SShortSingh.

Programming

0
ProgrammingDEV Community ·

Vectorless RAG Uses Document Structure Instead of Embeddings for Smarter Retrieval

Traditional Retrieval-Augmented Generation (RAG) splits documents into chunks, converts them into vector embeddings, and retrieves the most semantically similar chunk to answer a query. However, this approach has known limitations, including fragmented context from chunking, semantic similarity not always equating to relevance, and opaque retrieval that lacks traceability. Vectorless RAG addresses these issues by treating a document as a structured hierarchy — similar to a table of contents — rather than a flat collection of chunks. An LLM navigates this document tree using reasoning to identify the most relevant section, enabling answers to be traced back to their exact source. The two approaches are not mutually exclusive; depending on the use case and data type, traditional vector search and Vectorless RAG can be used together.

0
ProgrammingDEV Community ·

Material Foundation Package Brings Adaptive Layouts to Flutter Apps Across Devices

A lightweight Flutter package called Material Foundation helps developers build interfaces that adapt across mobile, tablet, and desktop screen sizes. The package provides two key widgets — DynamicLayoutBuilder and DynamicScaffold — which select appropriate layouts based on available screen width. It uses Flutter's LayoutBuilder constraints, switching between mobile, tablet, and desktop compositions at customizable breakpoints of 740px and 1200px. Developers supply separate widget bodies for each screen size, keeping adaptive logic cleanly separated from layout details. The package is available via GitHub and integrates into standard Flutter projects through pubspec.yaml.

0
ProgrammingDEV Community ·

Seven Infrastructure Patterns That Keep AI Apps Running at Scale

Software developer Maneshwar, building an AI code review tool called LiveReview, outlines seven architectural patterns critical to running AI applications under heavy user load. The core argument is that a well-tuned AI model means little if the surrounding infrastructure cannot handle thousands of simultaneous requests. Key patterns discussed include deploying an API gateway to manage incoming traffic, applying admission control to cap not just request counts but also token usage and concurrent model calls, and using caching to avoid redundant retrieval and embedding work for repeated queries. The article distinguishes between an API gateway, which governs user traffic into the app, and a model gateway, which manages outbound calls to providers like OpenAI or Anthropic. It emphasizes that production failures at scale typically stem from infrastructure bottlenecks rather than model quality, making these engineering patterns essential for any AI product moving beyond demo stage.

0
ProgrammingDEV Community ·

Why Your .NET App Silently Fails: The TCP Port Exhaustion Trap

ASP.NET Core applications can suffer mysterious outbound API failures even when CPU, memory, and database metrics appear normal, with the root cause hidden at the TCP layer. Each new HttpClient instance created per request opens a fresh TCP connection requiring a unique local port, and closing connections does not immediately free those ports due to TCP's TIME_WAIT state. Under high traffic, an application can exhaust the finite pool of available outbound ports, triggering SocketException errors that disappear temporarily after a restart. Microsoft explicitly warns against creating and disposing HttpClient instances per request, recommending instead that developers reuse HttpClient or use IHttpClientFactory to share connection pools. Understanding this TCP lifecycle explains why the pattern is dangerous and helps developers recognize the same failure even across different technologies or frameworks.

0
ProgrammingDEV Community ·

Bug report blamed wrong tool and gave false error code, hiding real anchoring flaw

A code verification tool flagged three of eight rows as mismatched and blamed a neighbouring tool, citing a BAD_SHAPE error — but both accusations were wrong. Investigation confirmed the accused tool never accessed the relevant file, ruling it out immediately. The real cause was that the verifier's own row-anchoring keys secretly embedded line numbers, despite its specification explicitly forbidding line-number anchoring. When an unrelated edit shifted a code block from line 2490 to 2515, the keys broke and rows went unmatched — yet the error code pointed developers toward rewriting perfectly valid rows. Two fixes were applied: stripping line numbers from keys and enforcing a strict one-block-per-row match, while deliberately reusing an existing error code to preserve the integrity of historical diagnostic comparisons.

0
ProgrammingHacker News ·

Real-SWE Benchmark Tests AI Coding Models on Private Enterprise Codebases

A new benchmark called Real-SWE has been introduced to evaluate AI models on real-world, private enterprise codebases. Unlike existing benchmarks that rely on public repositories, Real-SWE aims to reflect the complexity of actual software engineering environments. The initiative was shared on Hacker News, where it attracted modest early attention with 20 points and 5 comments. The benchmark is published by Specific, accessible via their website, and targets a more rigorous standard for assessing AI coding capabilities in professional settings.

0
ProgrammingDEV Community ·

Developer builds Python CLI music player using Deezer API, yt-dlp, and SQLite

A developer built a command-line music player in Python as a practical alternative to typical beginner projects like to-do apps. The tool integrates with the Deezer API for song and artist search, uses yt-dlp and VLC for audio playback, and stores playlists and favorites in a local SQLite database. During development, a persistent 'database locked' error turned out to be caused by a leftover connection from an older version of the app still running in the background. The fix involved killing the rogue process and implementing proper connection closing on exit. The developer has shared the project on GitHub and plans to build a GUI version next.

