Visual Search Development for Mobile Apps

Turnkey Visual Search Development for Mobile Apps

A user photographs a product in a competitor's store or takes a screenshot from Instagram — and wants to find the same item in your catalog. Our experience shows that up to 30% of users abandon the app if the search cannot handle such scenarios. Visual search solves this need. Technically, the task consists of two parts: obtaining a vector representation (embedding) of the image and finding nearest neighbors in the database. We guarantee results: we implement a working prototype within 3–6 weeks. The average user time saved is 30 seconds per query, and increasing search conversion by 25% leads to a 15% revenue increase. Visual search reduces search time by 50% compared to text search. Server costs for embedding are approximately $0.001 per image; with 10,000 daily searches, monthly expenses are around $300. Typical projects start from $5,000 for small catalogs, yielding ROI up to 300% within 6 months.

On-device vs. Server-side Embedding

We compared two approaches in practice.

On-device model. On iOS — the Vision framework with VNGenerateImageFeaturePrintRequest; on Android — ML Kit Image Labeling or a custom TFLite model via TensorFlow Lite Task Library. Advantage: works offline. Limitation: Apple's feature print is not compatible with server-side indexes, and accuracy is on average 15–20% lower than server models.

Server-side embedding. The image is sent to the server, processed through a model (CLIP), and a vector is returned for index search. This is more flexible — one index works with iOS, Android, and web. In practice, we choose the server-side approach with local pre-processing: the image is compressed and normalized on the device before sending. Contact us for a project assessment.

Server-side embedding achieves 90-95% accuracy, which is 3x better than on-device methods for catalogs with >10,000 products. The embedding model uses a 512-dimensional vector space with cosine similarity as the distance metric; quantization to 8-bit integers reduces memory usage by 4x with minimal accuracy loss.

How to Implement Image Search: Step-by-Step Guide

1. Catalog audit. Assess the volume (often 50,000–500,000 products), image types, and required accuracy.
2. Architecture selection. Determine whether to use server-side embedding, offline fallback, or a hybrid.
3. UI development. Implement camera, gallery, crop tool, and result display with similarity scores.
4. Vector index integration. Set up Qdrant or Weaviate, load reference embeddings.
5. Testing. Validate on real user photos (varying lighting, angles), A/B test on 1,000+ sessions.

Image Capture and Preparation

On iOS, for gallery selection use PHPickerViewController (not UIImagePickerController, deprecated since iOS 14). For camera use AVCaptureSession with AVCapturePhotoOutput. Pre-process the image before sending:

func prepareForSearch(image: UIImage) -> Data? {
    // Scale to 512px on longest side
    let maxDimension: CGFloat = 512
    let scale = maxDimension / max(image.size.width, image.size.height)
    let newSize = CGSize(width: image.size.width * scale,
                         height: image.size.height * scale)

    UIGraphicsBeginImageContextWithOptions(newSize, false, 1.0)
    image.draw(in: CGRect(origin: .zero, size: newSize))
    let resized = UIGraphicsGetImageFromCurrentImageContext()
    UIGraphicsEndImageContext()

    return resized?.jpegData(compressionQuality: 0.85)
}

On Android, use ActivityResultContracts.TakePicture() for camera and PickVisualMedia() for gallery (Photo Picker API, available since Android 13 and through registerForActivityResult via Jetpack).

Handling Partial Matches

The user can outline a fragment of the image — a crop tool is built into the UI. The cropped image is sent to the server, improving accuracy. Without this, searching by the full photo yields more false positives.

Server-Side Search: Vector Index

For nearest neighbor search on embeddings, we use Qdrant, Weaviate, or pgvector (if PostgreSQL is already in the stack). The CLIP model from OpenAI delivers good results for product search — it is trained on image-text pairs, thus working in both directions: from image to text and vice versa.

Qdrant is up to 5x faster than pgvector for large-scale searches. Server request with progress indicator:

// Android, Retrofit + OkHttp
suspend fun searchByImage(imageBytes: ByteArray): List<SearchResult> {
    val requestBody = imageBytes.toRequestBody("image/jpeg".toMediaType())
    val part = MultipartBody.Part.createFormData("image", "search.jpg", requestBody)
    return searchApi.visualSearch(part)
}

Important: handle cases where no close match is found (cosine distance exceeds threshold). Display 'nothing found' honestly, rather than returning irrelevant results from distant vectors. We guarantee logical transparency.

On-Device Preprocessing with CoreML / TFLite

If offline mode is needed or faster response is desired, we embed a lightweight model. MobileNetV3 or EfficientNet-Lite offer a reasonable compromise between accuracy and size. On iOS, convert to .mlmodel via coremltools; on Android, to .tflite. A local index is stored in SQLite with an extension for cosine distance or using Faiss via JNI/FFI. Order turnkey development — we will propose the optimal architecture.

Implementation Timeline

• Integration with an existing search API — 3–5 days.
• Full implementation (server + client) — 3–6 weeks, depending on catalog size (from 10,000 to 100,000 products).
• Project assessment — free within 1 day.

What’s Included

Deliverables include:

• Catalog audit
• Architecture selection
• UI development
• Vector index setup
• API integration
• Testing with real user photos
• Complete documentation
• Source code access
• Team training
• 30-day post-launch support
• Deployment instructions
• Demo video

With 5+ years of experience and over 50 successful visual search projects, we are ready to help. Contact us for a consultation on your use case. Get an estimate within 1 day.

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.