
Backend orchestrates each request: save the round's inputs, upload required documents to Files, reference them in the model call, enable built-in web tools when public facts are needed. The returned analysis is kept as a draft; after user confirmation or amendment, it is saved as a selection record.
Responsibility split:
Project page : Edit conditions, pick materials, ask follow-ups, view and save comparisons.
Application backend : Assemble the round's input, call APIs, handle success/failure, persist records.
Database & file storage : Store condition versions, material originals, sources, conversation history, comparison results.
DeepSeek API : File parsing, model reasoning, and any search-and-read needed for this round.
This responsibility split lets a first version run with one backend, one database, and one file store.
The official chat demo is an even lighter starting point: pure static frontend, browser calls API directly, session and attachments stored in the browser. Good for personal prototypes—study its file upload, reference rendering, and streaming handling. For multi-user, move project records and the shared API key to the backend.
With these records, reopening the project lets the team continue adding materials or changing conditions without re-assembling context from chat logs.
Make Comparisons Concrete
First request can be specific:
Compare self-hosted and cloud-managed against confirmed project conditions. Focus on: can reconciliation re-consume last 7 days' messages; what ops work does a two-person team own. Prioritize official docs for the exact versions. Distinguish what materials state vs. what is inferred; list gaps as open items.
Because this round crosses deployment, replay, and ops, try deep thinking search. Include project conditions and product versions in the request to reduce the chance of the model pulling docs for a different version.
First round doesn't score options; it lists what materials already answer and what is missing. Using the hypothetical scenario, records can be organized as:
Historical replay : Cloud docs state retention period; consumer re-consumption mechanism still open. Both options need duplicate-handling impact check.
Storage capacity : Missing daily message volume and avg size; cannot estimate 7-day space yet.
Ops workload : Self-hosted has deployment steps, lacks recovery drills; cloud availability undecided.
"Can replay" needs another layer: which consumers are triggered? Replaying only reconciliation vs. also re-running Warehouse and Notification have different impacts. Messages being readable again does not equal business-safe re-execution. Consumer offset reset answers "how to read again"; whether consuming apps can detect duplicates and avoid double inventory deduction or duplicate notifications depends on application code. The assistant can list these verification items; message queue docs only support part of the judgment.
Capacity suits deterministic code. Assuming 10M msgs/day, 1 KB avg payload, decimal units: 7 days ≈ 70 GB; with 3× replication ≈ 210 GB. Indexes, metadata, headroom, and compression not included—cannot be used directly for disk sizing.
Model helps find parameters and explain omitted factors; formulas and computed results are generated by code and saved with the record. When volume changes, re-run the same formula for easy audit.
Citations must sit next to the specific judgment. The built-in tool's cited_text is the snippet marked in the answer, not the full source excerpt. The UI can show source links; to show "what the original actually said" the application must fetch the original. Even cited conclusions need version and original-text verification.
When Conditions Change
User adds: "Message queue must run in our own data center."
Application writes the change back to the condition area, confirms it as a new version, then triggers a new comparison. Cloud-managed drops out; self-hosted still needs replay and disaster-recovery verification. Only one candidate left does not mean verification is done.
New record explicitly states: "Deployment location changed from undecided to on-premise; cloud-managed excluded; 7-day replay requirement unchanged; self-hosted recovery plan still pending." Colleagues reviewing later know cloud was excluded by a condition change, not by technical inferiority.
Condition changes can also happen while a previous generation is still in flight. Record the condition version and material list at request time to handle this race. Each result belongs to the condition version at request time; a late-arriving old response must not overwrite the newer condition's result. Old results remain queryable; new comparison saved as a new version. First version can manage this with request logs and version numbers.
Also distinguish "received some text" from "this round's analysis is complete." The official demo's Messages streaming waits for message_stop; early termination signals interruption. The UI can show partial analysis marked incomplete; before promoting to a full draft, verify protocol ended cleanly and required content is present. Existing selection records stay intact.
Next follow-up must give the model the materials it needs. After DeepSeek explicitly reads a link, the page body is not returned in plain text; success returns a signature. Subsequent turns must echo that signature per protocol so the server can restore the page context. In Chat Completions, the assistant message's server_tool_calls must also be preserved; if an SDK or gateway strips them, the request may succeed but previously read pages drop out of context.
Plain search results and pages opened during deep search are not automatically retained for later turns—their retention rules differ from explicit link reading.
As discussed in "Agent Harness Context Management: Chat History or Working Set?", the visible chat history and the actual materials the model receives this turn are two different things. For this follow-up, the new condition must be included, and the self-hosted recovery steps' sources must remain available. The application must save a human-reviewable full record and select the actual payload sent to the model each turn.
Which Materials to Manage Yourself
Next step: team prepares to review self-hosted; product version is pinned; need to archive the exact retention policy, replay method, and recovery docs for that version.
At review time, you must cite which documents were used and be able to retrieve that exact version later.
