September 2024 Summaries

Anthropic, a leading AI research company, has introduced Claude Sonnet 3.5, which is quickly becoming the default large language model (LLM) due to its intelligence, speed, and cost-effectiveness. In this article, we learn how to use Claude Sonnet 3.5 and pgvector to build an AI image gallery that lets users search for images and ask questions. The architecture includes a vector database hosted on Timescale Cloud, which allows efficient indexing and querying of high-dimensional vectors. We also explore the Flickr30k dataset as a source of diverse images for our smart image gallery application. By combining these tools with OpenAI's CLIP model, we can create an AI-powered image search gallery that retrieves similar images based on a text query and uses them as context for Sonnet 3.5.

The AI chatbot race initiated by OpenAI with the release of ChatGPT is seeing new competitors like Gemini, Cohere, and many others. Anthropic, a prominent AI research company dedicated to developing safe and ethical AI systems, is making waves in the field with its Claude family of large language models (LLMs). The new Claude launch is a notable upgrade from its predecessor and claims it can surpass OpenAI's GPT-4o model on key benchmarks like GPQA (Graduate-Level Google-Proof Q&A), multilingual math (MGSM), and more. In this article, we discussed Claude Sonnet 3.5, its strengths and usefulness, and compared it to other models in the Claude family. We also used Sonnet 3.5 and pgvector to build a retrieval-augmented generation (RAG) application.

Timescale's tiered storage architecture is designed to address scalability challenges in managing PostgreSQL databases with massive amounts of data. It enables seamless management by moving older, less frequently accessed data to cheaper, slower storage without sacrificing query performance. The tiered storage backend allows users to move data from the high-performance tier (block storage) to the low-cost bottomless storage tier (Amazon S3 object storage). Recent optimizations have led to up to 400x performance improvements for point queries, making it easier to write flexible queries while still benefiting from a significant performance boost.

Postgres introduced columnar compression in 2019, which significantly reduced storage costs and boosted query performance. However, frequent updates, inserts, and deletes on compressed data posed a challenge to performance. With TimescaleDB 2.16.0, significant improvements have been made, delivering up to 500x faster updates and deletes and 10x faster upserts on compressed data. These optimizations make compressed data behave more like uncompressed data without sacrificing performance or flexibility. The compression tuple filtering feature allows for efficient filtering of data at multiple stages during DML operations, reducing the amount of work needed to execute queries and improving overall application performance.

A developer's customer challenge often becomes their own. In this case, one of the customers, Ndustrial, an industrial energy optimization platform, reported that they had to resort to workarounds due to suboptimal PostgreSQL upsert performance when using TimescaleDB. The solution was a 300x increase in upsert performance over compressed data by modifying the upsert process to use existing indexes to find all batches of rows that need to be decompressed. This optimization not only resolved the immediate performance issues but also opened the door for Ndustrial to manage their data more efficiently and confidently.

PostgreSQL partitioning is a technique that divides large tables into smaller, more manageable pieces called partitions, enhancing performance and efficiency. However, not all queries are ideally suited for partitioning, especially when non-partitioned columns are used in WHERE clauses. To address this issue, TimescaleDB 2.16.0 introduces chunk-skipping indexes that can be defined on columnar sections of hypertables for specific data types. These indexes track the minimum and maximum values for a column in each chunk, allowing exclusion of chunks where queries would find no relevant data. This feature results in 7x faster performance with an 87% smaller storage footprint, thanks to compression.

TimescaleDB has introduced new features designed to improve performance, including chunk-skipping indexes for faster queries and compressed tuple filtering for more efficient updates and deletes. These optimizations can result in up to 7x faster queries, 500x faster updates, and 400x faster access to tiered data stored in S3. The company plans to explore real-world use cases and delve deeper into the technology behind these performance improvements.

Postgres has evolved into a versatile database over the last two decades. However, today's applications often require handling high-performance workloads and real-time analytics. TimescaleDB shines in this area by empowering developers to address both their relational and real-time analytics needs within the trusted Postgres platform. The hybrid row-columnar storage engine, hypercore, allows TimescaleDB to automatically handle high-speed ingestion of new data and efficient querying of large datasets while maintaining flexibility and performance for real-time workloads. Key capabilities of hypercore include chunk micro-partitions, SIMD vectorization, skip indexes, and compression.

- Developed a question-answering system over GitHub issues using OpenAI's Parallel Tool Calling API. - Implemented text-to-SQL tools and semantic search tools to build the question-answering system. - Used Pydantic models for separation of concerns, making it easier to evaluate tool selection separately from implementation. - Utilized Instructor for testing function calling capabilities before moving to implementation. - Discussed best practices for text-to-SQL generation and provided a prompt example for TimescaleDB-specific query generation.

This case study demonstrates how pgai, an AI extension for PostgreSQL, can be used to improve customer satisfaction and operational efficiency in a virtual pizza shop scenario. By integrating AIGC with ChatGPT, businesses can leverage AI-driven insights to enhance their CRM operations. The process involves running pgai, collecting and preparing customer feedback data, embedding the feedback, asking questions, generating answers, and finally creating a business report. This approach allows for quick identification of potential issues and helps improve overall customer satisfaction and brand reputation.

In-Database AI Agents are computer programs designed to interact with their environment by perceiving data, interpreting it through algorithms, and acting on it to achieve specific goals. With the advent of large language models (LLMs), AI agents can now combine LLMs with execution chains to perform specific actions. The Claude family of LLMs is equipped to use external tools and APIs, making it easier to develop custom AI agents. In this article, we show how to create AI agents within PostgreSQL's database using the pgai PostgreSQL extension to reduce latency and simplify architecture. AI agents serve many benefits in various industries such as finance, power industry, transportation, and healthcare. The combination of pgai with Claude offers enhanced efficiency, scalability, and cost-effectiveness.