January 2025 Summaries

DeepSeek-R1 is a cheaper and open-source model that excels at agentic reasoning, superior multilingual capabilities, large context windows, and generalization across domains. It was tested in real-world examples with Inngest as an orchestrator library and interface to explore its Multi-Lingual marketing content generator, Agentic arXiv Research Assistant, and other use cases. The model demonstrated strong performance in understanding and generating content in multiple languages while maintaining cultural nuances. Its large context window capability enabled it to provide relevant search queries for arXiv research papers, even when asked questions in French. However, the model still lacks support for tool calling and system prompts, which are essential for creating agentic applications.

The tutorial demonstrates how to create a custom Hacker News summary app using AI agents, workflows, and Inngest. The app is designed to periodically search for relevant stories based on user-defined topics and questions, generate embeddings of the content, and send an email summary. The app's architecture consists of three main components: a PostgreSQL database with pgvector extension, an indexer cron job that fetches new Hacker News stories, and a full-stack Next.js web app that hosts the UI and backend logic. Inngest orchestrates the workflow, which includes two AI agents: a Search Agent that uses tools to extract relevant information from the vector database, and a Summarizer Agent that summarizes the search results. The app can be deployed on Render and extended with additional features.

Ingest is being used to build agentic workflows with Weaviate, a vector database that enables flexible and robust AI applications. The Dinner Generator workflow uses Ingest's composable API to create an agentic pipeline that generates the perfect dinner menu based on tastes, allergies, and the number of attendees. The workflow relies on three datasets available on Huggingface and leverages Weaviate's batching and vectorizer APIs to quickly process large amounts of data. Ingest enables the creation of AI workflows with simple TypeScript API, while Weaviate enables quick retrieval of data from vectorized collections. The agentic nature of the workflow resides in its ability to fetch additional information when allergies are mentioned in the food preferences. The combination of Weaviate and Ingest makes it easy to build such agentic and data-intensive workflows.