Table of Contents
ShapedQL (by Shaped.ai)
The SQL Engine for AI Relevance
Stop gluing together Pinecone, Redis, and fragile Python scripts. ShapedQL is a specialized SQL engine designed to handle the complex logic of “relevance.” It powers “For You” feeds, semantic search, and RAG memory in minutes rather than months. Instead of managing infrastructure, developers write simple SQL queries that the engine compiles into real-time ranking pipelines. These pipelines handle retrieval, filtering, scoring, and reordering based on live user behavior—replacing thousands of lines of maintenance code with about 30 lines of declarative SQL.
Key Features
- SQL-Based Relevance: Define complex ranking logic (e.g., “boost trending items but hide seen ones”) using familiar SQL syntax.
- 4-Stage Pipeline: Automatically handles the standard RecSys flow: Retrieve (Candidates) → Filter (Rules) → Score (ML Models) → Reorder (Diversity).
- Automated MLOps: The system trains, deploys, and refreshes models automatically based on your data stream, removing the need for a manual training pipeline.
- Multi-Modal Embeddings: Native support for vectorizing text, images, and user sessions without external embedding services.
User Workflow
- Connect: Stream data from your warehouse (Snowflake, BigQuery) or event bus (Kafka, Segment).
- Query: Write a ShapedQL query to define how items should be ranked and filtered (e.g.,
SELECT * FROM items ORDER BY likelihood_to_click). - Deploy: The query becomes a real-time API endpoint that your app calls to fetch personalized results.
Use Cases
- “For You” Feeds: creating TikTok-style social feeds that adapt to user interactions in real-time.
- Semantic Search & RAG: Building search bars that understand intent and retrieving context for LLM agents.
- E-commerce Ranking: Re-sorting product grids to maximize conversion based on user history and inventory status.
Pros & Cons
- Pros: Drastic reduction in code complexity (declarative vs. imperative), “Glass Box” visibility (you see why items are ranked), predictable flat-fee pricing.
- Cons: Requires SQL knowledge (not a no-code tool), smaller ecosystem than massive clouds like AWS, overkill for simple keyword search.
Pricing
- Self-Serve: Flat monthly fee (tiered by usage volume).
- Enterprise: Custom volume pricing with white-glove support.
- Note: Shaped emphasizes a flat-fee model to avoid the “bill shock” common with per-request pricing like Algolia.
How Does It Compare?
vs. Algolia
Algolia is the leader in keyword search. It is fast and easy but struggles with deep personalization and “discovery” (showing users things they didn’t search for). Algolia also charges per-request, which gets expensive at scale. Shaped is built for ranking and personalization first, using a flat-fee model that encourages heavy usage without penalty.
vs. AWS Personalize
AWS Personalize is a “Black Box” service. You throw data in, and recommendations come out, but you have very little control over why or how it works. ShapedQL is a “Glass Box”—you explicitly define the logic using SQL, giving you granular control over the ranking rules while still automating the hard machine learning parts.
vs. Building Yourself (Pinecone / Weaviate / Redis)
Using a vector database (Pinecone) is just one piece of the puzzle. To build a feed, you still need to write the logic to fetch vectors, rerank them, filter duplicates, and manage user history. ShapedQL replaces this entire “glue code” layer, providing a managed engine that sits on top of the raw data so you don’t have to build the orchestration yourself.
Final Thoughts
ShapedQL represents a shift from “Managing Infrastructure” to “Managing Logic.” For engineering teams that need sophisticated personalization (like TikTok or Instagram) but lack a team of 50 ML engineers, Shaped provides the perfect middle ground: the power of a custom build with the simplicity of a managed service. If you are tired of debugging complex Python pipelines just to sort a list of items, ShapedQL is the modern solution.