Multi-Model API Cost Governance with the Inference Router

Using LlamaIndex for RAG in Python – Real Python

Khóa học hướng dẫn xây dựng ứng dụng RAG (Retrieval-Augmented Generation) bằng Python với LlamaIndex, từ thiết lập môi trường, tải tài liệu, tạo và lưu trữ indexes tìm kiếm, cấu hình nhà cung cấp AI (local/remote) đến chạy truy vấn nhằm giảm thiểu ảo giác (hallucination) từ LLM.

Là người phát triển Python muốn xây dựng hệ thống xử lý thông tin tự động với chất lượng cao, LlamaIndex giúp bạn tự động hóa việc kết hợp tìm kiếm và sinh lời từ mô hình ngôn ngữ lớn, giảm thiểu sai sót và tối ưu hóa hiệu suất cho ứng dụng của mình.

Python Is Not Enough: Why Pythonistas Love Rust (Podcast)

A podcast episode recap covering why Python developers are adopting Rust, featuring discussion of Rust tooling (Ruff, uv, Polars, Pydantic core), how Rust's strictness benefits AI agent guardrails, and a skeptical take on vibe coding. The author argues real AI productivity gains are closer to 1.2-2x rather than 10x, warns about rubber-stamping AI-generated code, and emphasizes that deep engineering fundamentals outlast hype cycles. The post also promotes a 6-week Python-to-Rust cohort building a JSON parser with PyO3 bindings.

TokenSpeed-Kernel: Portable APIs and High-Performance Kernels for Multi-Silicon LLM Inference – PyTorch

TokenSpeed-kernel is an open-source, standalone subsystem that provides a clean layered API and registry system for LLM inference kernels across multiple hardware backends. It decouples the high-level runtime from hardware-specific kernel implementations using a decorator-based registration system where kernels declare their platform capabilities, tensor format signatures, and priorities. The selector then dispatches to the best available implementation at runtime. Using GPT-OSS 120B on AMD MI355X (CDNA4) as a validation target, the post demonstrates how Gluon-backed attention and MoE kernels achieve 1.6–3.6x end-to-end throughput improvements over portable Triton baselines, while NVIDIA paths (via FlashInfer/TensorRT-LLM wrappers) use the same public APIs. The AMD-specific kernels are published as a standalone pip package (tokenspeed-kernel-amd) reusable by other inference engines like vLLM.

Bitbucket Packages adds PyPI and NuGet support

Bitbucket Packages bổ sung hỗ trợ PyPI (Python) và NuGet (.NET), tích hợp quản lý package vào Bitbucket cùng các registry sẵn có (container, Maven, npm). Tất cả năm registry chia sẻ chung mô hình quyền, thanh toán và dung lượng, nhưng PyPI/NuGet chỉ khả dụng trên gói Standard/Premium trả phí.

Lập trình viên phát triển ứng dụng Python hoặc .NET sẽ tìm hiểu Bitbucket Packages để tiết kiệm thời gian quản lý và chia sẻ gói phụ thuộc một cách hiệu quả hơn, tránh việc phụ thuộc vào các nền tảng bên ngoài như PyPI hoặc NuGet độc lập.

Why I wrote PEP 832 -- virtual environment discovery

Brett Cannon explains the motivation behind PEP 832, a proposal for standardizing Python virtual environment discovery. The core problem is that tools like VS Code have no reliable way to know which workflow tool (uv, Poetry, Hatch, etc.) a project uses, or where its environments are stored. Cannon proposes a .python-envs file listing environment paths and a [workflow] table in pyproject.toml to specify a workflow server protocol (WSP) over JSON-RPC, enabling editors to communicate with any workflow tool in a standardized way rather than requiring bespoke extensions per tool.

Explicit Lazy Imports Are Coming to Python 3.15

Python 3.15 introduces explicit lazy imports via PEP 810, allowing modules to be deferred until their attributes are first accessed. This reduces startup latency and memory consumption, especially useful in large applications and CLI tools that conditionally use heavy libraries. The post demonstrates the feature using Python 3.15.0b1 (installable via uv or pyenv), profiles import times with the tuna tool showing pandas being skipped when unused, and walks through the internal mechanism using pdb breakpoints to observe sys.lazy_modules and sys.modules state changes during reification. PyCharm 2026.1 does not yet fully support the new lazy keyword syntax.

The Hot Path Belongs to GBDTs, Agents Own the Cold Path: A Payment-Fraud Benchmark

A reproducible benchmark comparing gradient-boosted decision trees (GBDTs) vs. LLM-based scoring for payment fraud detection across three dimensions: latency, cost, and determinism. On a single CPU core, GBDTs hit p99 latency of 0.15ms vs. ~1,200ms for LLMs — well outside the 100ms ISO 8583 authorization budget. Cost-wise, GBDTs run ~$54/hour at 50K TPS vs. $16,200–$351,000 for LLM tiers. Determinism is the most critical issue for regulated environments: GBDTs return identical scores on identical inputs while LLMs produce hundreds of distinct outputs even at temperature=0. The recommended architecture keeps deterministic tree ensembles on the synchronous hot path and deploys LLM agents on the asynchronous cold path for SAR drafting, evidence gathering, and agent-as-a-judge validation before human review. All benchmark code is open-source and reproducible on a laptop.