Federated Search: Architecture, Ranking & Evaluation

What Is Federated Search?什么是联邦搜索？

This focused article is part of the federated queries and data virtualization guide; use the pillar guide to compare related concepts, methods, and implementation decisions across the full topic.

本文是联邦查询与数据虚拟化指南内容集群中的专题文章；如需比较完整主题下的相关概念、方法与实施决策，请返回基石指南。

Federated search is an information-retrieval architecture that accepts one query, selects relevant autonomous search collections, submits adapted queries to them, and merges their returned results into one ranked response. The sources retain their own indexes, search engines, ownership, permissions, availability, and ranking behavior.

联邦搜索是一种信息检索架构：它接收一次查询，选择相关的自治搜索集合，把适配后的查询提交给这些集合，并将返回结果合并为一个有序响应。各来源保留自己的索引、搜索引擎、所有权、权限、可用性与排序行为。

Microsoft Research also calls the field federated information retrieval or distributed information retrieval. Its overview identifies three core problems: representing collections well enough to understand them, selecting the collections likely to answer a query, and merging their results. NIST’s TREC Federated Web Search work similarly evaluates resource selection and results merging as separate tasks.

Microsoft Research也把该领域称为联邦信息检索或分布式信息检索，其综述指出三个核心问题：充分表示集合以理解其内容、选择可能回答查询的集合，以及合并各集合结果。NIST的TREC联邦Web搜索工作同样把来源选择和结果融合视为两个独立任务。

The phrase describes behavior, not a guaranteed feature set. A federated search engine may query unlike external services through adapters; a same-engine cross-cluster feature may query remote clusters with shared syntax; a hybrid enterprise-search system may combine a central index with live remote providers. Confirm which model, content types, operators, permissions, ranking controls, and failure semantics a candidate actually supports.

该术语描述行为，并不保证统一功能集。联邦搜索引擎可能通过适配器查询不同的外部服务；同引擎跨集群功能可能用共享语法查询远程集群；混合企业搜索系统则可能把中央索引与实时远程提供方结合。应确认候选系统实际支持的模型、内容类型、操作符、权限、排序控制与故障语义。

Federated Search Architecture联邦搜索架构

A production architecture can be understood as four cooperating planes. A product may combine them in one service, but each responsibility still needs an owner, interface, observable state, and failure policy.

生产架构可以理解为四个协作平面。产品可以把它们合并到一个服务中，但每项职责仍需明确所有者、接口、可观察状态与故障策略。

Experience plane体验平面

Search box, API, query parser, filters, pagination, source indicators, partial-result warnings, and accessible result presentation.

搜索框、API、查询解析器、筛选、分页、来源标识、部分结果警告与无障碍结果呈现。

Federation plane联邦平面

Source registry, collection representation, resource selection, query routing, deadlines, normalization, deduplication, ranking, and provenance.

来源注册表、集合表示、来源选择、查询路由、截止时间、归一化、去重、排序与来源证明。

Adapter plane适配平面

Connectors translate operators, fields, identity context, pagination, errors, snippets, scores, and result metadata for each source.

连接器为各来源转换操作符、字段、身份上下文、分页、错误、摘要、评分与结果元数据。

Source plane来源平面

Autonomous search services, collections, indexes, owners, ranking models, permissions, capacity, release cycles, and availability.

自治搜索服务、集合、索引、所有者、排序模型、权限、容量、发布周期与可用性。

The source plane remains authoritative. The federation plane should not silently imply that every collection was searched, every operator was preserved, or every result was observed at the same time. The response contract must expose source coverage, timing, transformations, and degraded states.

