How to Build Portable Semantic Search for Private SaaS Docs with RAG

·1 views

Developers building private fintech knowledge bases can use a four-step pipeline: chunking documents, generating embeddings, retrieving relevant passages, and grounding chat completions in those passages with cited chunk IDs. Access-control filters must be applied at the retrieval stage before any content enters the prompt, as serving answers from inaccessible policy documents is a security failure regardless of accuracy. Reranking between retrieval and prompt assembly is optional and should only be added when evaluation data shows first-pass semantic search is missing useful results. For provider flexibility, teams can choose between direct API access to OpenAI, Anthropic, or Gemini, or use aggregation layers like OpenRouter, LiteLLM, or Infrai depending on operational preferences. Token counting during chunking and prompt assembly is recommended over character counts to enforce usage limits reliably.

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)

Basketball Browser Game's Hidden Weight Table Reveals a Perfect Positional Spectrum

Build a Hooper is a browser-based basketball simulation game where players draft athlete attributes across five positions, each governed by a published 13-attribute weight table. A mathematical analysis of the table found that the L1 distance between any two position columns increases consistently the further apart those positions are on the court, forming a near-linear one-dimensional spectrum. Small forward emerged as the most versatile position with weights spread across nearly all attributes, while center showed the most concentrated profile, heavily favoring interior skills. High-variance attributes like Blocks, Interior Defense, and Ball Handling proved to be the clearest markers of positional identity, with spreads ranging up to 13 times between their lowest and highest values. Notably, none of this structural pattern is explicitly documented in the game — it emerges entirely from analyzing the published numbers.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Developer builds TuxCleaner, an open-source safety-first disk cleanup tool for Linux

A developer has released TuxCleaner, an open-source terminal application designed to help Linux users manage disk space by cleaning caches, removing unused packages, and identifying large files across different distributions. The tool was created out of frustration with having to remember separate commands for apt, dnf, pacman, Flatpak, Docker, and various build tools. A core design principle separates the discovery and deletion steps, requiring explicit user confirmation and path validation before any files are removed. Every operation supports a dry-run preview, and automation flags can skip prompts but cannot bypass safety checks. Built with Ratatui for a persistent terminal interface, TuxCleaner aims to make routine Linux maintenance more approachable without sacrificing transparency or control.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Eight Terminal Tricks to Cut Daily Command-Line Repetition

A developer productivity guide published on DEV Community outlines practical terminal techniques designed to reduce repetitive typing and speed up daily workflows. Key recommendations include creating shell aliases for frequently used commands like git and Docker, and installing fzf to enable fuzzy searching through command history and files. The guide also covers tmux, a terminal multiplexer that keeps sessions alive after SSH disconnections and allows pane splitting for multi-project work. Additional tips include using ripgrep over the cat-pipe-grep pattern, leveraging directory-jumping tools like z or autojump, and customizing the shell prompt to display the current git branch. Collectively, these tools and habits aim to eliminate common inefficiencies for developers who spend significant time working in the terminal.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

GitLab Runner's Default concurrent=1 Setting Can Silently Stall Your CI Pipelines

A CI pipeline appearing stuck in a pending state for minutes can be caused by GitLab Runner's default configuration, not a system failure. GitLab Runner ships with a concurrent = 1 setting in its config.toml file, which limits all projects sharing that runner to a single execution lane at a time. This means one active build from any repository can block all others, while the runner still reports a healthy status. Engineers can identify the culprit by inspecting running Docker containers to read which project holds the active lane. The fix is straightforward: increasing the concurrent value to 2 or higher, then restarting the runner, typically unblocks queued pipelines within about 30 seconds.

0 comments Read more at DEV Community