libvector Skill

When to Use

• Building semantic search functionality
• Implementing RAG retrieval pipelines
• Finding similar documents by embedding
• Filtering vector results by metadata

Key Concepts

VectorIndex: Storage-backed index for vectors with cosine similarity search and metadata filtering.

VectorProcessor: Processes documents into embeddings and indexes them.

calculateDotProduct: Utility function for computing dot product (cosine similarity for normalized vectors) with loop unrolling for performance.

Usage Patterns

Pattern 1: Search by vector

import { VectorIndex } from "@copilot-ld/libvector/index.js";

const index = new VectorIndex(storage, "content");
const results = await index.search(queryVector, {
  limit: 10,
  threshold: 0.7,
  filter: { type: "document" },
});

Pattern 2: Add vectors

await index.add({
  id: "doc-123",
  vector: embedding,
  metadata: { type: "document", title: "Example" },
});

Pattern 3: Calculate similarity directly

import { calculateDotProduct } from "@copilot-ld/libvector";

// For normalized vectors, dot product equals cosine similarity
const similarity = calculateDotProduct(vectorA, vectorB);

Integration

Used by Vector service. Embeddings generated via LLM service. Stored in data/vectors/.