来源平面仍然是权威。联邦平面不应静默暗示所有集合都已搜索、所有操作符都被保留，或所有结果都在同一时刻被观察。响应契约必须暴露来源覆盖、时间、转换和降级状态。

Core Components of a Federated Search Engine联邦搜索引擎的核心组件

Every component needs an explicit contract and evidence每个组件都需要明确契约与证据
Component组件	Responsibility职责	Evidence to require应要求的证据
Query broker查询代理	Accepts requests, sets context and deadlines, coordinates fan-out, and assembles responses接收请求、设置上下文与截止时间、协调扇出并组合响应	End-to-end trace with per-stage timing含各阶段时间的端到端追踪
Source registry来源注册表	Records collections, owners, topics, fields, protocols, limits, health, and versions记录集合、所有者、主题、字段、协议、限制、健康与版本	Versioned capability and ownership records带版本的能力与所有权记录
Resource selector来源选择器	Chooses sources likely to return useful results for each query and user为每个查询和用户选择可能返回有用结果的来源	Selection rationale, coverage metrics, and safe fallback选择理由、覆盖指标与安全回退
Query translator查询转换器	Maps fields, syntax, filters, operators, locale, and pagination into source-native requests把字段、语法、筛选、操作符、语言环境与分页映射为来源原生请求	Transformation trace and unsupported-feature policy转换追踪与不支持功能策略
Connector or protocol adapter连接器或协议适配器	Carries identity, submits searches, handles throttling, cancellation, retries, and response parsing传递身份、提交搜索，并处理限流、取消、重试与响应解析	Source/version test matrix and error mapping来源／版本测试矩阵与错误映射
Normalizer and deduplicator归一化与去重器	Creates a common result envelope and detects identical or near-identical items创建公共结果封装，并检测完全相同或近似相同项目	Canonicalization rules, confidence, and false-merge tests规范化规则、置信度与误合并测试
Result merger and reranker结果融合与重排器	Calibrates unlike scores or ranks and produces one ordered result list校准不同评分或名次，并生成一个有序结果列表	Versioned ranking policy and relevance evaluation带版本的排序策略与相关性评估
Security and operations plane安全与运维平面	Identity, authorization, secrets, audit, health, budgets, alerts, change, and recovery身份、授权、密钥、审计、健康、预算、告警、变更与恢复	Allow/deny tests, dashboards, runbooks, and drills允许／拒绝测试、仪表盘、运行手册与演练

Build a Source Registry and Capability Matrix First先建立来源注册表与能力矩阵

A source registry is operating state, not a list of URLs. For each collection, record the business owner, technical owner, content scope, language, audience, authority, index or API version, authentication method, supported fields and operators, ranking behavior, pagination, rate limits, expected latency, maintenance windows, data classification, result-retention rules, and an escalation path.

来源注册表是运行状态，而不是URL列表。对每个集合，应记录业务所有者、技术所有者、内容范围、语言、受众、权威性、索引或API版本、认证方式、受支持字段与操作符、排序行为、分页、速率限制、预期延迟、维护窗口、数据分类、结果保留规则与升级路径。

The capability matrix must be specific. “Supports search” does not tell the broker whether phrase queries, negation, field filters, date ranges, facets, highlighting, semantic retrieval, locale-aware stemming, stable pagination, total counts, or cancellation will behave as expected. Mark each capability as native, safely translated, approximated with a visible warning, or rejected.

能力矩阵必须具体。“支持搜索”并不能说明短语查询、否定、字段筛选、日期范围、分面、高亮、语义检索、语言环境词干处理、稳定分页、总数统计或取消是否按预期工作。应把每项能力标记为原生支持、安全转换、带可见警告的近似处理，或拒绝。

Collection representation supports resource selection. It may include curated topic labels, field and vocabulary statistics, sampled results, freshness, language, authority, or learned embeddings. Treat those descriptions as versioned derived data: stale representations can route queries away from the best source even while the source itself remains healthy.

集合表示为来源选择提供依据，可以包含人工维护的主题标签、字段与词汇统计、采样结果、新鲜度、语言、权威性或学习得到的向量。应把这些描述视为带版本的派生数据：陈旧表示会把查询错误地路由离开最佳来源，即使来源本身仍然健康。