0
ProgrammingDEV Community ·

Node.js CSV diff toolkit rejects duplicate IDs before comparing records

A lightweight Node.js toolkit (csv-toolkit v0.2.0) has been released to compare CSV snapshots by matching records on named key columns rather than row position or column order. The tool deliberately halts any comparison when it detects duplicate or blank key values, preventing ambiguous record pairing. Running the included demo generates nine output files — including HTML audit and change reports — showing added, removed, changed, and unchanged records across two snapshots. An audit function also flags data quality issues such as missing fields or invalid numeric values without modifying the source files. The toolkit has no third-party dependencies and requires Node.js 22 or later to run.

0
ProgrammingDEV Community ·

Generative AI Writes, Agentic AI Acts: Why the Difference Matters

A developer reflection sparked by an IBM/Coursera course on agentic AI draws a sharp line between generative and agentic AI systems. Generative AI takes a prompt and returns text — it translates, summarizes, or drafts code, but leaves no trace in any system. Agentic AI, by contrast, operates in a loop: it observes, decides, calls tools, and takes real-world actions like opening pull requests, calling APIs, or processing payments. The author warns that mislabeling generative outputs as 'agentic' is a common mistake, and that acting — unlike generating — carries genuine consequences. Security considerations such as least privilege, human-in-the-loop controls, and audit trails become essential once an AI model can write to the world, not just describe it.

0
ProgrammingDEV Community ·

LangChain vs LangGraph: Choosing the Right Orchestration Layer for AI Agents

A developer completing IBM's Agentic AI course on Coursera draws a practical distinction between LangChain, LangGraph, and IDE-based agent harnesses. LangChain uses a linear chain model — input, output, done — with no built-in loop, persistence, or human-in-the-loop support. LangGraph adds stateful nodes, conditional cycles, checkpointing, and pause-and-resume capabilities, making it suited for true agentic workflows. The author notes that using LangChain's agent shortcut does not eliminate the graph layer; the runtime still runs one under the hood. The key takeaway is that teams should define their flow's structure before picking a library, since a stateless invoke pipeline is not an agent, regardless of which tool is used.

0
ProgrammingDEV Community ·

Developer Launches Free Browser-Based File Transfer Tool for Cross-Platform Use

A developer has built nowiretransfer.xyz, a free, browser-only file and text transfer tool designed to work across any two devices regardless of operating system. The tool requires no app installation or account creation, and supports file transfers up to 250 MB. It uses peer-to-peer technology over local networks, which the developer claims is faster than Bluetooth and avoids the upload-download delays of cloud storage services. All transfers are AES-256 encrypted in transit, and the platform collects no user data beyond what is needed to keep the service running. The developer is seeking community feedback on edge cases such as unusual network configurations, large files, and browser compatibility issues.

0
ProgrammingDEV Community ·

Developer Builds Dubai Business Directory by Tiling City into 1,760 Grid Segments

A developer in Dubai built an open-source business directory toolkit after finding no existing tool could comprehensively browse or filter Google Maps data for an entire city. The core challenge was Google Maps' hard limit of roughly 200 results per query, which made a single API call insufficient to enumerate all businesses in a large city. To work around this, the developer divided Dubai into 44 geographic squares crossed with 40 business categories, creating 1,760 tile-and-category pairs that were crawled using SearchApi's Google Maps engine. A total of 1,400 API requests returned 15,246 unique businesses, with the pipeline using place_id as a stable deduplication key and GPS coordinates to reassign each business to its actual location. The MIT-licensed toolkit, deployed live at directory.pooyagolchian.com, was built as part of SearchApi's developer ambassador programme and is designed so that pagination depth is the primary cost variable.

0
ProgrammingHacker News ·

Study Compares CadQuery and OpenSCAD Performance for AI-Driven CAD Tasks

A benchmark study published on ModelRift's blog evaluates CadQuery and OpenSCAD, two popular open-source CAD tools, for use in agentic or AI-assisted design workflows. The comparison examines how well each tool performs when driven programmatically by automated agents rather than human users. CadQuery is a Python-based scripting library, while OpenSCAD uses its own declarative language, making them distinct approaches to code-based 3D modeling. The study aims to help developers and researchers choose the most suitable tool when building AI-powered CAD pipelines. The article attracted modest early attention on Hacker News with six points and no comments at the time of posting.

0
ProgrammingDEV Community ·

Developer releases browser-only LLM chat interface with no backend or Docker required

A developer has built 'Lab', a serverless, client-side web application designed as a lightweight alternative to heavyweight open-source LLM chat interfaces. Unlike typical setups that require Python or Node.js backends, databases, and multi-container Docker deployments, Lab runs entirely within the user's browser. All data — including chats, API keys, and model configurations — is stored locally using IndexedDB via Dexie.js, with nothing sent to external servers. The tool supports multiple AI providers including OpenAI, Anthropic, Google Gemini, OpenRouter, and local Ollama models, and can be installed as a Progressive Web App on desktop or mobile. The project is open-source under the MIT license, with a live demo available at labstudio.tech and source code hosted on GitHub.

