How Semantic Caching Works

Picture a mobile app with hundreds of thousands of users. Every day it generates thousands of similar LLM queries: “How to add a contact?”, “How to create a new contact?”, “How to enter a contact in the list?”. Without semantic caching, each such request hits the API, multiplying costs. We solve this by deploying a mechanism that stores responses together with the vector representation of the query. On repeat requests, the system searches for semantically close embeddings and returns the stored answer, bypassing the LLM. In practice, this reduces API costs by 40–60% and cuts latency from seconds to milliseconds.

Problems Semantic Cache Solves

Standard exact-key caching is powerless against synonyms and rephrasings. Users phrase the same question differently — and you pay each time. LLM APIs are expensive at high frequency of repetitive requests; a typical hit rate without caching is near zero. Response latency of 2–5 seconds hurts UX in a mobile app, especially on slow channels. Semantic cache solves all three problems at once.

How We Do It: FastAPI Middleware

The server side is FastAPI middleware. On each request, we generate an embedding via OpenAI, search for the nearest one in a vector store. If cosine similarity exceeds a threshold, we return the cache; otherwise, we call the LLM and save the new embedding. Example code:

import numpy as np
from openai import AsyncOpenAI

client = AsyncOpenAI()
cache: list[dict] = []  # In production: Redis + pgvector or Pinecone

async def get_embedding(text: str) -> list[float]:
    response = await client.embeddings.create(
        model="text-embedding-3-small",
        input=text
    )
    return response.data[0].embedding

def cosine_similarity(a: list[float], b: list[float]) -> float:
    a_arr, b_arr = np.array(a), np.array(b)
    return float(np.dot(a_arr, b_arr) / (np.linalg.norm(a_arr) * np.linalg.norm(b_arr)))

async def semantic_cache_lookup(query: str, threshold: float = 0.92) -> str | None:
    query_emb = await get_embedding(query)
    for entry in cache:
        similarity = cosine_similarity(query_emb, entry["embedding"])
        if similarity >= threshold:
            return entry["response"]
    return None

Threshold is a critical parameter. At 0.85 the cache becomes too aggressive: semantically different questions get the same answer. At 0.97 it barely works. The optimal range for most domains is 0.90–0.95, tuned on real queries.

Step-by-Step Semantic Cache Setup

1. Log user queries in production (at least 1000).
2. Generate embeddings using the chosen model (we use text-embedding-3-small).
3. Build a vector index: HNSW for fast search.
4. Tune threshold on a holdout set: analyze hit rate and quality.
5. Deploy middleware on the server side.
6. Monitor hit rate, savings, and false positives.

Why Threshold 0.92 Is an Optimal Start

At threshold 0.92, false positives are minimal, and hit rate on typical questions reaches 40–60%. Lower values give more matches but degrade answer quality. Higher values sharply reduce cache effectiveness. We always fine‑tune the exact value on your logs for the best balance.

When Semantic Cache Does Not Work

For dynamic data — user balance, order status, exchange rates — caching is useless. We identify such requests with a classifier and exclude them from the cache. Also, caching is ineffective if questions are unique and never repeated.

Redis vs pgvector: Which to Choose

Redis with RediSearch is 3× faster than pgvector for caches up to 50k entries, but pgvector scales to millions without precision loss.

For cosine similarity we use the standard formula: cosine of the angle between vectors via dot product.

Invalidation and TTL

Semantic cache must be invalidated when the system prompt or base model is updated — old answers may not match new behavior. Recommended TTL: 7–30 days for stable FAQ-like questions. For time‑sensitive questions, we do not apply caching.

What’s Included in the Work

• Architecture diagram for integrating semantic cache into a mobile app (iOS/Android).
• Setup of embedding generation and model selection (OpenAI, Cohere, SentenceTransformers).
• Tuning similarity threshold on your query logs.
• Implementation of server‑side middleware (FastAPI, Node.js, Go).
• Monitoring of hit rate and cost savings.
• Documentation and team training.

Our experience includes 5+ deployments for apps with audience from 10k to 1M DAU. We guarantee a hit rate of at least 40% on stable questions.

Typical Mistakes When Implementing

• Choosing too low a threshold — the cache starts confusing semantically different queries.
• Ignoring invalidation when the prompt changes — users get outdated responses.
• Lack of fallback: if vector search fails, the request must go directly to the LLM.
• Incorrect choice of vector index (flat vs HNSW) for the cache size.

Timeline Estimates

Basic semantic cache on Redis + OpenAI Embeddings — 2–3 days. With threshold tuning on real data and hit‑rate monitoring — 3–5 days. If integration into existing mobile infrastructure is needed, contact us for an estimate. Also, order an audit of your current AI costs — we will calculate potential savings and propose an architecture for your load. Get a consultation from an engineer who has already deployed such solutions.