How Federated Search Works, Step by Step联邦搜索如何工作：完整步骤

Establish the request contract.建立请求契约。
Authenticate the caller; capture locale, purpose, filters, requested content types, deadline, pagination token, and whether partial results are acceptable.
验证调用者身份，并记录语言环境、用途、筛选、请求内容类型、截止时间、分页令牌以及是否接受部分结果。
Parse and normalize the query.解析并规范化查询。
Separate terms, phrases, entities, fields, filters, dates, and intent without discarding the user’s explicit constraints.
分离词项、短语、实体、字段、筛选、日期与意图，同时不丢弃用户明确约束。
Filter sources by policy and capability.按策略与能力筛选来源。
Remove collections the caller may not discover or that cannot safely satisfy mandatory operators, data classes, or deadlines.
移除调用者无权发现，或无法安全满足必需操作符、数据分类或截止时间的集合。
Select and prioritize resources.选择并确定来源优先级。
Use collection representations and business rules to choose likely sources; retain a deterministic fallback for uncertain selection.
利用集合表示与业务规则选择可能相关的来源，并为不确定选择保留确定性回退。
Translate and dispatch in parallel.转换并并行分发。
Create source-native requests, pass the approved identity context, apply per-source budgets, and propagate cancellation.
生成来源原生请求，传递获批身份上下文，应用各来源预算并传播取消。
Normalize responses.归一化响应。
Map results into a common envelope containing source, stable identity, title, snippet, content type, observation time, rank or score, and access-safe metadata.
把结果映射为公共封装，包含来源、稳定身份、标题、摘要、内容类型、观察时间、名次或评分及可安全访问的元数据。
Deduplicate and merge.去重并融合。
Resolve exact and near duplicates, calibrate heterogeneous signals, apply a versioned ranking policy, and preserve provenance.
解析完全重复与近似重复，校准异构信号，应用带版本的排序策略并保留来源证明。
Return an honest response.返回诚实响应。
Present results with searched, skipped, failed, and timed-out sources, query transformations, freshness, ranking version, and a safe next action.
呈现结果时说明已搜索、跳过、失败与超时的来源、查询转换、新鲜度、排序版本及安全后续操作。

Preserve Query Meaning Across Different Search Engines跨不同搜索引擎保留查询含义

Query translation is semantic, not just syntactic. The same token can be a full-text term in one source, a controlled subject heading in another, and a product attribute in a third. Date fields may mean publication, modification, event, or ingestion time. A filter that cannot be represented exactly must not be silently dropped.

查询转换是语义问题，而不只是语法问题。同一词项在一个来源中可能是全文词，在另一个来源中是受控主题词，在第三个来源中则是产品属性。日期字段可能表示发布、修改、事件或采集时间。无法精确表示的筛选不得被静默丢弃。

Reject拒绝

Use when an unsupported constraint is mandatory for correctness, security, or user intent.

当不受支持的约束对正确性、安全或用户意图不可缺少时使用。

Approximate visibly显式近似

Translate to the closest safe behavior and show the changed semantics and affected sources.

转换为最接近的安全行为，并展示语义变化及受影响来源。

Split execution拆分执行

Send a supported core query, then perform a bounded post-filter only when access rules and result limits permit.

先发送受支持的核心查询，仅在访问规则和结果限制允许时执行有边界后筛选。

Route elsewhere改路由

Choose a source or central index that can honor the required operator exactly.

选择能够精确遵守必需操作符的来源或中央索引。

Version translators with source APIs and test representative queries, including punctuation, empty terms, very long terms, non-Latin text, locale changes, field aliases, quoted phrases, negation, filters, and pagination. Store the translated request in the trace without exposing secrets.

查询转换器应与来源API共同版本化，并测试代表性查询，包括标点、空词、超长词、非拉丁文本、语言环境变化、字段别名、引号短语、否定、筛选与分页。追踪中应保存转换后的请求，但不得暴露密钥。

Resource Selection: Search the Right Collections来源选择：检索正确集合

Querying every source can increase latency, cost, throttling, and noise. Resource selection ranks collections by their probability of returning useful, authorized results. Inputs may include declared scope, language, content type, coverage, sample queries, collection statistics, freshness, user context, historical judgments, and current health.

查询所有来源会增加延迟、成本、限流与噪声。来源选择按照返回有用且获授权结果的概率对集合排序。输入可包括声明范围、语言、内容类型、覆盖、采样查询、集合统计、新鲜度、用户上下文、历史判断与当前健康。

Selection quality needs two complementary checks. Coverage asks whether relevant sources were included; efficiency asks whether unnecessary sources were avoided. A selector that queries only one fast source may look efficient while hiding relevant material. A selector that always queries everything avoids selection errors but transfers the problem to latency and source protection.

