Zod .refine() Rules Are Silently Dropped From JSON Schema Sent to LLMs

·1 views

Passing a Zod 4 schema to structured-output helpers such as the OpenAI Node SDK or Vercel AI SDK silently discards every .refine() rule. The SDKs convert schemas with z.toJSONSchema(), and the converter only maps checks it recognizes, while custom refinements are stored under an unrecognized name and skipped. The refinements still run when the response is parsed, but the model is never told about them. The author illustrates this with a refund extraction schema, showing that regex, positive-number, enum and date rules survive while a field-level refinement and a cross-field date check do not. Transforms are also lost in input mode, leaving only a plain string.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Call Center Jobs Are Shrinking, Yet Analysis of 14,000 Calls Finds Humans Still Beat AI

A report from BoldAgents examines call center employment, which has been declining as companies adopt AI tools. Its authors reviewed roughly 14,000 calls to compare how human agents and AI systems perform. The analysis concludes that human agents still outperform AI in these interactions. The source material is a brief Hacker News post linking to the full article, so further methodological detail is not available here.

0 comments Read more at Hacker News

ProgrammingDEV Community ·

Handling Edge Cases in TTS Applications: Chunking, Cleaning, and Rate Limits

Text-to-speech tools have advanced to sound natural, but real-world use exposes edge cases such as very long paragraphs, odd punctuation, emojis, code snippets, API rate limits, and mixed languages. The article argues that ignoring these issues erodes user trust in voice products. Its main technique is splitting long text on sentence boundaries to stay under per-request character limits, with short pauses between chunks and local caching for reuse. It also recommends preprocessing input to convert emojis and strip markdown so engines read them intelligibly. The examples use Python with the ElevenLabs API, and the piece goes on to discuss handling per-minute request caps.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Free AI Tools With No Trials or Credit Card Required, Tested for Real Use

A reviewer tested more than two dozen AI tools to find ones that remain free without expiring trials, credit card requirements, or tight generation caps. Tools were judged on whether a free tier could complete real tasks, how high its usage ceiling runs, and whether it retains user inputs for training. Most candidates failed on at least one point, including card-required sign-ups, short trials, and watermarked outputs. The author acknowledges that free tiers still carry fair-use limits such as daily message caps and peak-hour queues. The surviving tools are described with their specific free-tier limits so readers can judge how reliable each one is.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

SchemaGuard Detects Runtime Schema Drift and Enforces Data Contracts in Distributed Systems

The article introduces SchemaGuard, a runtime engine that detects schema drift and enforces data contracts across distributed architectures such as Kafka, gRPC, REST, and GraphQL. It argues that static schemas and registries fail when live payloads diverge, causing pipeline corruption, dead letter queue floods, and slow incident resolution. SchemaGuard runs as embedded middleware or a sidecar proxy, inspecting traffic with low-overhead parsing and flagging both structural mismatches and subtler semantic changes such as range shifts or precision loss. Its shadow mode compares production traffic against API specifications without blocking requests, while quarantine gateways can isolate malformed payloads and propose corrective schema changes.

0 comments Read more at DEV Community