@anchor/tag-walker

Semantic search without embeddings. No GPU. No cloud.

Overview

Tag-Walker is a graph-based FTS (Full-Text Search) retrieval system that replaces traditional vector databases. It uses a deterministic tag-based retrieval approach instead of embedding-based search, making it completely CPU-based with zero GPU/VRAM requirements.

The system implements a 3-phase search algorithm:

1. Anchor: Direct Full-Text Search for initial results
2. Pivot: Tag extraction and expansion using synonym rings
3. Walk: Associative expansion through graph-based exploration

Installation

npm install @anchor/tag-walker

Quick Start

import { TagWalker } from '@anchor/tag-walker';
import { PGlite } from '@electric-sql/pglite';

// Initialize PGlite database
const db = await PGlite.create();

// Initialize TagWalker
const tagWalker = new TagWalker({ 
  db,
  maxResults: 50,
  tokenBudget: 4096 
});

// Ingest content with tags
await tagWalker.ingest(
  "Artificial intelligence is transforming industries worldwide.",
  ["AI", "technology", "innovation"],
  "source_document_1"
);

// Search for relevant content
const results = await tagWalker.search("machine learning technology");
console.log(results);

How It Works

Tag-Walker replaces embedding-based vector search with deterministic tag-based retrieval. Instead of relying on neural networks to create semantic vectors, it uses:

• Synonym Rings: Static lookup tables for semantic expansion
• Graph Navigation: Associative expansion through shared tags
• Full-Text Search: Fast indexing and retrieval using PostgreSQL FTS

This approach provides semantic search capabilities without the computational overhead of embedding models, making it ideal for local-first applications.

API Reference

new TagWalker(config)

Initialize the Tag-Walker with configuration options:

• db: PGlite database instance
• synonymsPath?: Path to custom synonym dictionary
• maxResults?: Maximum number of results to return (default: 50)
• tokenBudget?: Maximum token budget for results (default: 4096)

tagWalker.search(query, options?)

Perform a search with the following parameters:

• query: Search query string
• options?: Optional search parameters
  • buckets?: Filter by buckets
  • tags?: Filter by tags
  • provenance?: Filter by provenance ('internal', 'external', 'quarantine', 'all')

Returns: Promise<SearchResult[]>

tagWalker.ingest(content, tags, source)

Ingest content into the database:

• content: Content string to store
• tags: Array of tags associated with content
• source: Source identifier for the content

Returns: Promise<string> - ID of the ingested content

Comparison: Tag-Walker vs Vector DB

| Feature | Tag-Walker | Vector DB | |---------|------------|-----------| | Hardware Requirements | CPU only | GPU recommended | | Embedding Models | Not required | Required | | Storage Efficiency | High | Lower (vectors) | | Search Speed | Fast (FTS + Graph) | Variable | | Semantic Understanding | Rule-based | Learned | | Local Operation | Yes | Depends on model | | Cost | Low | Potentially high |

When to Use Tag-Walker

• ✅ Local-first applications without cloud dependencies
• ✅ CPU-only environments without GPU access
• ✅ Applications requiring deterministic search results
• ✅ Systems with limited computational resources
• ✅ Privacy-focused applications (no external APIs)
• ✅ Fast prototyping without ML model setup

License

Elastic License v2.0 (ELv2)