Knowledge Share
Technical articles, tutorials, and insights
One Open Source Project a Day (No. 225): NanoJev — A 0.6B Parallel Decision Model That Outputs Probability Distributions Instead of Generating Tokens
NanoJev is a lightweight recreation of the "Jev" parallel decision model system. Built on a Qwen3-0.6B backbone with dedicated decision heads, it takes states and questions as input and outputs complete probability distributions directly — with zero output-token decoding. It supports three decision head types (dynamic choice, boolean, ordered score) and has been validated on maze navigation and Snake game benchmarks. 954 Stars, MIT License.
One Open Source Project a Day (No. 220): WeKnora — Tencent's Enterprise Knowledge Framework, From RAG Q&A to a Self-Evolving Wiki
WeKnora is Tencent's open-source LLM knowledge framework, combining fast RAG Q&A, multi-step ReAct agent reasoning, and a Wiki mode that auto-generates an interlinked knowledge base. Supports 20+ LLM providers, 10+ document formats, and enterprise-grade multi-workspace RBAC — it's the core technology behind the WeChat Dialog Open Platform. 25.4k Stars, MIT License.
One Open Source Project a Day (No. 215): Langflow — Visual Drag-and-Drop AI App Builder
Langflow is a low-code AI application platform built on a visual flow editor. Connect LLMs, vector databases, tools, and Agents by dragging and dropping nodes, building production-ready RAG pipelines and multi-Agent systems without writing backend code.
DeepSeek Harness Series (04): Agent Loop — How a Conversation Turn Actually Runs
The user sends a message, the Agent replies and calls three tools — but what's actually happening under the hood? This article fully unpacks the dsh Agent Loop: the Turn/Step structure, how the driver works, what each extension point does, and where you can insert control logic.
DeepSeek Harness Series (05): Sessions and Memory — How Conversation History Survives
What happens when your Agent crashes mid-run? Does history survive a process restart? Can you rewind to an earlier step and try a different path? This article unpacks dsh's Session design: the append-only log, derived history, format versioning, the Fork API, and why the model never sees dirty failed attempts.
DeepSeek Harness Series (06): System Prompt Assembly — Engineering Dynamic Prompts
Before every model request, how does dsh stitch together prompt fragments from dozens of plugins into one coherent system prompt? This article unpacks ctx.systemPrompt's registration mechanism, section ordering, dynamic variables, automatic tool schema injection, and how Prompt Caching fits in.
DeepSeek Harness Series (08): Multi-Agent Collaboration — Subagents and Agent Teams
One Agent calls another Agent — the result can be plain text or structured JSON. This article covers dsh's multi-Agent mechanisms: launching subagents, continuable subagents, tool filtering, personas, and the experimental Agent Teams.
DeepSeek Harness Series (03): The Tool System — Giving Your Agent Hands
Tools are what separate an Agent from a chatbot. This article covers the complete dsh tool system: registering tools, type-inferred schemas, the three-phase execution pipeline (pre-execute / execute / post-execute), the approval mechanism for dangerous actions, and scope-based tool isolation. By the end you'll be able to write a production-ready custom tool.
DeepSeek Harness Series (01): What Is It — A Production Agent Runtime in Full View
DeepSeek Harness (dsh) is an open-source Agent runtime by DeepSeek AI. Its 'everything-is-a-plugin' architecture tackles the engineering problems that stand between an Agent demo and production. This opening article covers what dsh can do, how the architecture is structured, how it differs from other frameworks, and when it's worth reaching for.
One Open Source Project a Day (No. 207): AirLLM — Run 70B Models on a 4 GB GPU
AirLLM streams transformer layers one at a time to the GPU, making trillion-parameter models runnable on consumer hardware — no quantization, distillation, or pruning required. 4 GB VRAM for Llama 3 70B, 8 GB for 405B, 12 GB for DeepSeek-V3 671B, and just 3.72 GB for Kimi K3 2.8T. 33.5k Stars, Apache 2.0.
One Open Source Project a Day (No. 202): Needle 2 — A 14 MB On-Device Tool-Calling Model
Needle 2 is a 45M-parameter tool-calling model from Cactus Compute, packaged as a single 14 MB binary that runs in a fixed 28 MB of session RAM. It matches FunctionGemma 270M on tool-calling benchmarks at one-sixth the size, and runs at 300+ tok/s on sub-$200 Android phones.
Code Agent Anatomy (21): Extending from Scratch — Connecting a New LLM Provider
OpenAI, DeepSeek, Kimi, Qwen... new LLM services appear every few months, and MyCodeAgent already supports 10 of them. Connecting a new provider isn't about changing the call chain — it's adding one row to a table. This article explains the provider routing mechanism in full, then walks you through the complete onboarding process.