ShapedQL

Product Introduction

1. Definition: ShapedQL is a specialized SQL engine for real-time relevance engineering, categorized under AI/ML infrastructure. It transforms SQL syntax into scalable pipelines for retrieval, ranking, and personalization tasks.
2. Core Value Proposition: It eliminates the need to stitch together fragmented tools (Pinecone, Redis, Python) by enabling developers to build "For You" feeds, semantic search, and RAG memory systems in minutes using SQL.

Main Features

1. Real-Time Ranking Pipelines:

   • How it works: Compiles SQL queries into low-latency pipelines that retrieve data, apply filters (e.g., genre filters), score results using ML models (e.g., predicted CTR), and reorder outputs based on live user interactions.
   • Technologies: Uses automated MLOps for model deployment and supports real-time collaborative filtering (e.g., collaborative_embedding).

2. Native Multi-Modal Embeddings:

   • How it works: Integrates pre-built embeddings for text (title_embedding), images (poster_embedding), and user behavior (people_also_liked), allowing hybrid queries like semantic + collaborative search in one SQL statement.
   • Example: text_search(query='$query', text_embedding_ref='title_embedding') combines keyword and vector search.

3. Declarative Relevance Engineering:

   • How it works: Replaces thousands of lines of Python/Redis code with 30-line SQL scripts. Supports parameterized templates (e.g., $query) and output formats (Masonry/List/Carousel) for rapid iteration.

Problems Solved

1. Pain Point: Fragmented infrastructure requiring manual integration of vector databases (Pinecone), caching layers (Redis), and custom scripts for ranking logic.
2. Target Audience:
   • ML engineers building recommendation systems.
   • Full-stack developers creating personalized feeds.
   • Data teams implementing RAG for AI agents.
3. Use Cases:
   • Real-time movie recommendations (e.g., Movielens dataset).
   • Hybrid e-commerce search combining text, images, and user behavior.
   • Breaking filter bubbles via multi-embedding diversification.

Unique Advantages

1. Differentiation: Unlike standalone vector databases (Pinecone) or LLM frameworks, ShapedQL unifies retrieval, scoring, and ranking in a single SQL interface—reducing DevOps overhead by 90%.
2. Key Innovation: SQL-to-pipeline compilation optimizes latency (e.g., 93ms for 50 results) while automating MLOps for embeddings and scoring models (e.g., click_through_rate).

Frequently Asked Questions (FAQ)

1. How does ShapedQL reduce RAG implementation complexity?
   ShapedQL replaces document retrieval scripts with SQL-native hybrid search, enabling real-time context injection for AI agents using collaborative + semantic embeddings.
2. Can ShapedQL handle visual data like product images?
   Yes, its poster_embedding feature converts images to vectors, allowing visual similarity searches in SQL queries alongside text/collaborative filters.
3. What makes ShapedQL faster than custom Python pipelines?
   Compiled SQL pipelines execute retrieval, scoring, and reordering as optimized native code, avoiding serialization overhead between Redis/Python/Pinecone.
4. Is ShapedQL suitable for high-traffic recommendation systems?
   Yes, automated MLOps and parameterized queries support scaling to millions of user interactions with sub-100ms latency.