🔍 Search Module - ThemisDB v1.4

Kategorie: Core Search
Version: v1.4.0
Status: ✅ Produktionsreif
Datum: Januar 2026

📑 Inhaltsverzeichnis

📋 Übersicht
✨ Search-Methoden
🚀 Schnellstart
📖 Detaillierte Dokumentation
💡 Best Practices
📊 Performance-Vergleich
📚 Siehe auch

📋 Übersicht

ThemisDB bietet drei leistungsstarke Search-Methoden: Full-Text Search, Vector Search und Hybrid Search. Jede Methode hat spezifische Stärken und Use Cases.

Vollständige Search-Dokumentation (v1.4)

Full-Text Search Guide - Inverted Index, BM25 Ranking, Analyzers, Fuzzy Search, Phrase Search mit 3 realen Use Cases
Vector Search Guide - HNSW Index, Similarity Metrics, Embeddings, Semantic Search mit 3 Beispielen
Hybrid Search Guide - Fusion Methods, Adaptive Strategies, Multi-Signal Ranking mit 3 kompletten Implementierungen
Search Feature Matrix - Vollständiger Vergleich, Decision Tree, Benchmarks, Limitations

✨ Search-Methoden

1. Full-Text Search

Best für: Exakte Keyword-Matches, Boolean Queries, Phrase Search

Features:

✅ BM25 Ranking
✅ Multi-Language Analyzers (DE, EN, FR, ES, etc.)
✅ Fuzzy Matching (Tippfehlertoleranz)
✅ Phrase Search
✅ Boolean Operators (AND, OR, NOT)
✅ Schnell (10-50ms)

Quickstart:

// Create index
CREATE FULLTEXT INDEX idx_content ON articles(content) ANALYZER "text_en"

// Search
FOR doc IN articles
  SEARCH ANALYZER(doc.content IN TOKENS(@query, "text_en"), "text_en")
  LET score = BM25(doc)
  SORT score DESC
  LIMIT 20
  RETURN {doc, score}

2. Vector Search

Best für: Semantic Search, Ähnlichkeitssuche, Multi-Modal Search

Features:

✅ HNSW Index (State-of-the-art ANN)
✅ Cosine, Euclidean, Dot Product Metrics
✅ Bis zu 2048 Dimensionen
✅ Sub-100ms Queries (bei Millionen Vektoren)
✅ Automatische Synonym-Erkennung
✅ Cross-Lingual Support

Quickstart:

// Create index
CREATE VECTOR INDEX idx_embedding ON articles(embedding)
  DIMENSIONS 768
  METRIC cosine

// Search
FOR doc IN articles
  LET similarity = COSINE_SIMILARITY(doc.embedding, @queryEmbedding)
  FILTER similarity > 0.7
  SORT similarity DESC
  LIMIT 20
  RETURN {doc, similarity}

3. Hybrid Search

Best für: Best of Both Worlds - Keywords + Semantik

Features:

✅ Kombiniert Full-Text + Vector
✅ Bessere Recall und Precision
✅ Multi-Signal Ranking
✅ Adaptive Fusion Strategies
✅ Robust gegen verschiedene Query-Typen

Quickstart:

// Full-Text Branch
LET ftResults = (
  FOR doc IN articles
    SEARCH ANALYZER(doc.content IN TOKENS(@query, "text_en"), "text_en")
    LET ftScore = BM25(doc)
    RETURN {doc, ftScore}
)

// Vector Branch
LET vecResults = (
  FOR doc IN articles
    LET vecScore = COSINE_SIMILARITY(doc.embedding, @queryEmbedding)
    FILTER vecScore > 0.7
    RETURN {doc, vecScore}
)

// Combine
FOR result IN UNION(ftResults, vecResults)
  COLLECT doc = result.doc
  AGGREGATE
    ftScore = MAX(result.ftScore || 0),
    vecScore = MAX(result.vecScore || 0)
  LET hybridScore = (ftScore * 0.4) + (vecScore * 0.6)
  SORT hybridScore DESC
  LIMIT 20
  RETURN {doc, ftScore, vecScore, hybridScore}

