AgentSkillsCN

vector-db-elixir

在 Elixir 中设计并实现向量数据库。当您需要构建嵌入式存储、相似性搜索、k-NN 检索，或当用户提及 Elixir 中的向量数据库、嵌入式表示，或语义搜索时，此功能将为您提供强大助力。

SKILL.md
--- frontmatter
name: vector-db-elixir
description: Design and implement a vector database in Elixir. Use when building embedding storage, similarity search, k-NN retrieval, or when the user mentions vector DB, embeddings, or semantic search in Elixir.

Vector Database in Elixir

Scope

A vector DB stores high-dimensional vectors (embeddings) and supports similarity search: find the k nearest vectors to a query vector by distance (cosine, Euclidean, dot product).

Core Operations

OperationDescription
insertAdd vector(s) with optional metadata (id, payload)
searchQuery vector → return k nearest vectors (by similarity)
deleteRemove by id or filter
getFetch by id (optional)

Implementation Workflow

  1. •

    Choose storage backend

    • •In-memory: ETS (fast, no persistence)
    • •Persistent: file (binary/term), or PostgreSQL with pgvector
    • •External: Qdrant client
  2. •

    Choose distance metric

    • •Cosine similarity (often for normalized embeddings)
    • •Euclidean (L2)
    • •Dot product (when vectors already normalized, equivalent to cosine)
  3. •

    Implement or wire

    • •Exact k-NN: compute distance to every vector (fine for < ~100k vectors in memory)
    • •Approximate: use pgvector indexes (IVFFlat, HNSW) or external engine for scale
  4. •

    Expose API

    • •GenServer or dedicated context module: insert/2, search/2, delete/1
    • •Optional HTTP/Web API on top

Elixir Stack (Typical)

  • •Nx or Scholar for vector math (cosine, L2, dot product)
  • •GenServer + ETS or :array for in-memory index
  • •pgvector + Ecto for persistent, indexed storage
  • •Application supervision tree to start the vector store process

Design Checklist

  • • Vector dimension fixed per collection (e.g. 1536 for OpenAI embeddings)
  • • Ids unique; support upsert by id if required
  • • Metadata/payload stored with vector for filtering or display
  • • Distance metric consistent for insert and search
  • • Persistence strategy (none, file, DB) decided and implemented

Additional Resources