选择质量需要两类互补检查。覆盖关注是否包含相关来源；效率关注是否避免不必要来源。只查询一个快速来源的选择器可能看似高效，却隐藏相关材料；始终查询全部来源虽避免选择错误，却把问题转移为延迟和来源保护。

Log why each source was selected, excluded by policy, skipped by health, or omitted by the selector.记录每个来源为何被选择、因策略排除、因健康跳过或被选择器省略。
Keep a small audited fallback set for novel or low-confidence queries.为新颖或低置信度查询保留一个经过审计的小型回退来源集。
Re-evaluate collection descriptions after content, taxonomy, API, ownership, or ranking changes.内容、分类、API、所有权或排序变化后重新评估集合描述。

Normalize, Deduplicate, and Rank Federated Results归一化、去重并排序联邦结果

Raw source scores are usually not directly comparable. One engine may return a bounded relevance score, another an unbounded score, another only rank positions, and another a business-prioritized order. Copying those values into one sort column creates a precise-looking but unjustified ranking.

原始来源评分通常不能直接比较。一个引擎可能返回有界相关性分数，另一个返回无界分数，第三个只提供名次，第四个则按业务优先级排序。把这些值直接放入一个排序列，会产生看似精确却缺乏依据的排名。

Result-merging decisions to make explicitly必须明确作出的结果融合决策
Decision决策	Practical options实际选项	Risk to test应测试风险
Common envelope公共封装	Stable ID, canonical URL, source, type, title, snippet, times, rank, score, and policy-safe fields稳定ID、规范URL、来源、类型、标题、摘要、时间、名次、评分与策略安全字段	Missing provenance or unsafe metadata leakage来源证明缺失或不安全元数据泄漏
Deduplication去重	Canonical IDs, normalized URLs, content fingerprints, or bounded similarity规范ID、规范化URL、内容指纹或有边界相似度	Merging distinct editions, locales, permissions, or time versions错误合并不同版本、语言、权限或时间版本
Score calibration评分校准	Per-source calibration, rank-based fusion, learned reranking, or typed result blocks按来源校准、基于名次的融合、学习重排或按类型结果块	Large or verbose sources dominating the list大型或冗长来源主导结果列表
Diversity and authority多样性与权威性	Source caps, type quotas, authority rules, freshness, and intent-aware blending来源上限、类型配额、权威规则、新鲜度与意图感知混排	Relevant authoritative material being displaced by duplicates or recency相关权威材料被重复项或新近内容挤出

Keep the original source rank and score alongside the merged rank. A result should explain where it came from, when it was observed, whether it was grouped with duplicates, and which ranking policy version placed it. Ranking changes are production changes: evaluate them against fixed judgments and representative traffic before promotion.

应在融合名次旁保留原始来源名次与评分。结果应说明来源、观察时间、是否与重复项分组，以及哪个排序策略版本决定其位置。排序变化属于生产变更，上线前必须用固定判断集和代表性流量评估。

Propagate Identity Without Expanding Access传递身份但不扩大访问

Define the identity chain from person or service, to federated session, to connector, to source principal. OpenSearch’s cross-cluster documentation provides one concrete example: the coordinating cluster authenticates the user and the remote cluster evaluates that user’s permissions. Other systems may use delegated tokens or fixed service accounts, so verify the exact chain rather than assuming equivalent behavior.

