Response Caching for AI Agents: Semantic Cache, Exact Cache, and TTL Strategies

Why Cache LLM Responses

LLM API calls are expensive and slow. A single GPT-4o call costs $2.50-$10 per million input tokens and takes 1-5 seconds. If 30% of your users ask variations of the same question, you are paying for the same computation repeatedly.

Caching stores previous LLM responses and serves them for identical or similar future queries. A well-designed cache can reduce LLM API costs by 20-50% and cut response times from seconds to milliseconds for cache hits.

Exact-Match Cache

The simplest cache: hash the input and store the output. If the exact same input appears again, return the cached output.

flowchart LR
    subgraph IN["Inputs"]
        I1["Monthly call volume"]
        I2["Average deal value"]
        I3["Current answer rate"]
        I4["Receptionist cost<br/>per month"]
    end
    subgraph CALC["CallSphere Captures"]
        C1["Missed calls converted<br/>at 24 by 7 coverage"]
        C2["Receptionist payroll<br/>displaced or freed"]
    end
    subgraph OUT["Outputs"]
        O1["Recovered revenue<br/>per month"]
        O2["Operating cost saved"]
        O3((Net ROI<br/>monthly))
    end
    I1 --> C1
    I2 --> C1
    I3 --> C1
    I4 --> C2
    C1 --> O1 --> O3
    C2 --> O2 --> O3
    style C1 fill:#4f46e5,stroke:#4338ca,color:#fff
    style C2 fill:#4f46e5,stroke:#4338ca,color:#fff
    style O3 fill:#059669,stroke:#047857,color:#fff

import hashlib
import json
import time
from typing import Any

class ExactCache:
    def __init__(self, redis_client, default_ttl: int = 3600):
        self.redis = redis_client
        self.default_ttl = default_ttl

    def _make_key(self, model: str, messages: list[dict], **kwargs) -> str:
        """Create a deterministic cache key from the request parameters."""
        payload = json.dumps(
            {"model": model, "messages": messages, **kwargs},
            sort_keys=True,
        )
        return f"llm:exact:{hashlib.sha256(payload.encode()).hexdigest()}"

    async def get(self, model: str, messages: list[dict], **kwargs) -> dict | None:
        key = self._make_key(model, messages, **kwargs)
        cached = await self.redis.get(key)
        if cached:
            return json.loads(cached)
        return None

    async def set(
        self, model: str, messages: list[dict], response: dict, ttl: int = None, **kwargs
    ):
        key = self._make_key(model, messages, **kwargs)
        await self.redis.set(
            key,
            json.dumps(response),
            ex=ttl or self.default_ttl,
        )

# Usage with an LLM client
class CachedLLMClient:
    def __init__(self, openai_client, cache: ExactCache):
        self.client = openai_client
        self.cache = cache

    async def complete(self, model: str, messages: list[dict], **kwargs) -> str:
        # Check cache first
        cached = await self.cache.get(model, messages, **kwargs)
        if cached:
            return cached["content"]

        # Cache miss — call the LLM
        response = await self.client.chat.completions.create(
            model=model, messages=messages, **kwargs
        )
        content = response.choices[0].message.content

        # Store in cache
        await self.cache.set(
            model, messages, {"content": content}, **kwargs
        )
        return content

Exact caching works well for deterministic queries like classification, extraction, and structured data processing where the same input always produces the same desired output.

Semantic Cache: Matching Similar Queries

Users rarely ask the exact same question. They ask "What is your return policy?" and "How do I return an item?" and "Can I send something back?" — all meaning the same thing. A semantic cache uses embedding similarity to match these variations.

import numpy as np
import json
import hashlib

class SemanticCache:
    def __init__(self, embedder, redis_client, similarity_threshold: float = 0.92):
        self.embedder = embedder
        self.redis = redis_client
        self.threshold = similarity_threshold
        self._embeddings: list[tuple[str, np.ndarray]] = []

    async def _load_index(self):
        """Load cached embeddings from Redis into memory."""
        keys = await self.redis.keys("llm:semantic:emb:*")
        self._embeddings = []
        for key in keys:
            data = json.loads(await self.redis.get(key))
            self._embeddings.append((
                data["cache_key"],
                np.array(data["embedding"]),
            ))

    async def get(self, query: str) -> dict | None:
        query_embedding = await self.embedder.embed(query)

        best_key = None
        best_score = 0.0

        for cache_key, stored_embedding in self._embeddings:
            score = np.dot(query_embedding, stored_embedding) / (
                np.linalg.norm(query_embedding) * np.linalg.norm(stored_embedding)
            )
            if score > best_score:
                best_score = score
                best_key = cache_key

        if best_score >= self.threshold and best_key:
            cached = await self.redis.get(f"llm:semantic:resp:{best_key}")
            if cached:
                return json.loads(cached)

        return None

    async def set(self, query: str, response: dict, ttl: int = 3600):
        embedding = await self.embedder.embed(query)
        cache_key = hashlib.sha256(query.encode()).hexdigest()[:16]

