Agentic Design

Patterns
🎯

Embedding-based Routing(EBR)

A semantic routing system that converts queries and route definitions into high-dimensional vector embeddings, using cosine similarity or other distance metrics to match incoming requests to the most semantically similar handler, enabling fuzzy matching, multi-lingual support, and context-aware routing beyond simple keyword matching

Complexity: mediumRouting

🎯 30-Second Overview

Pattern: Route requests by comparing vector embeddings in semantic space

Why: Captures meaning beyond keywords, enabling fuzzy matching and cross-lingual routing

Key Insight: Same embedding model + cosine similarity + smart thresholds = semantic routing

⚡ Quick Implementation

1Embed:Convert query to vector representation
2Compare:Calculate similarity to route embeddings
3Select:Choose route with highest similarity
4Threshold:Verify similarity meets minimum
5Route:Direct to matched handler/agent
Example: query_embedding = encode(query) → similarities = cosine_sim(query_embedding, route_embeddings) → if max(similarities) > 0.85: route()

📋 Do's & Don'ts

✅Use same embedding model for queries and routes
✅Normalize embeddings for consistent cosine similarity
✅Set appropriate similarity thresholds through testing
✅Cache embeddings for frequently used routes
✅Monitor embedding drift and update periodically
❌Mix different embedding models without retraining
❌Use cosine similarity blindly without validation
❌Ignore magnitude when it carries meaning
❌Skip normalization if model expects it
❌Use fixed thresholds across all route types

🚦 When to Use

Use When

  • • Semantic understanding is crucial
  • • Routes have fuzzy boundaries
  • • Need language-agnostic routing
  • • High-dimensional intent spaces

Avoid When

  • • Exact keyword matching suffices
  • • Embedding computation is costly
  • • Routes require strict boundaries
  • • Low-latency requirements (<10ms)

📊 Key Metrics

Similarity Score
Distribution of route matches
Embedding Latency
Time to generate vectors
Route Precision
% correctly routed queries
Route Recall
% of queries finding a route
Cache Hit Rate
% embeddings served from cache
Threshold Effectiveness
False positive/negative rates

💡 Top Use Cases

Semantic Search: query → embed → find most similar documents/FAQs
Intent Classification: user input → embed → match to intent clusters
Multi-lingual Support: translate meaning, not words → unified routing
Knowledge Base Navigation: question → embed → relevant article section
Dynamic Tool Selection: task description → embed → appropriate tool match

References & Further Reading

Deepen your understanding with these curated resources

Academic Papers

Is Cosine-Similarity of Embeddings Really About Similarity? (2024)

Semantic Routing for Enhanced Performance of LLM-Assisted Networks (2024)

Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks

Efficient Natural Language Response Suggestion for Smart Reply

Implementation Guides

Semantic Router Documentation - Aurelio Labs

Vector Similarity Explained - Pinecone

Semantic-Router Integration - Qdrant

LLM Semantic Router - Red Hat Developer

Tools & Libraries

Semantic Router - Superfast AI Decision Making

Sentence Transformers - State-of-the-art Embeddings

FAISS - Efficient Similarity Search

ChromaDB - Embedding Database

Community & Discussions

Netflix Study on Cosine Similarity Limitations

Why You MUST Know Semantic Router in 2024

Vector Database Similarity Metrics - Zilliz

Hugging Face Sentence Transformers Forum

Contribute to this collection

Know a great resource? Submit a pull request to add it.

Contribute

Patterns

closed

Loading...

Built by Kortexya