July 3, 2025

Building AI-First Apps with Harper using Fast Semantic Search

Welcome to Community Posts
Click below to read the full article.
Arrow
Summary of What to Expect
Table of Contents

Today's AI-first applications live or die by their ability to understand meaning, not just match keywords. Whether you're building a RAG system that needs to find relevant context for an LLM, powering semantic search across documentation, or creating recommendation engines that truly understand user intent, vector embeddings have become the foundation of intelligent retrieval. These mathematical representations capture the semantic relationships between concepts, enabling applications to find "similar" content in ways that traditional keyword search simply can't match. But here's the problem: while embeddings unlock semantic understanding, searching through them at scale becomes a computational nightmare that can bring even the most promising AI application to its knees.

The challenge hits fast and hard. For a few thousand vectors, brute force similarity search would work fine. But scale to hundreds of thousands or millions of embeddings, and suddenly your "intelligent" search takes seconds instead of milliseconds, destroying user experience and making real-time applications impossible. Most developers quickly discover the brutal tradeoff: you can have fast vector search, or you can have accurate results, but getting both requires either accepting significant compromises or adding yet another specialized service to your already complex architecture. This forces teams into an uncomfortable choice - sacrifice the semantic intelligence that makes their AI application special, or accept the performance penalties and infrastructure complexity that come with dedicated vector databases. The very capability that should differentiate your AI application becomes its biggest bottleneck.

The breakthrough that changed everything for vector search is an algorithm called HNSW, short for Hierarchical Navigable Small Worlds. Unlike earlier approaches like LSH or IVF that struggled to balance speed and accuracy, HNSW introduced a graph-based structure that makes searching large sets of vectors both fast and precise. It organizes data into multiple layers of small-world graphs, allowing the algorithm to efficiently navigate through the vector space and zoom in on the nearest neighbors without scanning everything. The result is logarithmic search time with high recall, even as your dataset scales into the millions.

Today, HNSW is the gold standard. It’s the default engine behind systems like FAISS, Vespa, and now, with the release of 4.6, it’s built directly into Harper’s core. This is a big win for performance and developer experience. HNSW makes vector search fast enough to support real-time UX and accurate enough to power mission-critical AI features. No more offloading to external vector stores or sacrificing quality for latency. With HNSW built in, Harper gives developers a way to do semantic search that actually keeps up with their users, and their product roadmap.

What makes Harper’s approach different is that vector indexing is now native to its core runtime. Not an add-on, not a plugin, and definitely not a separate service. You can index any vector-valued property, whether it comes from text, images, or multi-modal embeddings, and query it right alongside metadata, application logic, and pub-sub events. Because everything (database, vector engine, cache, messaging, and even hosted functions) lives inside a single process, there’s no network overhead or context-switching between layers of your stack. That gives developers full query flexibility with higher performance, and it shortens the distance between data, inference, and action.

Harper’s native vector indexing unlocks a wide range of practical, high-impact AI use cases that previously required complex multi-service setups. Here are some examples:

  • You can now build semantic document search directly into your application, retrieving content that aligns with user intent even when there’s no keyword match. 
  • For teams building retrieval-augmented generation (RAG) pipelines, Harper enables real-time vector lookups feeding LLMs with relevant context without compromising speed. 
  • It also supports similarity detection across massive datasets, making it easy to identify near-duplicates, group related content, or detect patterns. 
  • Personalized experiences become more powerful too, as you can compare user behavior or preferences as vectors, not just tags or categories. 
  • And because Harper can store and query multi-modal embeddings (text, images, audio, and more) you can power rich, unified search experiences across different content types. 

All of this happens in a single query layer, using intuitive APIs that feel like working with standard JSON data but deliver intelligence at scale.

The future of software is undeniably AI-native, but building truly intelligent applications still depends on having the right infrastructure behind the scenes. One that doesn’t just support AI, but is designed for it. Harper 4.6 brings that vision closer to reality by combining structured data, vector search, real-time messaging, caching, and app logic into a single, high-performance runtime. This isn’t just about convenience, it’s about giving developers a coherent, low-latency environment where semantic intelligence is built-in, not bolted on. Instead of juggling multiple services and tradeoffs, teams can now prototype faster, scale smarter, and deliver AI experiences that are both context-aware and lightning-fast. If you’re building apps that think in embeddings, Harper now thinks with you. Try it here, and we can’t wait to see what you build.