🚀 Schnellstart

Full-Text Search Example

// 1. Create index
await db.query(`
  CREATE FULLTEXT INDEX idx_blog ON blog_posts(title, content)
    ANALYZER "text_en"
`);

// 2. Insert documents
await db.query(`
  FOR post IN @posts
    INSERT post INTO blog_posts
`, {posts: blogPosts});

// 3. Search
const results = await db.query(`
  FOR doc IN blog_posts
    SEARCH ANALYZER(
      BOOST(doc.title IN TOKENS(@query, "text_en"), 3.0) OR
      doc.content IN TOKENS(@query, "text_en"),
      "text_en"
    )
    LET score = BM25(doc)
    SORT score DESC
    LIMIT 10
    RETURN {doc, score}
`, {query: "machine learning"});

Vector Search Example

from sentence_transformers import SentenceTransformer

# 1. Generate embeddings
model = SentenceTransformer('all-MiniLM-L6-v2')
texts = ["Quantum computing explained", "Machine learning basics"]
embeddings = model.encode(texts)

# 2. Insert with embeddings
for i, text in enumerate(texts):
    await db.query("""
        INSERT {
            content: @content,
            embedding: @embedding
        } INTO articles
    """, {
        "content": text,
        "embedding": embeddings[i].tolist()
    })

# 3. Search
query_embedding = model.encode("quantum physics")
results = await db.query("""
    FOR doc IN articles
      LET similarity = COSINE_SIMILARITY(doc.embedding, @queryEmbedding)
      FILTER similarity > 0.7
      SORT similarity DESC
      LIMIT 5
      RETURN {doc, similarity}
""", {"queryEmbedding": query_embedding.tolist()})

📖 Detaillierte Dokumentation

Complete Guides

Guide	Beschreibung	Seiten
Full-Text Search	Inverted Index, BM25, Analyzers, Fuzzy, Phrase	~60
Vector Search	HNSW, Similarity Metrics, Embeddings	~70
Hybrid Search	Fusion Methods, Adaptive Strategies	~80
Feature Matrix	Comparison, Decision Tree, Benchmarks	~55

Code Examples

Jeder Guide enthält:

✅ 3+ komplette Use Case Implementierungen
✅ Production-ready Code in Python, JavaScript, AQL
✅ Performance-Benchmarks
✅ Best Practices und Common Pitfalls

💡 Best Practices

Wann welche Methode?

Full-Text Search:

✅ Exakte Keywords wichtig
✅ Boolean Logic erforderlich
✅ Low Latency kritisch (< 50ms)
✅ Code/Technical Docs

Vector Search:

✅ Semantik wichtiger als Keywords
✅ Multi-Language Support
✅ Recommendation Systems
✅ Image/Audio Search

Hybrid Search:

✅ Best of Both Worlds
✅ E-Commerce
✅ Enterprise Search
✅ High Quality Requirements

Performance Tips

Index nur benötigte Felder
Tune HNSW Parameters (M, efConstruction, efSearch)
Use Pre-Filtering (Metadata filter vor Vector Search)
Enable Caching für häufige Queries
Monitor Query Latency mit Prometheus/Grafana

📊 Performance-Vergleich

Test Setup: 1M Dokumente, 16 CPU cores, 64GB RAM

Method	Avg Latency	Throughput	Recall@20	NDCG@20
Full-Text	25ms	600 QPS	0.72	0.68
Vector	35ms	400 QPS	0.78	0.74
Hybrid	55ms	250 QPS	0.89	0.85

Trade-offs:

Full-Text: Schnellste, aber begrenzte Semantik
Vector: Beste Semantik, höhere Latenz
Hybrid: Beste Qualität, höchste Latenz

Index Size Overhead

Method	Index Size	Build Time (1M docs)
Full-Text	~15-30%	2-5 min
Vector	~50-100%	8-12 min
Hybrid	~65-130%	10-17 min

📚 Siehe auch

External Resources

🔧 Zusätzliche Files

Datei	Beschreibung
fulltext_api.md	Full-Text API Reference
hybrid_fusion_api.md	Hybrid Fusion API
performance_tuning.md	Performance Optimization
migration_guide.md	Migration from older versions
stemming.md	Stemming Configuration

