US Neural builds sub-10ms local semantic tool registry to replace LangChain routing

·1 views

Engineers at US Neural have developed an open-source tool called Mycelium, designed to replace LangChain's built-in tool routing for AI agents. The team argues that LangChain's AgentExecutor becomes slow and unwieldy when managing 50 or more tools, resulting in brittle logic and high latency. Mycelium achieves a cold tool-discovery latency of 9.6ms, compared to 194ms reported for BM25-based approaches. The project reframes tool discovery as a distributed networking problem rather than a task delegated to a large language model. The code and benchmarking scripts have been published publicly on GitHub for independent review.

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)

Truck Driver Turned Coder Discovers RSS After 20 Web Scraping Blocks

A Japanese truck driver teaching himself Python repeatedly hit 403 Forbidden errors while attempting to scrape news data from a major website using BeautifulSoup and Requests. Despite trying over 20 workarounds — including rotating User-Agent headers, adding randomized delays, and adjusting request intervals — the site's bot detection and web application firewall continued blocking access. After his twentieth failed attempt, he reconsidered his approach and discovered an RSS feed linked in the website's footer. Using the Python library feedparser, he retrieved the same structured data in just a few lines of code, with no HTML parsing or browser disguise needed. This experience became the founding moment for his AI summarization application, puoppo.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How a Silent HTTP 200 Error Exposed Deep Flaws in an AI Meme Generator

A developer building 'Punchline Ops', an AI-powered meme generation tool, discovered that a successful-looking HTTP 200 response from storage did not guarantee a valid image was actually saved. Inspecting the raw bytes revealed a broken file, prompting a complete overhaul of the system's validation logic. The fix introduced strict WebP header verification, binary byte checks, and a re-fetch of stored objects before any gallery record is created. The project also uncovered a separate data migration flaw where template metadata — including image dimensions and caption regions — had been lost while embeddings were preserved. Observability tooling was added to trace each pipeline stage, allowing engineers to pinpoint failures in retrieval, rendering, or storage without guessing at the model or prompt.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Common JSON Errors in JS, Python, and Postgres Mapped to Their Fixes

A comprehensive reference guide maps the most frequent JSON errors to their root causes and solutions across JavaScript, Python, and PostgreSQL. In JavaScript, common issues include unexpected tokens, circular structure references, and truncated documents when using JSON.parse() or JSON.stringify(). Python developers often encounter JSONDecodeError variants caused by single-quoted keys, Python dict literals, or non-JSON response bodies. PostgreSQL throws syntax errors when JSON values are inserted using single quotes, empty strings, or double-encoded data. Across all platforms, the most widespread mistakes include trailing commas, unquoted or single-quoted keys, JavaScript-style comments, wrong boolean/null literals, and invalid numeric formats like NaN or leading zeros.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Developer Builds Custom OpenTelemetry Layer to Debug Slow Streaming Voice AI

A developer identified a 3-second latency issue in a Groq-powered voice assistant that standard APM tools could not diagnose. To isolate bottlenecks across Speech-to-Text, LLM token streaming, and Text-to-Speech stages, they built a custom instrumentation layer called Zooid using OpenTelemetry. The solution used a Python decorator to manually manage trace spans across async streaming pipelines, capturing voice-specific metrics like Time to First Audio and per-turn token costs. Data was routed to SigNoz for visualization, though custom span attributes required additional schema handling to appear correctly in dashboards. The project highlights the gap between standard auto-instrumentation tooling and the complex, asynchronous nature of real-time Voice AI systems.

0 comments Read more at DEV Community