search-patterns

name	search-patterns
description	Search: Full-text search, Elasticsearch patterns, filters, and relevance tuning.
triggers	{"keywords":["search","Elasticsearch","Algolia","full-text","indexing","relevance","facets","fuzzy"]}
auto_load_when	Implementing search functionality
agent	architect
tools	["Read","Write","Bash"]

Search Patterns

Focus: Search implementation, query optimization, relevance

---

1. When to Use Search Engine

Use dedicated search engine when:

├── Full-text search needed
│   └── Fuzzy matching, synonyms
│   └── Partial matches, prefix search
│   └── Relevance ranking required
│
├── Complex filtering
│   └── Multiple faceted filters
│   └── Range queries, geo queries
│   └── Complex boolean logic
│
├── High-volume queries
│   && Many concurrent searches
│   && Sub-second response required
│   && Don't want to overload DB
│
└── Analytics on search
    └── Search suggestions
    && Popular queries
    └── Zero-result analysis

Stick with database when:
├── Exact matches only
├── Simple filters
├── Results fit in one query
└── Already have database infrastructure

---

2. Index Design

Index design patterns:

├── One index per entity type
    && Products, users, articles
    && Separate mappings
    && Different analyzers
│
├── Time-based indices
    && Logs, events, time-series
    && Delete old data by dropping index
    && Performance: smaller indices
│
└── Parent-child (rarely)
    && When updates to parent rare
    && When child queries independent
    └── Avoid: high join cost

---

3. Query Patterns

When to use query type:

├── Match query
    └── Full-text search
    && Fuzziness, slop
    && Use for: search box
│
├── Term query
    └── Exact value
    && No analysis
    && Use for: filters, exact match
│
├── Bool query
    && Multiple conditions
    && Must, should, must_not, filter
    && Use for: complex searches
│
├── Range query
    └── Numeric/date ranges
    && gte, lte, gt, lt
    && Use for: filters, faceted search
│
└── Prefix/wildcard
    && Prefix matching
    && Performance: slower than term
    && Use for: autocomplete

---

4. Filters vs Queries

When to use filters:

├── Filters (cached)
    && Boolean conditions
    && No scoring needed
    && Fast, cached automatically
    && Examples: category, status, date range
│
└── Queries (not cached)
    && Full-text relevance
    && Scoring matters
    && Examples: search text, title match

---

5. Performance Optimization

How to optimize search:

├── Limit returned fields
    && _source: include/exclude
    && Reduces network transfer
    && Use when: don't need all fields
│
├── Pagination
    && Use: from/size (shallow)
    && Deep pagination: search_after
    && Avoid: offset > 10k
│
├── Aggregations
    && Faceted search
    && Pre-computed at query time
    && Group by filters
│
└── Optimize text fields
    && norms: false (if no relevance needed)
    && index: false (if not searchable)
    && use keyword for sorting

---

Key Patterns

1. Index design first — Think about query patterns upfront
2. Filters for boolean, queries for relevance — Separate concerns
3. Denormalize — Store related data together, avoid joins
4. Reindex strategy — Plan for reindexing (new index, alias swap)
5. Relevance tuning — Use function_score, boost specific fields

---

Anti-Patterns

❌ LIKE '%query%' on large tables (full scan)
✅ Full-text search index: PostgreSQL tsvector or Elasticsearch

❌ Re-indexing entire dataset on every document update
✅ Incremental indexing — queue updates to search index

❌ Returning raw search scores without relevance tuning
✅ Boost by recency, popularity, and field weight

❌ No search analytics — unknown what users can't find
✅ Log zero-result queries and click-through rate

❌ Typos breaking search entirely
✅ Fuzzy matching with configurable edit distance

---

Quick Reference

Scale	Solution	Note
Small (<100k docs)	PostgreSQL FTS	Built-in, no extra infra
Medium	Typesense / Meilisearch	Self-hosted, fast
Large	Elasticsearch / OpenSearch	Complex, powerful
Hosted	Algolia	SaaS, fast setup
Vector search	pgvector / Pinecone	Semantic / AI search

🌍 Universal Language Support

• Turkish Native: This skill natively supports Turkish. If the user prompt is in Turkish, all analysis, formatting, and output MUST be entirely in Turkish. You do not need explicit "write in Turkish" instructions.