0
ProgrammingDEV Community ·

7 Key Checks Before Applying for a Tech Job in 2026

The tech job market in 2026 is evolving rapidly, with AI tools reshaping workflows and employers updating the skills they expect from candidates. Experts advise job seekers to look beyond a posting's title and carefully review actual day-to-day responsibilities, required technologies, and the broader tech stack involved. Comparing multiple job listings for similar roles can help identify which skills are consistently in demand, giving candidates a clearer picture of market expectations. Understanding the company's business context — including its products, customers, and engineering team size — is also recommended before submitting an application. The overall advice is to apply selectively and strategically rather than broadly, focusing on roles that genuinely align with one's existing skills and career goals.

0
ProgrammingDEV Community ·

Seed4J Validates Tech Stack Compatibility by Auto-Generating 25 Project Configurations

Seed4J is an open-source tool designed to address the persistent challenge of dependency and framework compatibility across modern software projects. It uses Renovate and GitHub Actions to automatically check for library updates daily, merging minor changes only after all required status checks pass. The tool currently generates 25 distinct projects spanning a wide range of technology combinations, including Angular, Vue, React, Spring Boot, and multiple databases such as PostgreSQL, MongoDB, MySQL, and Cassandra. Six of these configurations intentionally deviate from defaults — for example, using Gradle, Java 26 early access, or pnpm — to test compatibility at the edges. The project landscape is publicly visible at start.seed4j.com and was last captured on September 12, 2026.

0
ProgrammingDEV Community ·

Bol's dev team shares hard lessons from migrating Backstage platform to pnpm

Developers at Dutch retailer bol migrated their internal Backstage developer platform from Yarn 4 to pnpm after five years, seeking better support for parallel Git worktree workflows used with coding agents. While the initial switch — pinning pnpm, replacing commands, and committing a new lockfile — was straightforward, CI pipelines repeatedly downloaded all 4,226 packages from scratch despite a cache being in place. The root cause was a mismatch between where pnpm wrote its store and where GitLab expected to find it, resolved by explicitly setting the store directory within the CI configuration. A further issue emerged when the cached archive ballooned to 1.42 GB across roughly 601,000 files because node_modules trees were included unnecessarily; dropping those reduced cache size by about 72%. The team concluded that package manager migrations surface hidden architectural assumptions, particularly around cache paths and what is actually worth caching.

0
ProgrammingDEV Community ·

Analysis of 285,000 Transactions Reveals Fraud Peaks at 2AM, Not 11AM

A data analyst examined the widely used Kaggle Credit Card Fraud Detection dataset, comprising 284,807 anonymized European transactions from September 2013, of which only 492 — just 0.173% — were fraudulent. The analysis found that fraudulent transactions tend to cluster at low amounts, rarely exceeding a few hundred dollars, suggesting fraudsters deliberately avoid large sums that trigger alerts. An initial hour-by-hour breakdown appeared to show fraud spiking at both 2AM and 11AM based on raw case counts, but recalculating by fraud rate per transaction volume told a different story. Hour 2 proved a genuine high-risk window at a 1.71% fraud rate — roughly 10 times the dataset average — while Hour 11's apparent spike largely reflected higher transaction volume rather than elevated risk. The project highlights a key analytical pitfall: raw counts can be misleading without normalization, a principle the analyst compared to reviewing financial margins without examining both numerator and denominator.

0
ProgrammingDEV Community ·

Why Frontend Developers Need Stateful Mock APIs Over Static Ones

Frontend developers often rely on static mock APIs like JSONPlaceholder during backend development, but these tools fail to persist data across requests, causing issues such as newly created items vanishing after a page refresh. Static mock servers simply echo responses without storing any state, meaning POST, PUT, PATCH, and DELETE operations have no lasting effect on subsequent GET calls. This limitation breaks modern frontend patterns like query caching, optimistic UI updates, and reactive state management. Stateful mock APIs address this by maintaining a per-session overlay layer that merges mutations with baseline seed data, simulating real backend behavior. Tools built on this architecture allow developers to test full CRUD flows, conditional UI states, and multi-step navigation without needing an actual database.

0
ProgrammingDEV Community ·

How One Developer Ships a Real Mobile App Using an Entirely Free Tech Stack

A developer has built and deployed a fully functional mobile app with real users without incurring any monthly infrastructure costs. The stack combines Expo for mobile builds, Supabase for database and authentication, Render for background jobs, and Vercel for landing page hosting — all on free tiers. Each service comes with limitations, such as Supabase pausing idle projects after seven days and Render services experiencing slow cold starts after 15 minutes of inactivity. The developer works around these constraints using scheduled pings and strategic build timing rather than paid upgrades. The setup demonstrates that, as of 2026, free tiers from major cloud platforms are generous enough to support a production-grade mobile app from launch onward.

← NewerPage 1409 of 3335Older →