应定义从人员或服务、到联邦会话、连接器、再到来源主体的身份链。OpenSearch跨集群文档给出一个具体示例：协调集群验证用户身份，远程集群评估该用户权限。其他系统可能使用委托令牌或固定服务账户，因此必须验证确切身份链，而不能假定行为等价。

Discovery control: do not reveal restricted source names, result counts, titles, facets, snippets, errors, or timing through the federation.发现控制：不得通过联邦暴露受限来源名称、结果数量、标题、分面、摘要、错误或时间。
Least privilege: connector identities receive only approved search operations and collections; secrets are rotated and never exposed to clients.最小权限：连接器身份仅获得获批搜索操作与集合；密钥应轮换且不得暴露给客户端。
Post-processing safety: deduplication, caching, highlighting, snippets, reranking, and analytics must preserve source restrictions.后处理安全：去重、缓存、高亮、摘要、重排与分析必须保留来源限制。
Revocation evidence: test how quickly sessions, caches, connection pools, and remote sources reject a revoked identity.撤销证据：测试会话、缓存、连接池和远程来源拒绝已撤销身份所需时间。

Correlate the user request, selected sources, translated queries, source authorization decisions, returned items, deduplication groups, and final rank in an audit trail. Redact query terms or snippets when they contain sensitive data; auditability does not justify creating a new data leak.

审计轨迹应关联用户请求、选定来源、转换查询、来源授权决定、返回项目、去重分组与最终名次。查询词或摘要含敏感数据时应脱敏；可审计性并不能成为制造新数据泄漏的理由。

Design Explicitly for Slow, Missing, and Partial Sources显式设计慢来源、缺失来源与部分结果

Fan-out makes user latency and availability depend on multiple services. Set an overall deadline and smaller per-source budgets; propagate cancellation; cap concurrency; apply rate limits; isolate connection pools; and distinguish safe retry from duplicate work. A slow source must not consume the entire request budget or trigger an unbounded retry storm.

扇出会使用户延迟与可用性依赖多个服务。应设置整体截止时间和更小的各来源预算，传播取消，限制并发，应用速率限制，隔离连接池，并区分安全重试与重复工作。慢来源不得耗尽整个请求预算或触发无界重试风暴。

Elastic’s cross-cluster search documentation illustrates why availability semantics matter: remote clusters can be optional or required, and responses can report skipped, failed, or partial cluster states. A general federated search system needs an equally explicit contract even if it uses different names.

Elastic跨集群搜索文档说明了可用性语义为何重要：远程集群可以设为可选或必需，响应可报告跳过、失败或部分集群状态。通用联邦搜索系统即使使用不同术语，也需要同样明确的契约。

Fail closed关闭式失败

Return no result when a mandatory authority, policy source, or completeness condition is unavailable.

必需权威来源、策略来源或完整性条件不可用时，不返回结果。

Return marked partial results返回已标记部分结果

Show which sources answered, failed, timed out, were excluded, or were never selected, plus how ranking changed.

展示哪些来源已响应、失败、超时、被排除或从未被选择，并说明排序如何变化。

Use a bounded cache使用有边界缓存

Serve only when identity, retention, freshness, invalidation, and stale-result labeling are defined.

仅在身份、保留、新鲜度、失效与陈旧结果标识均已定义时提供缓存。

Route to a fallback路由到回退

Use an approved central index or alternate source and disclose the changed coverage.

使用获批中央索引或替代来源，并披露覆盖变化。

Evaluate Federated Search Before Production生产前评估联邦搜索

Connectivity proves only that requests can travel. Production validation must separately test source selection, translation, authorization, result merging, user relevance, latency, partial states, and operations. NIST’s TREC FedWeb materials are useful because they separate resource-selection judgments from result-merging judgments instead of hiding both behind one aggregate score.

连通性只能证明请求可以传输。生产验收必须分别测试来源选择、转换、授权、结果融合、用户相关性、延迟、部分状态与运维。NIST的TREC FedWeb材料具有参考价值，因为它把来源选择判断与结果融合判断分开，而不是隐藏在一个汇总分数后。

