Switchboard Router Boosts AI Tool Selection Accuracy from 21% to 88%

·2 views

Developers have built Switchboard, a tool-routing layer designed to help AI agents manage large numbers of MCP (Model Context Protocol) server connections more efficiently. The core problem it addresses is that connecting many MCP servers to a single agent inflates token costs, reduces tool-selection accuracy, and creates operational fragility. Switchboard uses a four-stage retrieval pipeline combining dense and sparse vector search, cosine-similarity filtering, and an LLM judge to dynamically select the right tools per request. In testing against 70 realistic scenarios, the router achieved 85–90% accuracy compared to just 21% for keyword search alone, while reducing token usage for tool descriptions by 99.6%. The system also supports dynamic backend registration, reactive health detection, and a Redis-backed cache to keep the tool index current without redundant reprocessing.

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)

Developer Cuts Next.js App Load Time by 47% With Three Optimization Fixes

A Next.js web application was loading in nearly 8 seconds on mobile connections, with a bloated JavaScript bundle of 1.85 megabytes causing the slowdown. The development team identified three core issues: loading all heavy libraries upfront regardless of user need, serving oversized images to mobile devices, and unnecessary component re-renders triggered by top-level state changes. They addressed these by splitting code into on-demand chunks, optimizing asset delivery, and reducing redundant rendering. The fixes cut the main bundle to under 980 kilobytes and halved the first meaningful render time from 4.2 to 2.1 seconds. The team recommends auditing performance early, shipping code only when needed, and prioritizing image optimization as the fastest path to speed gains.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

DeepSeek's Smaller Flash Model Beats Its Flagship Pro on Agent Benchmarks

DeepSeek released V4-Flash-0731 last week, a 284-billion-parameter model with only 13 billion activated parameters per token — far fewer than V4-Pro's 49 billion activated parameters. Despite no changes to its architecture, the updated Flash model now outperforms V4-Pro-Preview on several agent benchmarks, including Terminal-Bench and DeepSWE. DeepSeek attributes the performance gains entirely to additional post-training, with no increase in model size. Because inference costs scale with activated parameters, Flash runs at roughly a quarter of Pro's compute cost while delivering comparable or better results on agent tasks. The model is MIT-licensed with open weights on HuggingFace, though some benchmark figures have not yet been fully independently verified.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

GGUF, GPTQ, AWQ: A Practical Guide to LLM Quantization Formats

Running large language models locally has become more accessible thanks to three quantization formats — GGUF, GPTQ, and AWQ — each suited to different hardware setups. GGUF, developed by the llama.cpp team, allows models to split workloads between GPU and system RAM, making it ideal for consumer laptops and Apple Silicon devices even with limited VRAM. GPTQ targets dedicated GPU environments, using a calibration dataset to compress model weights to 4-bit integers for fast inference, but requires all data to fit entirely within VRAM. AWQ, the newest of the three, improves on GPTQ's accuracy by identifying and preserving roughly 1% of critical model weights during quantization, making it well-suited for enterprise tasks like complex reasoning and code generation. Choosing the right format depends on available hardware — GGUF for flexibility, GPTQ for GPU-bound production APIs, and AWQ where accuracy is the top priority.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

AWS Launches Dogwood to Enforce Sequence-Aware Policies for AI Agents

AWS released Dogwood this week, an open-source policy language licensed under Apache 2.0 that adds temporal, sequence-aware authorization for AI agents. It extends Cedar, AWS's existing stateless authorization language, which can only evaluate individual requests in isolation without knowledge of prior actions. Dogwood addresses this gap by tracking event histories, allowing policies to count in-flight requests, sum transferred values, and verify approvals across a time window before permitting an action. For example, an agent restricted to $5,000 per hour can be properly rate-limited even across concurrent requests that Cedar alone would each approve individually. The reference implementation is available on GitHub for exploration, while production use requires additional infrastructure such as trusted timestamps and durable trace storage; AWS is currently seeking language design feedback rather than code contributions.

0 comments Read more at DEV Community