semantic-cache（语义缓存）

Semantic Cache

Cache LLM responses by meaning using Redis vector search. Similar questions return cached answers instantly instead of making expensive API calls.

How It Works

1. User asks a question or makes an LLM request
2. The question is embedded into a vector using OpenAI text-embedding-3-small
3. Redis vector search finds semantically similar cached queries (cosine similarity > 0.80)
4. Cache hit: Return the cached response instantly (~100ms)
5. Cache miss: Pass through to the LLM, cache the response for future similar queries

Commands

Cache a query and response

node scripts/cache.js store "What is the capital of France?" "The capital of France is Paris."

Check cache for a similar query

node scripts/cache.js lookup "What's France's capital city?"

Cache stats

node scripts/cache.js stats

Clear all cached entries

node scripts/cache.js clear

Interactive mode — wraps any LLM call with caching

node scripts/cache.js query "Your question here"

This checks cache first. On miss, calls OpenAI, caches the result, and returns it.

When to Use This Skill

• Before making any LLM API call, check if a semantically similar query was already answered
• When building agents that answer repetitive questions (support bots, FAQ systems)
• When you want to reduce OpenAI/Anthropic API costs by 40-80%
• When you need faster response times for common queries

Configuration

Set these environment variables:

• REDIS_URL — Redis connection string with vector search support (Redis Cloud or Redis Stack)
• OPENAI_API_KEY — For generating embeddings
• SEMANTIC_CACHE_THRESHOLD — Similarity threshold 0-1 (default: 0.80, higher = stricter matching)
• SEMANTIC_CACHE_TTL — Cache TTL in seconds (default: 86400 = 24 hours)

Example Workflow

User: "How do I reset my password?"
  -> Embed query -> Search Redis -> MISS
  -> Call LLM -> Get response -> Cache it -> Return response

User: "I forgot my password, how do I change it?"
  -> Embed query -> Search Redis -> HIT (92.7% similar)
  -> Return cached response in 8ms (saved ~2 seconds + API cost)

Performance

• Cache lookup: ~5-15ms (vs 1-5 seconds for LLM call)
• Embedding generation: ~50-100ms
• Storage per entry: ~6KB (1536-dim vector + metadata)
• Supports millions of cached entries