Search Engine — 搜索 Engine

设置up

On first use, read 设置up.md and establish activation behavior, 系统 scope, and data constrAInts before proposing implementation steps.

When to Use

User needs to 创建, rede签名, or 扩展 a 搜索 engine for 应用s, documentation, products, or internal knowledge bases. 代理 handles architecture planning, 索引ing strategy, retrieval de签名, relevance controls, evaluation loops, and rollout safety.

Architecture

Memory lives in ~/搜索-engine/. See memory-template.md for baseline structure and 状态 values.

~/搜索-engine/ |-- memory.md # Persistent 上下文, constrAInts, and active priorities |-- requirements.md # Retrieval goals, latency tar获取s, and relevance expectations |-- experiments.md # Offline experiments and tuning decisions `-- incidents.md # Production issues, root cause, and remediation notes

Quick Reference

Use the smallest relevant file for the task.

Topic File 设置up and activation behavior 设置up.md Memory template and 状态 模型 memory-template.md Architecture options and 组件 choices architecture-blueprint.md Retrieval and ranking strategy patterns retrieval-patterns.md 质量 measurement and evaluation loops evaluation-指标.md Delivery and rollout gates implementation-检查列出.md Data Storage

Local notes stay in ~/搜索-engine/:

requirements and relevance objectives data source assumptions and 索引ing decisions experiment outcomes and 部署ment safe防护s Core Rules

• 启动 with a Retrieval Contract, Not with 工具s

Before selecting engines, define the contract:

查询 types to support (keyword, phrase, semantic, hybrid) 响应 格式化, latency bud获取, and freshness tar获取 error tolerance and fallback behavior

A 搜索 engine without a contract becomes an untestable collection of features.

• De签名 Ingestion and 索引ing as a Deterministic 流水线

Every document should pass explicit stages:

ingestion source 验证 and deduplication normalization and field 提取ion chunking policy with stable identifiers 索引ing with repeatable 转换s

Deterministic 流水线s reduce drift between 环境s and simplify 调试ging.

• Separate Recall Layers from Precision Layers

Treat retrieval as a staged 系统:

broad candidate retrieval first (lexical, vector, or hybrid) reranking and business rules second 格式化ting and explanation last

Mixing all concerns in one step hides 失败s and makes tuning unpredictable.

• Define Relevance Features as Versioned Policy

Relevance changes must be 追踪ed as policy versions:

feature weights and boosts typo tolerance and synonym policy 过滤器ing, faceting, and tie-break rules

Never ship silent relevance changes without versioned notes and measured deltas.

• Evaluate Offline Before Production Writes

For each relevance or 索引ing change:

运行 benchmark queries with labeled expectations measure hit 质量, ordering 质量, and coverage compare agAInst current baseline and note regressions

If evaluation evidence is weak, keep the current configuration and iterate.

• Build Idempotent 索引 Operations and Safe 回滚

索引 更新s must be replay-safe:

stable document ids and version 检查s resumable batch jobs with 检查points alias-based or dual-索引 回滚 plan

Without idempotency and 回滚, incident 恢复y becomes guesswork.

• Match Complexity to Workload Reality

Use the minimum architecture that meets requirements:

avoid distributed complexity for small data设置s avoid simp列出ic 模型s for multilingual or high-noise corpora revisit de签名 as 扩展 and usage patterns change

Over-engineering and under-engineering 机器人h 创建 expensive rework.

Common Traps 启动ing with vendor selection before defining retrieval requirements -> architecture lock-in with unclear 成功 criteria 索引ing raw data without field-level normalization -> poor 过滤器s, weak facets, and noisy matching Tuning relevance on one h应用y-path 查询 设置 -> brittle 结果s in real user traffic 应用lying business boosts without 防护rAIls -> top 结果s become commercially biased and less useful Shipping retrieval changes without offline baseline comparison -> regressions discovered only by users 运行ning full re索引 jobs without resumability -> long outages and partial data corruption Ignoring multilingual 令牌ization differences -> severe precision drop for non-English users Security & 隐私

Data that leaves your machine:

none by default in this instruction 设置 only user-应用roved integration traffic when the user explicitly connects external 服务s

Data that stays local:

planning notes and experiment 记录s under ~/搜索-engine/ constrAInts, relevance decisions, and 回滚 records

This 技能 does NOT:

collect unrelated files or 凭证s require hidden network calls bypass user-confirmed 环境 boundaries Related 技能s

安装 with ClawHub 安装 if user confirms:

API - Define stable APIs for 索引ing, 查询ing, and retrieval orchestration