Source: Redis Stack documentation

Machine Learning in Mobile Apps: CoreML, TFLite, and On-Device Models

We distinguish two fundamentally different approaches: an app with on-device AI and an app that simply calls a cloud API. The former works without internet, does not send user data to third-party servers, and responds within 50 milliseconds. The latter depends on network latency and pricing plans. Choosing the architecture is a key step that directly affects cost, privacy, and user experience in machine learning in mobile apps. Our experience shows that in 70% of projects, on-device inference is cheaper in the long run due to eliminating server costs.

How to Choose Between CoreML and TFLite for On-Device Inference?

CoreML — Apple's native framework for running ML models on device. Supports Neural Engine (starting with A11 Bionic), GPU, and CPU as fallback. Models are converted to .mlmodel format via coremltools from PyTorch, ONNX, or TensorFlow. Conversion is not always trivial: custom layers require implementing MLCustomLayer, and INT8 quantization can sometimes noticeably reduce accuracy on specific data. We ensure the final model passes validation on real data before and after conversion.

TensorFlow Lite — cross-platform alternative for Android and Flutter. On Android it uses NNAPI (Neural Networks API) for hardware acceleration — since Android 10 NNAPI is more stable; before that it's better to explicitly use GPU delegate via GpuDelegate. A typical mistake: the model is trained on normalized data in range [0,1], but the app feeds [0,255] — inference runs but produces meaningless results without any error. We include an automatic input data validation module in the SDK.

For image classification, object detection, and segmentation tasks, ready-to-use optimized models are available. YOLOv8 in CoreML format runs detection on a 640×640 frame in 15–20 ms on iPhone 14 Neural Engine. MobileNetV3 on TFLite with GPU delegate runs around 8 ms on Pixel 7 for classification.

What If You Need Text Generation On-Device?

Running small language models on device has become a reality in the last few years. Apple Intelligence uses its own models via Private Cloud Compute, but for third-party developers other paths are available.

llama.cpp with Metal backend on iOS is a working approach for phi-3-mini (3.8B parameters, 4-bit quantization, ~2.3 GB). Inference: 15–25 tokens/second on iPhone 15 Pro. For integration in Swift, use the Swift Package llama.swift or a wrapper via C interface llama.h. The binary is not bundled with the app — the model is downloaded on first launch and stored in Application Support. Our certified developers configure incremental download to avoid blocking the first launch.

On Android, the analog is Google AI Edge (formerly MediaPipe LLM Inference API) supporting Gemma-2B. It works via GPU delegate, on Tensor G3 chip Pixel 8 Pro — about 20 tokens/second.

Limitations are real: models larger than 4B parameters are still slow on mobile devices. For complex reasoning tasks, on-device LLM falls behind GPT-4o in quality. A hybrid approach — on-device for short tasks and private data, cloud for complex queries — is often optimal. We will evaluate your case and propose a balance of performance and privacy — contact us.

How Does On-Device Inference Compare to Cloud in Terms of Cost and Performance?

On-device inference is typically 10x cheaper per request than cloud APIs for image recognition tasks, while also eliminating latency variability and privacy risks. The table below summarizes the trade-offs.

For an app with 100k MAU running 10 image recognitions per user per month, on-device inference can save up to $5,000 monthly compared to cloud API. Get a free consultation on your ML architecture today.

Integrating OpenAI API and Other Cloud Models

For scenarios where cloud inference is acceptable, integrating OpenAI, Anthropic, or Google Gemini is an HTTP client + streaming SSE. In Swift, AsyncThrowingStream is convenient for streaming responses. In Kotlin, use Flow.

Critically: API keys must never be stored in the app bundle. Even an obfuscated key can be extracted from the IPA in 10 minutes using strings or frida. Correct architecture: mobile app → your own backend → OpenAI API. The backend controls rate limiting, logs requests, and protects the key.

What Is Included in the Work (Deliverables)

• Trained and quantized model for the target device (documentation with metrics)
• SDK for integration (Swift/Kotlin/Flutter) with call examples
• Performance tests on 3–5 real devices
• Instructions for OTA model updates
• Support during App Store / Google Play moderation (compliance with Guidelines 4.2, 5.1)
• 2 weeks of technical support after release

Typical Project Pipeline

1. Task analysis — measure latency, privacy, size, supported devices.
2. Model prototyping — in Python, evaluate accuracy on target data.
3. Conversion and quantization — for CoreML/TFLite with validation.
4. Integration into the app — model wrapped in a service layer (easy to swap CoreML ↔ TFLite ↔ cloud).
5. Testing — on real devices, measure FPS, RAM, battery.
6. Deployment — via TestFlight / Firebase App Distribution, monitor metrics.

Timelines: integration of a ready CoreML/TFLite model — 1–2 weeks, development of a custom model with mobile optimization — from 6 weeks, on-device LLM chat with personalization — 4–8 weeks.

Why We Take on Complex Cases?

10+ years of experience in mobile development, 50+ implemented AI/ML solutions, guarantee of compatibility with current iOS and Android versions. All projects undergo code review and load testing. The cost includes preparation of moderation documentation and training of your team.

Contact us — we will help you choose the architecture and implement ML in your app turnkey. Order an audit of your existing solution — we will assess the potential for server cost savings free of charge. In some projects, savings can reach significant amounts per month.