what is dev-pulse fr?

DevPulse is a project that lets you ask plain English questions about software release notes and get back answers that cite the exact version where a change happened. It was built as a capstone project for a course called LLM Zoomcamp. The idea addresses a real problem: developers rely on many open source tools that release updates constantly, and nobody has time to read every changelog, yet missing an important breaking change can cause problems in production. The system works by pulling release notes from the GitHub Releases page of ten popular open source projects, including Node.js, React, Next.js, Docker Compose, and TypeScript, covering close to 1,800 individual releases. Those notes are broken into chunks and indexed two ways: a traditional keyword based search and a vector search that captures meaning rather than exact words, with the two combined for better results and then reordered by a separate reranking step to surface the most relevant chunks. An AI agent built with Pydantic AI decides which project to search for a given question, can run multiple searches if needed, and always cites the specific release version behind each fact it states. A scheduled workflow refreshes this whole knowledge base automatically every week so answers stay current. The project includes a formal evaluation comparing different search methods, finding that combining keyword and vector search with reranking performed best, and comparing two different prompt styles for the AI, where a simpler prompt actually gave better answers than an overly cautious one. Every question and search is also logged with detailed tracing, feeding a built in monitoring dashboard that shows request volume, response times, token usage, estimated cost, and user feedback collected as thumbs up or down on each answer. Running the project locally requires a Python package manager called uv, plus Ollama for running the underlying language model, and involves downloading models, fetching the release data, building the search indexes, and starting a chat interface built with Streamlit. Docker Compose is also supported as an alternative way to run the whole system, including the scheduled weekly refresh job.