Почему OpenAI Assistants не подошёл для сложных AI-агентов
OpenAI Assistants не хватило гибкости для сложной логики. Свой бэкенд превратился в операционный ад — 50% времени на поддержку вместо разработки. Agent Builder создавал vendor lock-in. Как мы решили эту проблему в Soft Skills Lab.
https://habr.com/ru/articles/993700/
#aiагенты #openai #llm #fastapi #edtech #gemini #claude #vendorlock
Python Trending (@pythontrending)
Grok2API를 FastAPI 기반으로 재구성한 프로젝트 발표. 최신 Web 호출 포맷을 완전 지원하며 스트리밍/비스트리밍 대화, 이미지 생성·편집, 고차원 추론('deep thinking')을 지원하고 계정풀 동시성 및 자동 로드밸런싱을 통합한 API 솔루션임.
Todays challenge: remove FastAPI and SQLModel. Its a great framework with great DX. But with the prototyping phase I learned that taking the "usual" choice is contrary to my approach of #HATEOAS. The cracks were showing.
IMHO FastAPI is a great tool for small apps. However, I can see already unnecessary complexity in keeping front and backend components flexible. Still sticking with Starlette (also great DX) and SQLAlchemy (SQLModel lacks behind in async).
I've finally released holm v0.9
The main feature is that you can now have your layouts in plain HTML files!
holm automatically loads HTML layouts, processes them using a simple, fully customizable pipeline, and swaps in child content for you.
The main benefit of this approach is performance: a large, mostly static HTML element tree can be replaced with single component.
If you're curious, check out the guides here: https://volfpeter.github.io/holm/guides/html-layout/
Ваша LLM стримит в никуда: разбираемся, как работать с дисконнектами в FastAPI
Что делает ваше FastAPI-приложение, когда клиент неожиданно закрывает соединение? По умолчанию — ничего. Если сервису важно экономить ресурсы (например, при работе с LLM), дисконнекты приходится обрабатывать явно, иначе GPU продолжит генерировать токены в пустоту. В FastAPI есть встроенные механизмы отмены обработки запроса при дисконнекте клиента. Однако просто включить их недостаточно: нужно подготовить ваш код к этому. Иначе возможны проблемы: зависшие транзакции и сломанный пул соединений с БД.
https://habr.com/ru/companies/tochka/articles/992134/
#fastapi #starlette #uvicorn #стримминг #sse #дисконнекты #отмена_запроса #asgi
New blog post is up! How I added search functionality to Rack Root, a FastAPI app.
I cover just about all the Python code I wish I had examples for when I was figuring this out. I go over the database class design, indexes/vectors a little bit, and the API design I ended up implementing.
#blogging #writing #development #search #rackroot #python #backend #fastapi #homelab #learning #databases #sqlalchemy
[GlyphLang - AI 최적화를 위한 프로그래밍 언어
GlyphLang는 AI 최적화를 위한 프로그래밍 언어로, LLM 코드 생성 효율성을 극대화하며 FastAPI 대비 23%, Java 대비 57% 적은 토큰 사용을 자랑합니다. 경량 문법, 정적 타입 시스템, 비동기 구문, 모듈 시스템, 고성능 런타임, 인프라 통합 기능, 보안 기능, 관측성 도구, 개발 도구 등을 지원하며 Apache License 2.0으로 공개되었습니다.
[Show GN: Streamlit의 직관성에 성능을 더하고 싶어서 시작한, Pure Python Web UI Framework 'Violit'
Violit은 Streamlit의 직관성과 성능을 결합하고 스타일링을 간편하게 한 Python 기반 웹 UI 프레임워크로, FastAPI와 shoelace(Lit)를 기반으로 합니다. 주요 특징으로는 Streamlit과 유사한 직관성, O(1) 반응성, 30가지 이상의 테마 지원, 하이브리드 엔진 및 멀티 플랫폼 지원 등이 있습니다. 현재 초기 단계이지만 긍정적인 반응을 얻고 있습니다.
