Velvet

Velvet is an AI-first data pipeline for software engineers to record and store large language model (LLM) requests to PostgreSQL. It lets you query, test and train AI features with fine-grained LLM data.

Some of the key features of Velvet include:

• Recording LLM calls: Record OpenAI requests in Postgres for analysis, testing, and training.
• Querying data: Get fine-grained access to LLM data without losing control.
• Forwarding logs: Forward LLM calls to any service with full data context.
• Analyzing performance: Drill down into usage and costs of models, features, and versions.
• Monitoring features: Catch errors, vulnerabilities, and feedback as you iterate.
• Testing and training: Experiment with data context and train custom models.

Velvet supports OpenAI and PostgreSQL, but the company plans to expand to other models and databases. Some common Velvet use cases include natively querying LLM calls to understand costs and usage, optimizing model usage by trying out prompts and historical data, testing and iterating on features locally, and forwarding requests to proxy services.

Velvet pricing begins with a free tier for individual contributors, with up to 10,000 requests per month, an AI SQL editor and email support. For early-stage startups, the "Startup" plan is $195 per month and includes up to 1 million requests per month, unlimited seats and email and Slack support. High-growth startups can sign up for a custom "Growth" plan with unlimited requests, premium support and compliance features.

Velvet is designed to make it easy to store and use LLM data, a process that's currently complex and that Velvet hopes will be more accessible and useful for teams to analyze, test and improve their AI features.

Published on June 13, 2024