Evidence-based acceptance matrix证据化验收矩阵
Area领域	Test测试	Block release when阻止发布条件
Resource selection来源选择	Use judged queries to measure relevant-source coverage and unnecessary fan-out用已判断查询衡量相关来源覆盖与不必要扇出	Authoritative or high-value sources are systematically omitted权威或高价值来源被系统性遗漏
Translation转换	Compare transformed and source-native results for operators, fields, locales, and filters针对操作符、字段、语言环境与筛选比较转换结果和来源原生结果	Mandatory semantics are lost without rejection or warning必需语义丢失且没有拒绝或警告
Ranking排序	Evaluate final lists with graded judgments, source authority, duplicate handling, and query segments用分级判断、来源权威、重复处理与查询分段评估最终列表	A source, type, locale, or duplicate pattern dominates without justification某来源、类型、语言或重复模式无合理依据地主导结果
Security安全	Test allow, deny, discovery, revocation, cache, snippet, facet, error, and audit paths测试允许、拒绝、发现、撤销、缓存、摘要、分面、错误与审计路径	Restricted existence or content leaks at any stage任何阶段泄漏受限存在性或内容
Resilience韧性	Inject slow, failed, throttled, stale, malformed, and cancelled sources注入慢、失败、限流、陈旧、格式错误与取消来源	The system hangs, retries without bound, or returns silent partial results系统挂起、无界重试或返回静默部分结果
Operations运维	Change a connector, schema, ranking version, credential, and source owner; then execute rollback变更连接器、Schema、排序版本、凭据与来源所有者，再执行回滚	Impact, owner, evidence, or recovery cannot be traced影响、所有者、证据或恢复无法追踪

Track metrics by query segment and source, not only overall. Useful measures include successful source coverage, selected-source precision, result-list relevance such as nDCG at a chosen cutoff, duplicate rate, zero-result rate, authorization-denial correctness, per-source and end-to-end latency percentiles, timeout rate, partial-response rate, and user reformulation or abandonment. Define cutoffs and goals from the actual workflow; do not import arbitrary benchmark numbers.

指标应按查询分段和来源跟踪，而不只看总体。实用指标包括成功来源覆盖、选定来源精确度、选定截断位置的nDCG等结果相关性、重复率、零结果率、授权拒绝正确性、各来源与端到端延迟分位数、超时率、部分响应率，以及用户改写或放弃。截断和目标应来自实际工作流，不得照搬任意基准数字。

Use InfiniSynapse for a Related Approved Multi-Source Analysis使用InfiniSynapse执行相关的获批多来源分析

The InfiniSynapse public product pageInfiniSynapse公开产品页面 presents direct connections to supported platforms including Snowflake, Supabase, PostgreSQL, MySQL, MongoDB, Redis, SQL Server, Oracle, and ClickHouse, together with multi-source and multi-modal joint analysis. That makes the application a related entry when the actual goal is approved analysis across supported data sources.

InfiniSynapse public product pageInfiniSynapse公开产品页面展示了对Snowflake、Supabase、PostgreSQL、MySQL、MongoDB、Redis、SQL Server、Oracle和ClickHouse等受支持平台的直接连接，以及多来源与多模态联合分析。当实际目标是在受支持数据来源之间开展获批分析时，该应用可作为相关入口。

This guide does not claim that InfiniSynapse is a general federated search engine, enterprise search index, source selector, query-translation SDK, search-protocol gateway, deduplication service, score-calibration layer, result-merging engine, or universal relevance platform. Those search responsibilities remain with the selected architecture and accountable owners.

本指南不声称InfiniSynapse是通用联邦搜索引擎、企业搜索索引、来源选择器、查询转换SDK、搜索协议网关、去重服务、评分校准层、结果融合引擎或通用相关性平台。这些搜索职责仍属于选定架构及其问责所有者。

Before opening the application, prepare confirmed supported connection details, data-owner approval, least-privilege credentials, network and certificate requirements, approved data scope, keys and definitions, observation-time expectations, query or analysis limits, and source-load budgets.

打开应用前，请准备已确认支持的连接信息、数据所有者批准、最小权限凭据、网络与证书要求、获批数据范围、键与定义、观察时间预期、查询或分析限制及来源负载预算。

Analyze approved sources within defined boundaries在明确边界内分析获批来源

When support, authorization, semantics, freshness, and workload limits are confirmed, open InfiniSynapse to perform the related multi-source analysis workflow.

确认支持、授权、语义、新鲜度与工作负载限制后，打开InfiniSynapse执行相关的多来源分析工作流。

Analyze approved data sources分析获批数据来源

Federated Search FAQ联邦搜索常见问题

What is federated search?