Status: ✅ All documentation complete and production-ready
Last Updated: April 2026
Version: 1.4.0 class VectorIndexManager { enum class Metric { L2, COSINE, DOT };

// API
Status init(objectName, dim, metric, M, efConstruction, efSearch, savePath);
Status addEntity(const BaseEntity& e, vectorField = "embedding");
Status removeByPk(std::string_view pk);

// KNN-Suche mit optionalem Whitelist Pre-Filtering
std::pair<Status, std::vector<Result>> searchKnn(
    const std::vector<float>& query,
    size_t k,
    const std::vector<std::string>* whitelistPks = nullptr
);

// Attribut-Filter (Post-Filtering)
struct AttributeFilter { field, value, op };
std::pair<Status, std::vector<Result>> searchKnnWithFilter(query, k, filters);

// Persistenz
Status saveIndex(directory);
Status loadIndex(directory);

};


### GraphIndexManager

```cpp
class GraphIndexManager {
    // Traversal
    enum class Direction { OUTBOUND, INBOUND, ANY };
    
    // API
    Status addEdge(from, to, type, properties);
    Status removeEdge(from, to, type);
    
    // Algorithmen
    std::vector<Path> bfs(start, maxDepth, direction, edgeFilter);
    std::vector<Path> dijkstra(start, end, weightField);
    std::vector<Path> astar(start, end, heuristic);
    
    // Graph Analytics
    double pagerank(nodeId);
    std::vector<Community> louvain();
    std::vector<std::string> shortestPath(from, to);
};

PropertyGraph

class PropertyGraph {
    // Nodes & Edges
    Status addNode(id, labels, properties);
    Status addEdge(from, to, type, properties);
    
    // Pattern Matching
    std::vector<Match> match(pattern);  // Cypher-ähnlich
    
    // Properties
    Status setProperty(id, key, value);
    Value getProperty(id, key);
};

TemporalGraph

class TemporalGraph {
    // Zeit-basierte Abfragen
    std::vector<Edge> getEdgesAt(timestamp);
    std::vector<Edge> getEdgesInRange(from, to);
    
    // Aggregation
    double aggregateEdgeProperty(node, edgeType, field, from, to, AggType);
};

Performance-Metriken

Index-Typ	Lookup	Insert	Memory
Secondary (B-Tree)	O(log n)	O(log n)	~100 bytes/entry
Vector (HNSW)	O(log n)	O(log n)	~500 bytes/vector
Graph (Adjacency)	O(1) + O(degree)	O(1)	~50 bytes/edge
Fulltext (Inverted)	O(k) k=terms	O(n) n=terms	~200 bytes/doc

Navigation Menu

🔍 Search Module - ThemisDB v1.4

📑 Inhaltsverzeichnis

📋 Übersicht

Vollständige Search-Dokumentation (v1.4)

✨ Search-Methoden

1. Full-Text Search

2. Vector Search

3. Hybrid Search

🚀 Schnellstart

Full-Text Search Example

Vector Search Example

📖 Detaillierte Dokumentation

Complete Guides

Code Examples

💡 Best Practices

Wann welche Methode?

Performance Tips

📊 Performance-Vergleich

Index Size Overhead

📚 Siehe auch

Related Documentation

External Resources

🔧 Zusätzliche Files

PropertyGraph

TemporalGraph

Performance-Metriken

Verwandte Dokumentation

FilesExpand file tree

README.md

Latest commit

History

README.md

File metadata and controls

🔍 Search Module - ThemisDB v1.4

📑 Inhaltsverzeichnis

📋 Übersicht

Vollständige Search-Dokumentation (v1.4)

✨ Search-Methoden

1. Full-Text Search

2. Vector Search

3. Hybrid Search

🚀 Schnellstart

Full-Text Search Example

Vector Search Example

📖 Detaillierte Dokumentation

Complete Guides

Code Examples

💡 Best Practices

Wann welche Methode?

Performance Tips

📊 Performance-Vergleich

Index Size Overhead

📚 Siehe auch

Related Documentation

External Resources

🔧 Zusätzliche Files

PropertyGraph

TemporalGraph

Performance-Metriken

Verwandte Dokumentation