SEC filings for AI agents need a retrieval contract面向 AI Agent 的 SEC 文件需要明确的检索契约
An agent can retrieve public filings without an EDGAR API key, but “public” does not mean unconstrained. The SEC asks automated clients to identify themselves with a declared User-Agent and currently limits request rates to no more than ten per second in aggregate. The system should cache responses, use bulk files for large jobs and retain the source URL.
Agent 可以在没有 EDGAR API key 的情况下获取公开文件，但“公开”不代表可以无限请求。SEC 要求自动化客户端通过 User-Agent 声明身份，目前聚合请求速率不得超过每秒 10 次。系统还应缓存响应、大批量任务使用 bulk 文件，并保留来源 URL。
Choose the EDGAR surface that matches the task按任务选择正确的 EDGAR 数据入口
The strongest architecture does not treat EDGAR as one endpoint. It selects a source according to whether the agent needs filing discovery, primary documents, structured facts or large-scale ingestion.
可靠的架构不会把 EDGAR 当成单一接口，而会根据 Agent 需要发现文件、读取原文、提取结构化事实还是进行大规模采集来选择数据源。
Keep three clocks separate: the fiscal period described, the filing date shown in indexes and the accepted timestamp recorded by EDGAR. A document may appear minutes after its timestamp under normal conditions, but availability is not a guaranteed real-time feed. For historical analysis, store when your own system first retrieved the record as well. That extra clock prevents a later backtest or agent evaluation from assuming evidence was available earlier than it actually was.
还要分开记录三种时间：文件描述的财务期间、索引中的 filing date，以及 EDGAR 的 accepted timestamp。正常情况下文件可能在时间戳后数分钟出现，但这并不是有保证的实时数据流。历史分析还应保存本系统首次获取记录的时间，避免后续回测或 Agent 评估错误假设证据更早可用。
Submissions API
Use a CIK to discover recent forms, accession numbers, filing dates and primary documents. Preserve the returned metadata before ranking documents.
用 CIK 发现近期表单、accession number、申报日期和主文档；排序前先保存返回的元数据。
EDGAR Archives
Retrieve the filed HTML, exhibits and filing directory. Narrative disclosures, footnotes and risk language live here and cannot be replaced by numeric facts alone.
读取已申报 HTML、附件与文件目录。叙述性披露、附注和风险语言都在这里，不能只靠数值事实替代。
Companyfacts / Frames
Use XBRL facts for normalized values, units, periods, forms and filed timestamps. Validate taxonomy, context and amendments before comparison.
利用 XBRL facts 获取规范化数值、单位、期间、表单与 filed 时间；比较前验证 taxonomy、context 和修订。
Bulk ZIP files
For broad ingestion, use the SEC’s nightly submissions and companyfacts archives rather than repeatedly crawling per-company endpoints.
大规模采集应使用 SEC 每晚更新的 submissions 与 companyfacts 压缩包，而不是反复抓取单公司接口。
Official references: EDGAR application programming interfaces, SEC developer resources and fair-access guidance, and accessing EDGAR data.
Build a seven-stage SEC filing RAG pipeline构建七阶段 SEC 文件 RAG 管道
Resolve the issuer解析发行人
Map name or ticker to CIK, then verify legal name and exchange. Ticker mappings can change; the CIK is the stable EDGAR identity.
把名称或 ticker 映射到 CIK，并核验法定名称和交易所。ticker 会变化，CIK 才是稳定的 EDGAR 身份。
Define the filing scope定义文件范围
Specify form types, filing-date window, report period, amendments and exhibits. Do not silently mix annual, quarterly and event disclosures.
明确表单类型、申报日期区间、报告期、修订与附件，不要默认混合年度、季度和事件披露。
Fetch politely合规获取
Declare an organizational User-Agent, rate-limit globally, retry with backoff, cache immutable filings and log request status.
声明组织 User-Agent，全局限速，退避重试，缓存不可变文件并记录请求状态。
Preserve originals保留原始证据
Store the accession number, source URL, primary HTML, relevant exhibits, retrieval time and content hash before parsing.
解析前保存 accession number、来源 URL、主 HTML、相关附件、获取时间和内容哈希。
Parse by section按章节解析
Segment 10-K and 10-Q items, tables, footnotes and exhibits while keeping headings, anchors and page context attached to every chunk.
按 10-K/10-Q Item、表格、附注与附件分块，并让每个 chunk 保留标题、锚点和页面上下文。
Retrieve with metadata带元数据检索
Filter by CIK, form, filed timestamp, period and amendment status before semantic ranking. Retrieve neighboring chunks for context.
语义排序前先按 CIK、表单、filed 时间、期间和修订状态过滤，并取回相邻 chunk 补充上下文。
Answer and verify回答并验证
Require section-level citations, distinguish quoted facts from inference, surface conflicts and route consequential conclusions to human review.
强制章节级引用，区分原文事实与推断，暴露冲突，并将高影响结论交由人工复核。
Teach the agent what each filing can and cannot prove让 Agent 理解每类文件能证明什么、不能证明什么
Make every answer traceable to a filing section让每个回答都能追溯到具体文件章节
Minimum evidence record最小证据记录
issuer_cik · legal_name · form · accession_number
filed_at · accepted_at · report_period · amendment
document_url · section · anchor · excerpt_hash
Separate fact, calculation and inference区分事实、计算与推断
A filing may state revenue; a margin is a calculation; a claim that pricing power improved is an inference. Label each layer, expose the formula and cite the exact filing text or XBRL fact used.