什么是联邦搜索？

Federated search is an information-retrieval architecture that accepts one query, selects relevant autonomous collections or search engines, translates and submits source-native requests, and normalizes, deduplicates, merges, and ranks returned results in one response. Sources retain their own indexes, permissions, ownership, and availability.

联邦搜索是一种信息检索架构：它接收一次查询，选择相关的自治集合或搜索引擎，转换并提交来源原生请求，再把返回结果归一化、去重、融合和排序为一个响应。各来源保留自己的索引、权限、所有权与可用性。

How does federated search work?

联邦搜索如何工作？

A broker authenticates the caller, parses the query, filters sources by policy and capability, selects likely collections, translates and dispatches requests in parallel, normalizes responses, removes or groups duplicates, merges unlike rankings, and returns results with provenance, freshness, and searched, skipped, failed, or timed-out source states.

查询代理验证调用者，解析查询，按策略与能力筛选来源，选择可能相关的集合，并行转换和分发请求，归一化响应，移除或分组重复项，融合不同排序，最后返回带来源证明、新鲜度，以及已搜索、跳过、失败或超时来源状态的结果。

How is federated search different from unified or enterprise search?

联邦搜索与统一搜索或企业搜索有什么区别？

Federated search names a query-time mechanism that searches autonomous sources. Unified search names the one-place user experience, while enterprise search names a product or organizational category. Either outcome may use a centralized index, live federation, same-engine cross-cluster search, or a hybrid, so verify the actual architecture.

联邦搜索命名的是查询时搜索自治来源的机制；统一搜索命名的是单一入口用户体验；企业搜索命名的是产品或组织类别。它们可以采用集中式索引、实时联邦、同引擎跨集群搜索或混合方式，因此必须验证实际架构。

Does federated search copy data into a central index?

联邦搜索会把数据复制到中央索引吗？

Not necessarily. Query-time federation can search source-owned indexes without copying full content into one central index. The broker may still retain collection descriptions, normalized metadata, snippets, identifiers, caches, telemetry, or audit logs. Define permissions, retention, freshness, invalidation, and deletion for every retained artifact.

不一定。查询时联邦可以搜索来源自有索引，而不把完整内容复制到一个中央索引。代理仍可能保留集合描述、归一化元数据、摘要、标识符、缓存、遥测或审计日志。必须为每种保留工件定义权限、保留、新鲜度、失效与删除规则。

How are federated search results ranked?

联邦搜索结果如何排序？

Source scores are rarely directly comparable. A merger may calibrate scores by source, combine rank positions, apply a learned reranker, group results by type, or use authority, freshness, diversity, and business rules. Keep original source rank, merged rank, provenance, and ranking version, then evaluate the final list with fixed relevance judgments.

来源评分通常不能直接比较。融合器可以按来源校准分数、组合名次、应用学习重排、按类型分组，或使用权威性、新鲜度、多样性与业务规则。应保留原始来源名次、融合名次、来源证明与排序版本，并用固定相关性判断评估最终列表。

What happens when a federated search source fails?

联邦搜索来源失败时会发生什么？

The response should follow a declared policy: fail closed when the source is mandatory, return clearly marked partial results, use an authorized bounded cache, or route to an approved fallback. It must show which sources were searched, skipped, failed, timed out, or never selected and must never imply complete coverage silently.

响应应遵循已声明策略：来源必需时关闭式失败，返回清晰标记的部分结果，使用获授权的有边界缓存，或路由到获批回退。系统必须说明哪些来源已搜索、跳过、失败、超时或从未被选择，绝不能静默暗示覆盖完整。

Official Sources and Verification Notes官方来源与验证说明

These sources support formal federated-search concepts, evaluation structure, protocol history, and concrete same-engine implementation examples. They do not establish one universal product feature set. Verify the exact source, API, connector, version, query operators, identity flow, ranking behavior, limits, availability, licensing, and commercial terms you plan to operate.

这些来源支持联邦搜索形式概念、评估结构、协议历史与同引擎具体实现示例，但不建立通用产品功能集。请针对计划运行的确切来源、API、连接器、版本、查询操作符、身份流程、排序行为、限制、可用性、许可与商业条款重新验证。