Mạng kỹ sư tạo công cụ tự động ghi chép cấu hình switch, phân tích log, trích xuất VLAN, routing. Miễn phí 3 tài liệu/tháng, Pro $9/tháng không giới hạn. Thử ngay, phản hồi tính năng, vendor cần thêm. #Network #NetworkEngineering #Documentation #Python #FastAPI #ViệtNam #Mạng #CôngCụ
Cách xử lý refactor lớn với người dùng trực tiếp? Phân chia frontend thành ứng dụng riêng, thay đổi schema cơ sở dữ liệu và refactor backend. #Refactor #SaaS #React #FastAPI #Postgres #Redis #TechDebt #XửLýRefactor #PhátTriểnSaaS
https://www.reddit.com/r/SaaS/comments/1qqp1ru/how_to_handle_major_refactor_with_live_users/
https://github.com/chaosslabs/whisper-server-apple-silicon is a #FastAPI wrapper that server an #OpenAI compatible server to run #SpeechToText transcriptions in a machine in your network so you can use #Whisper to transform your audios into text :D
Open-Source AI Фитнес-Тренер: 27 MCP-инструментов, 3 провайдера и граф упражнений
Выделили из production-проекта и открыли в open-source PWA-приложение для персонального фитнес-тренера с AI. Дисклеймер: это open source, в нем могут быть недостатки, заходите, предлагайте идеи, исправления. Публикую тут в ознакомительных и образовательных целях. Выпилил этот кусок в open source из части личного проекта, о котором писал тут . Весь код писал полностью Claude Code на Opus 4.5 с thinking режимом. Обсудить
https://habr.com/ru/articles/990694/
#python #fastapi #llm #chatgpt #claude #nextjs #pwa #opensource #fitness #rag
#jerecrute enfin pas moi, mais dans une école où j'enseigne : https://discuss.afpy.org/t/developpeur-full-stack-oteria/2926
Yesterday, I came across a GitHub repo with an #FastAPI project template for authorization. It was #vibecode by an engineer in a mid-size company building software for government websites. Just a cursory glimpse into the codebase revealed the utter deep shit we are in.
How can we handle this? How can society handle this? This company will write code for websites that we potentially trust our credentials with. Users cannot see what happens underneath. #AI is a #war against society.
A few days I posted an article containing some Python/web-app benchmarks. Those benchmarks questioned the benefits of async for python web-apps. Jamie Chang has done his own benchmarks and discovered some interesting things that make async look better than it did in the first set of results. It's still not 'async good, non-async bad' but there are things that are worth knowing which make a big differences to the results.
I'm sure there's more to be known so I'm looking forward to the next round of investigations.
I do maintain that unless your use case is one where performance is paramount, the focus should be on reducing cost of development, and cost of maintenance, and that using async generally increases both of those.
https://blog.changs.co.uk/asyncio-is-neither-fast-nor-slow.html
🎙️ Python Bytes 467: Toads in my AI
with @mkennedy and @brianokken
https://pythonbytes.fm/467
#Python #GreyNoise #tprof #TOAD #FastAPI #Digg
🎮 Lenswalker Dev Update
Productive day! New features shipped:
@Mentions in Comments
- Type @ for player suggestions
- Click @name to open profile
- Mentioned players get notified Comment Improvements
- ⏱️ Timer shows remaining edit time
- 📜 "Load more" for long threads
- 🚫 Blocked users hidden from view
- 💬 Notifications show comment preview
Plus iOS viewport fixes after commenting.
#IndieDev #GameDev #SoloDev #WebDev #PWA #TypeScript #Python #FastAPI #BuildInPublic #DevLog
Assuming a REST backend and an app that talks a fair deal to that backend. Are there any real time saving benefits to using WebSockets for the queries instead of setting up new sockets all the time? I don't really see any big performance problems, but it would be interesting to see. Especially during startup some 20+ queries are performed (up to 50 for a totally new client app) and there has to be some connection and SSL overhead setting up all the sockets.
Đã tái tạo SubGrid thành một PWA dễ thương, hoạt động offline, đồng bộ dữ liệu qua thiết bị với xác thực passkey (không mật khẩu, không email). Cài được trên điện thoại và máy tính. Cảm ơn hoangvu12 vì ý tưởng gốc! #PWA #SubGrid #snoutpink #ứng_dụng_nhẹ #công_nghệ #passkey #offlineapp #Angular #FastAPI
https://www.reddit.com/r/SideProject/comments/1qlh1po/i_rebuilt_subgrid_as_a_cute_pwa/
I wrote up some notes on serving a monorepo with separate frontend (React, Svelte, etc.) and API-only backend (FastAPI, Django, etc.) apps with dokku, since I couldn't find much info on this setup specifically. I've been running it for awhile and it's working great.