Current built-in search-and-read is model-driven: no tool_choice to force invocation, no allowed_domains to restrict sources, and fetched full text is not returned to the application. Simply prompting "please read these official links" cannot guarantee the application has obtained and stored the specified materials.
Therefore, the assistant will handle this material on the application side: known links → direct Fetch; unknown links → independent Search first. Search with need_content: true returns body text; if sufficient, skip Fetch; if empty or missing key sections, re-fetch or try alternate source. summary and snippet are summaries—store them separately from full text.
After fetching, assign each material an internal ID, record product version, source URL, fetch timestamp, and link comparison items to that material. When the official site updates, the old selection record still points to the saved snapshot; re-evaluation explicitly selects a new material version.
Files have a similar boundary: currently uploaded parsed documents do not support downloading the original file or the parsing result. Hence the assistant saves the original before upload, mapping its internal material ID to the file_id. Later it can show the original page, re-parse, or switch models—all from the original.
When scanned pages need conversion to searchable text, call OCR independently. The application saves the OCR output and feeds it into subsequent requests. If critical content lives in architecture diagrams, cross-check arrows and hierarchy against the original page—text recognition completeness does not guarantee relationship accuracy.
Application-managed materials also mean handling fetch failures, missing sections, and version upgrades. Start by applying this to review-grade evidence; ad-hoc fact-finding still uses built-in search-and-read.
Scale by Task
First version has only a few short packs; reference Files directly. As material grows, let the user pick which packs participate in this comparison. file_id is the reference entry. Per current docs, documents referenced in multi-turn history stay fully in context; exceeding the limit returns an error—the API does not auto-select relevant passages. Parsing incurs a per-page fee once; subsequent references are billed by actual input tokens consumed.
Karpathy's June 2025 X post on context engineering emphasizes preparing the right information for the model's next step, and how the application orchestrates control flow and post-generation verification. Applied here: if the next step only verifies self-hosted replay, send current conditions + retention policy + consumer-offset docs; archive the old cloud pack.
After filtering by project, version, and question, if material is still large, add chunking and retrieval. For multi-user, material IDs and file_id s must map to each member's access scope; switching API keys under the same account does not automatically isolate business users' files.
If analysis must continue after the user leaves the page, or resume after failure, then introduce a background task queue. Voice can be added on demand: dictated conditions, once confirmed, enter the same project record. Current ASR accepts up to 120 seconds of complete audio, no streaming—real-time voice interaction needs separate design.
Cost is tracked per complete selection: model usage, parsing, search-and-read, retries grouped under one task to know total spend from first comparison through condition changes and material additions. Billing reads usage fields from responses; UI search-result block counts do not equal billable calls.
The first version's goal: produce a record that lets a colleague see why cloud-managed was excluded, what self-hosted still lacks, and which configs and application behaviors the "7-day replay" depends on. Next material addition, condition change, or new reviewer starts from that record.
References
DeepSeek: 全家桶应用 API 公告; Search/Fetch 开放公告 (https://api-docs.deepseek.com/zh-cn/news/news261113/; https://api-docs.deepseek.com/zh-cn/news/news261017/)
DeepSeek: 应用 API 发布的官方 X 原帖, 2026-11-13 (https://x.com/deepseek_ai/status/2121247896918253875)
DeepSeek: 构建 Chatbot 与联网工具, 含协议、多轮与计费规则 (https://api-docs.deepseek.com/zh-cn/guides/build_a_chatbot/; https://api-docs.deepseek.com/zh-cn/guides/web_tools/)
DeepSeek: 文件解析与 Files API (https://api-docs.deepseek.com/zh-cn/guides/document_parsing/; https://api-docs.deepseek.com/zh-cn/guides/files_api/)
DeepSeek: 独立 Search/Fetch 与 OCR (https://api-docs.deepseek.com/zh-cn/guides/search/; https://api-docs.deepseek.com/zh-cn/guides/ocr/)
DeepSeek: 模型与价格、ASR、TTS (https://api-docs.deepseek.com/zh-cn/quick_start/pricing/; https://api-docs.deepseek.com/zh-cn/guides/asr/; https://api-docs.deepseek.com/zh-cn/guides/tts/)
官方 chat-demo 源码, 阅读版本 5d5dd754ebc1948e977ddb880a9eb5aa25f37b4c (https://github.com/deepseek-ai/chat-demo/tree/5d5dd754ebc1948e977ddb880a9eb5aa25f37b4c)
Addy Osmani: Agent Harness Engineering, 2026-04-19 (https://addyosmani.com/blog/agent-harness-engineering/)
Andrej Karpathy: 关于 context engineering 的 X 帖子, 2025-06-25 (https://x.com/karpathy/status/1937902205765607626)
Signed-in readers can open the original source through BestHub's protected redirect.
This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactand we will review it promptly.
How this landed with the community
Was this worth your time?
0 Comments
Thoughtful readers leave field notes, pushback, and hard-won operational detail here.