        # Store the embedding for future similarity lookups
        await self.redis.set(
            f"llm:semantic:emb:{cache_key}",
            json.dumps({"cache_key": cache_key, "embedding": embedding.tolist()}),
            ex=ttl,
        )
        # Store the response
        await self.redis.set(
            f"llm:semantic:resp:{cache_key}",
            json.dumps(response),
            ex=ttl,
        )
        self._embeddings.append((cache_key, embedding))

The similarity threshold is critical. Set it too low (0.80) and you serve wrong answers. Set it too high (0.98) and you rarely get cache hits. Start at 0.92 and tune based on your domain.

TTL Strategies: When to Invalidate

Different types of cached data need different expiration strategies.

from enum import Enum

class CacheTTL(Enum):
    # Static knowledge: rarely changes
    FACTUAL = 86400        # 24 hours
    # Company-specific: changes occasionally
    POLICY = 3600          # 1 hour
    # User-specific: changes frequently
    PERSONALIZED = 300     # 5 minutes
    # Real-time data: changes constantly
    LIVE_DATA = 30         # 30 seconds

class SmartCache:
    def __init__(self, exact_cache: ExactCache, semantic_cache: SemanticCache):
        self.exact = exact_cache
        self.semantic = semantic_cache

    def classify_ttl(self, messages: list[dict]) -> int:
        """Determine appropriate TTL based on query characteristics."""
        last_message = messages[-1]["content"].lower()

        if any(w in last_message for w in ["price", "stock", "available", "weather"]):
            return CacheTTL.LIVE_DATA.value
        elif any(w in last_message for w in ["my account", "my order", "my"]):
            return CacheTTL.PERSONALIZED.value
        elif any(w in last_message for w in ["policy", "return", "shipping"]):
            return CacheTTL.POLICY.value
        else:
            return CacheTTL.FACTUAL.value

    async def get(self, messages: list[dict]) -> dict | None:
        # Try exact cache first (fastest)
        result = await self.exact.get("gpt-4o", messages)
        if result:
            return result

        # Fall back to semantic cache
        query = messages[-1]["content"]
        return await self.semantic.get(query)

Hit Rate Optimization

Track and optimize your cache hit rate with structured metrics.

from dataclasses import dataclass, field

@dataclass
class CacheMetrics:
    exact_hits: int = 0
    semantic_hits: int = 0
    misses: int = 0

    @property
    def total_requests(self) -> int:
        return self.exact_hits + self.semantic_hits + self.misses

    @property
    def hit_rate(self) -> float:
        if self.total_requests == 0:
            return 0.0
        return (self.exact_hits + self.semantic_hits) / self.total_requests

    @property
    def cost_savings_pct(self) -> float:
        return self.hit_rate * 100

    def report(self) -> str:
        return (
            f"Hit rate: {self.hit_rate:.1%} "
            f"(exact: {self.exact_hits}, semantic: {self.semantic_hits}, "
            f"miss: {self.misses}) | "
            f"Est. cost savings: {self.cost_savings_pct:.0f}%"
        )

FAQ

What similarity threshold should I use for semantic caching?

Start with 0.92 for general-purpose agents. For high-stakes domains like medical or legal, use 0.96 or higher to minimize incorrect cache hits. For casual conversational agents, 0.88-0.90 can work well. Monitor your false-positive rate — cases where the cache serves a response that does not actually answer the user's question — and adjust accordingly.

Should I cache streaming responses?

Yes, but cache the complete response after streaming finishes, not the stream itself. On a cache hit, you can either return the full response instantly or simulate streaming by emitting the cached text in chunks with small delays to maintain a consistent UX.

How do I handle cache invalidation when my knowledge base changes?

Use versioned cache keys that include a content hash or version number. When your knowledge base updates, increment the version. Old cache entries expire naturally via TTL while new queries hit the updated knowledge base. For critical updates, implement active invalidation by scanning and deleting affected cache keys.

---

#Caching #SemanticSearch #Redis #CostOptimization #Python #AgenticAI #LearnAI #AIEngineering