Dev builds multilingual chat widget using local ONNX embeddings to cut API costs

·1 views

A developer created 'Ask Leya', a business-focused chat widget that answers queries from a company's own content and escalates to a human agent when needed. Unlike conventional chatbots, it uses the paraphrase-multilingual-MiniLM-L12-v2 model run locally via ONNX to embed queries across 50+ languages without external API calls or per-query embedding costs. The stack combines Node.js, Express, Postgres with pgvector, Socket.IO for live handoffs, and gpt-4o-mini via OpenRouter for generating responses. Key engineering challenges included managing a shared database with unreleased modules, handling async rejections in Express 4, and a 90-second production hang caused by a macOS binary being deployed on a Linux server. Integrations for WhatsApp, Instagram, and Messenger are complete but held back pending a polished self-serve release.

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)

How USDC Escrow Contracts Enable Trustless Payments for AI Agent Services

Developers building autonomous AI agents face a core payment problem: neither the payer nor the agent can safely transact without risking non-payment or fund theft. A smart contract escrow solution locks USDC tokens until a verifiable proof of task completion is submitted, removing the need for a trusted intermediary. The approach, detailed for the Base blockchain, uses an ERC-20-compatible Solidity contract that transitions through defined states such as Funded, Completed, and Refunded. A timelocked refund mechanism protects payers if the agent fails to deliver, while the agent is guaranteed payment upon verified completion. The design deliberately avoids complex features like multi-signature or upgradeability to minimize the attack surface and keep gas costs predictable.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

FlatBB Launches Multilingual PHP Forum Software With No Build Tools Required

FlatBB is a new open-source forum platform built on plain PHP 8.1+ that requires no Composer, Docker, or build steps, making it accessible for small communities. The software ships with eight built-in interface languages — including Persian with full right-to-left layout support — and allows both visitors and registered members to set their preferred language. Core features include nested categories, group permissions, Markdown editing, full-text search, image uploads, notifications, and a responsive three-column layout with dark mode. An admin panel covers user management, plugins, scheduled jobs, and layout customization, while a marketplace extends functionality with SEO tools, social login, two-factor authentication, and more. FlatBB is released under the MIT license and is available for download on its official website and GitHub.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

ingress-nginx Retired in 2026: Key Pitfalls Teams Face When Migrating to HAProxy

The ingress-nginx project was officially retired in March 2026, ending all releases and security patches despite being used in roughly half of all Kubernetes clusters. Installations of the retired project proceed without any warnings, meaning automated pipelines can continue deploying it silently. A hands-on migration test revealed that switching to HAProxy Kubernetes Ingress Controller is not simply an annotation-mapping exercise — subtle but serious issues emerge around IP address forwarding and protocol headers. Each network hop rewrites the source address, causing the original client IP to be buried under intermediate addresses, while X-Forwarded-Proto is incorrectly set to HTTP even for HTTPS connections, potentially breaking redirects and secure cookies. An additional practical hazard was found in Cloudflare IP list generation, where a missing newline between IPv4 and IPv6 ranges causes HAProxy to reject its own configuration at load time.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Mostik Claims AI Models Can Share Thoughts Directly, Skipping Text Entirely

Startup Mostik has developed a 'latent bridge' that allows a large 753-billion-parameter AI model to pass internal hidden states directly to a smaller 4-billion-parameter model, bypassing text generation entirely. The company claims this approach retains 80% of the larger model's accuracy while delivering 20 times faster performance at lower inference costs. Unlike standard AI handoffs where one model generates text that another then reads, Mostik's protocol transfers raw numerical representations beneath the language layer, with neither model's original weights modified. However, Mostik has not released reproducible benchmarks, and key technical details — including which layers are captured and how representations are mapped between models — remain undisclosed. The approach raises both efficiency promises and transparency concerns, as the internal communication channel cannot currently be inspected by engineers.

0 comments Read more at DEV Community