vector-db-patterns

name	vector-db-patterns
description	Vector Database: Embedding storage, semantic search, similarity metrics, RAG architectures.
triggers	{"extensions":[".py",".ts"],"directories":["vector/","embeddings/","rag/"],"keywords":["vector","embedding","pinecone","chroma","weaviate","milvus","qdrant","pgvector","faiss","semantic search"]}
auto_load_when	Building RAG systems or semantic search features
agent	ai-engineer
tools	["Read","Write","Bash"]

Vector Database Patterns

Focus: Embedding storage, semantic search, similarity, RAG

1. Vector Database Selection

Which vector DB to use?
├── Scale: Small (<100K vectors) → Chroma, in-memory
├── Scale: Medium (100K-10M) → Pinecone, Weaviate, Qdrant
├── Scale: Large (10M+) → Milvus, Elasticsearch
│
├── Infrastructure:
│   ├── Cloud-native → Pinecone, Elasticsearch
│   ├── Self-hosted → Weaviate, Qdrant, Milvus
│   └── PostgreSQL extension → pgvector (if already using PG)
│
├── Features needed:
│   ├── Hybrid search (vector + keyword) → Weaviate, Elasticsearch
│   ├── Multi-tenancy → Pinecone, Milvus
│   └── Time-based filtering → Most support
│
└── Cost: Open source (Weaviate, Qdrant, pgvector) vs Managed (Pinecone)

---

2. Embedding Strategy

Embedding Pipeline:
├── Text preprocessing
│   ├── Chunking strategy (fixed size, sentence, recursive)
│   ├── Chunk size: 256-512 tokens for semantic search
│   └── Overlap: 10-20% to preserve context
│
├── Embedding model selection
│   ├── General purpose → text-embedding-3-small, bge-small
│   ├── Code → codellama, codex
│   └── Multilingual → bge-m3, multilingual-e5
│
├── Batch processing
│   ├── Batch size: 100-1000 for efficiency
│   ├── Async processing for large datasets
│   └── Progress tracking for long jobs
│
└── Storage
    ├── Store raw text + embeddings
    ├── Metadata for filtering
    └── Original source for citation

---

3. Search Implementation

Semantic Search Flow:
├── Query preprocessing
│   ├── Same embedding model as indexing
│   ├── Clean, truncate if needed
│   └── (Optional) Query expansion
│
├── Vector search
│   ├── Top-K retrieval (k=3-10 for RAG)
│   ├── Similarity metric: Cosine (default), Dot, Euclidean
│   └── (Optional) Filtering by metadata
│
├── Reranking (optional but recommended)
│   ├── Cross-encoder reranking for accuracy
│   └── Re-rank top 20 to top 5
│
└── Post-processing
    ├── Extract source documents
    └── Format for LLM context

---

4. RAG Architecture Patterns

RAG Variants:
├── Naive RAG
│   └── Retrieve → Pass directly to LLM
│   └── Simple but can retrieve irrelevant docs
│
├── Advanced RAG
│   ├── Query preprocessing (rewriting, expansion)
│   ├── Chunking optimization
│   ├── Hybrid search (vector + keyword)
│   └── Reranking
│
├── Modular RAG
│   ├── Routing (choose data source based on query)
│   ├── Fusion (combine multiple retrievers)
│   └── Memory (store conversation context)
│
└── Agentic RAG (LangChain/LlamaIndex agents)
    ├── Multi-step retrieval
    ├── Tool use (web search + DB)
    └── Iteration until answer found

---

Key Patterns

1. Chunk strategically - Fixed size with overlap, optimize for semantic units
2. Hybrid search - Vector + keyword (BM25) for better recall
3. Rerank - Cross-encoder for top results
4. Metadata filtering - Filter by date, source, category
5. Cache embeddings - Don't re-embed same text

---

Anti-Patterns

❌ Using wrong chunk size — too small loses context, too large adds noise
✅ Experiment with 256-1024 tokens, measure recall

❌ No metadata filtering — searching everything for every query
✅ Add filters (date, source, category) to reduce search space

❌ Same embedding model for all — code needs code model, text needs text model
✅ Choose model based on content type

❌ Storing only vectors — losing original text for citations
✅ Store original text + metadata alongside vector

❌ No fallback — search fails silently
✅ Fallback to keyword search or cached results

---

Quick Reference

Task	Solution	Note
Semantic search	Vector similarity (cosine)	Top-K retrieval
Hybrid search	Vector + BM25	Weaviate, Elasticsearch
Reranking	Cross-encoder	Re-rank top-20 to top-5
Filtering	Metadata + vector	Combine in query
Scaling	Sharding + replication	DB-specific
Cost optimization	Quantization	FP16 → INT8 → binary

🌍 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.