文件可以披露收入；利润率是计算结果；“定价权增强”则是推断。应标记每一层、展示公式，并引用使用的准确原文或 XBRL fact。
Resolve conflicts before synthesis综合前先解决冲突
Prefer the correct issuer, period and latest applicable amendment, but retain conflicting evidence. A later filing may update prior disclosure without erasing what was known at the earlier date.
优先选择正确发行人、期间与适用的最新修订，同时保留冲突证据。后续文件可以更新早期披露，但不能抹去早期时点已知内容。
Treat filing text and exhibits as untrusted input将文件正文和附件视为不可信输入
SEC-hosted content is authoritative as a filing record, but it is still external content supplied by filers. An AI agent must never follow instructions embedded in a document, reveal secrets, change tool permissions or execute links merely because the text appears in EDGAR.
SEC 托管内容作为申报记录具有权威性，但它仍然是申报方提供的外部内容。AI Agent 不能因为文字出现在 EDGAR 中，就执行文档内嵌指令、泄露密钥、改变工具权限或自动访问链接。
Sandbox parsing沙箱解析
Disable active content, scripts and arbitrary file execution. Normalize HTML while retaining evidence anchors.
禁用活动内容、脚本与任意文件执行；规范化 HTML 同时保留证据锚点。
Bound tool authority限制工具权限
Retrieval tools should be read-only, domain-scoped and unable to access unrelated credentials or systems.
检索工具应只读、限制域名，并且无法访问无关凭证或系统。
Validate outputs验证输出
Check every citation, number, unit, period and issuer against stored evidence before release.
发布前逐一核验引用、数字、单位、期间和发行人是否与已存证据一致。
Escalate impact高影响升级
Legal, compliance and investment conclusions require qualified human review. A fluent answer is not proof of completeness.
法律、合规与投资结论需要合格人员复核；表达流畅不代表信息完整。
Use QVeris to govern SEC and financial-data tools使用 QVeris 治理 SEC 与金融数据工具
Authorized discovery授权发现
QVeris can help an agent select approved SEC, FMP and research tools rather than inventing endpoints or mixing providers silently.
QVeris 可帮助 Agent 选择已批准的 SEC、FMP 与研究工具，避免编造接口或静默混用数据提供商。
Bounded execution受控执行
Policies can constrain domains, methods, dates, quotas and credentials while logging which tool produced each artifact.
策略可以约束域名、方法、日期、配额和凭证，并记录每个产物由哪个工具生成。
Provenance handoff来源交接
The workflow can preserve filing metadata and evidence links for the model and reviewer. QVeris does not certify completeness or provide legal or investment advice.
工作流可把文件元数据与证据链接交给模型和审核者。QVeris 不保证完整性，也不提供法律或投资建议。
Continue with QVeris guides on financial RAG, financial statement analysis, and what a 13F filing shows.
继续阅读 QVeris 指南：金融 RAG、财务报表分析以及13F 文件包含哪些信息。
SEC filings for AI agents: common questions面向 AI Agent 的 SEC 文件常见问题
Can AI agents access SEC filings?AI Agent 可以访问 SEC 文件吗？
Yes. Agents can use public EDGAR APIs and archives, subject to SEC fair-access guidance, declared User-Agent identification, rate limits and appropriate caching.
可以。Agent 可使用公开 EDGAR API 与档案，但需要遵守 SEC 公平访问规范、声明 User-Agent、限速并合理缓存。
Is the EDGAR API free and does it require an API key?EDGAR API 免费吗，需要 API key 吗？
The public data.sec.gov APIs do not require authentication or an API key. That does not remove rate, identification and usage obligations.
公开 data.sec.gov API 不需要认证或 API key，但仍须遵守速率、身份声明与使用规则。
Which SEC filings should an AI agent use?AI Agent 应使用哪些 SEC 文件？
Choose by question: 10-K for annual context, 10-Q for interim change, 8-K for material events, 13F for limited institutional holdings, plus amendments and exhibits when applicable.
按问题选择：10-K 看年度全貌，10-Q 看中期变化，8-K 看重大事件，13F 看有限范围的机构持仓，并纳入适用的修订与附件。
How should an agent cite a filing?Agent 应如何引用 SEC 文件？
Include issuer, form, filing date, accession number, document URL and exact section or anchor. Preserve the supporting excerpt and distinguish inference from filed fact.
引用应包含发行人、表单、申报日期、accession number、文档 URL 与准确章节或锚点，并保存支持性摘录、区分推断与申报事实。
How do I reduce hallucinations and prompt injection?如何减少幻觉与提示注入风险？
Retrieve bounded evidence, require citations, validate numbers and metadata, treat document instructions as data, restrict tool authority and send consequential outputs to review.
检索有限证据、强制引用、核验数字与元数据，把文档指令当作数据，限制工具权限，并将高影响输出交给人工审核。
Can QVeris guarantee complete SEC research?QVeris 能保证 SEC 研究完整吗？
No. QVeris can govern authorized tools and provenance, but completeness, interpretation, legal compliance and investment judgment remain the responsibility of the research process and reviewer.
不能。QVeris 可治理授权工具与来源，但完整性、解释、法律合规和投资判断仍由研究流程与审核人员负责。