Post

HKUDS000: PENGYI_HKUDS_STUDYMAP

HKUDS000: PENGYI_HKUDS_STUDYMAP

这是 PENGYI_HKUDS_STUDYMAP 的第一篇。

1
HKUDS000 -> Study Map

这一篇不是单个项目详解,而是先给 HKUDS 项目宇宙建立第一张地图。

本地 HKUDS 工作区已经找到了:

1
E:\2026\B面\香港大学\PENGYI笔记\PENGYI superCODEX PROJECT笔记!\hkuds

这个工作区里已经集中保存了 HKUDS / 黄超老师 Lab 的公开 GitHub 项目本地副本。

当前本地索引显示:

ItemCount / Status
HKUDS public repos indexed87
standard shallow git clones84
Windows repaired clone1
Windows sparse clone1
Windows source snapshot1

所以 HKUDS 不是一个小项目,而是一个完整的 AI research ecosystem。

Why HKUDS

前面我们已经完成了 LLMQuant 第一阶段:

1
LLMQuant = AI-native finance research stack

HKUDS 的性质不一样。

我现在会这样区分:

1
2
LLMQuant = finance-domain research and workflow system
HKUDS    = AI research infrastructure and agent ecosystem

LLMQuant 更贴近金融任务:

1
2
3
4
5
market data
finance workflows
quant knowledge
trading agents
backtest / strategy research

HKUDS 更像底层 AI 研究能力地图:

1
2
3
4
5
6
7
8
agents
RAG
graph learning
recommendation
spatio-temporal intelligence
multimodal / video
AI research automation
LLM efficiency and reasoning

这对我们很重要。

因为我们要做的不是单纯量化项目,而是:

1
2
3
4
5
6
AI Scientist
  + Quant Research OS
  + R&D Agent
  + personal agent workspace
  + source-grounded knowledge system
  + public research output

HKUDS 正好给了 AI research infrastructure 的训练场。

Current focus

HKUDS 项目很多,不能一上来全拆。

第一阶段先聚焦三个:

1
2
3
LightRAG
Vibe-Trading
nanobot

原因很直接:

RepoFirst role for us
LightRAGknowledge / RAG / project memory infrastructure
Vibe-Tradingtrading agent and quant research workspace
nanobotlightweight personal agent shell

这三个刚好对应我们自己的三条主线:

1
2
3
Knowledge layer
Trading research layer
Personal agent layer

如果把它们和 LLMQuant 接起来,可以得到一个很清楚的系统方向:

1
2
3
4
LightRAG        -> Research memory / source-grounded retrieval
Vibe-Trading    -> Trading research / backtest / agentic quant workflow
nanobot         -> Personal always-on agent shell
LLMQuant stack  -> finance data / skills / quant knowledge / ecosystem map

Local snapshot

我本地看到的三个重点 repo 规模如下:

RepoFilesMarkdownPythonTypeScriptJSON
LightRAG6834840610023
Vibe-Trading1547374992757
nanobot6426241511816

nanobot 当前有本地未提交改动:

1
2
M nanobot/agent/tools/long_task.py
M nanobot/providers/openai_compat_provider.py

所以后续读 nanobot 时要注意:只读不覆盖,不随便 pull,不改动这些未提交文件。

HKUDS macro map

本地 REPO_MEANINGS.md 把 HKUDS 大致分成六条主线。

我会先采用这个分法:

LineMeaningRepresentative repos
Agent and automation systemspersonal agents, coding agents, coworker agents, research automationnanobot, AutoAgent, OpenHarness, ClawTeam, ClawWork, DeepCode, AI-Researcher
RAG and knowledge systemsretrieval, KG, multimodal RAG, video RAGLightRAG, MiniRAG, RAG-Anything, VideoRAG
Graph / LLM4Graphgraph foundation models and graph language modelsOpenGraph, GraphGPT, HiGPT, GraphAgent, AnyGraph
Recommendation systemsSSL, GNN, diffusion, LLM recommendation, explainable recommendationSSLRec, LightGCL, RecLM, RecGPT, XRec, LLMRec
Spatio-temporal / urban computingtraffic prediction, urban intelligence, ST graph learningUrbanGPT, OpenCity, GPT-ST, FlashST, EasyST
LLM efficiency / reasoning / multimodalreasoning transfer, compression, video agent, AI phoneLightReasoner, SepLLM, VideoAgent, OpenPhone, ViMax

这个宏观地图告诉我们:HKUDS 不是只有 agent。

它的底盘其实是:

1
agent + RAG + graph + recommendation + urban intelligence + multimodal

这对顶会研究非常有价值。

因为它不只是“产品 demo”,它背后有一组持续发表、持续开源、持续工程化的研究方向。

LightRAG

LightRAG 的一句话定位:

1
LightRAG = simple and fast graph-based RAG infrastructure

README 里讲得很明确:LightRAG 是一个 lightweight knowledge-graph RAG framework,也是 Microsoft GraphRAG 的高效替代方向。

它的核心不是普通向量检索,而是:

1
2
3
4
5
6
7
8
documents
  -> chunks
  -> entities
  -> relationships
  -> knowledge graph
  -> vector embeddings
  -> local / global / hybrid retrieval
  -> grounded generation

本地顶层目录显示它已经是一个完整工程:

1
2
3
4
5
6
7
lightrag/
lightrag_webui/
docs/
examples/
k8s-deploy/
prompts/
tests/

重点特性包括:

1
2
3
4
5
6
7
8
9
10
API server
WebUI
KG / vector / storage backend
role-specific LLM configuration
multimodal document processing
reranker
citations
document deletion and KG regeneration
PostgreSQL / Neo4j / MongoDB / OpenSearch and other storage options
Docker / k8s deployment

对我们来说,LightRAG 应该先放在:

1
Pengyi Research OS / Knowledge Memory Layer

它可以支撑:

1
2
3
4
5
6
paper memory
project memory
CV / PS / RP material retrieval
quant research notes retrieval
source-grounded answer generation
website learning archive search

后续 HKUDS001 可以专门拆:

1
2
3
4
5
LightRAG indexing pipeline
LightRAG query pipeline
storage abstraction
API server
how to connect it to Pengyi Research OS

Vibe-Trading

Vibe-Trading 的一句话定位:

1
Vibe-Trading = personal trading agent workspace

README 的定位是:

1
2
Your Personal Trading Agent
One Command to Empower Your Agent with Comprehensive Trading Capabilities

它对我们非常直接。

因为它不是抽象 agent framework,而是把交易研究工作流做成了可运行系统:

1
2
3
4
5
6
7
8
9
10
11
natural-language market research
market-data loaders
strategy generation
backtesting
reports
exports
persistent research memory
multi-agent research teams
broker / data connectors
MCP tools
Web UI

本地顶层目录:

1
2
3
4
5
agent/
frontend/
wiki/
tools/
scripts/

README 里最值得我们关注的能力包括:

1
2
3
4
5
6
7
8
9
10
11
self-improving trading agent
multi-agent trading teams
cross-market data and backtesting
shadow account
research autopilot
alpha zoo
run cards
MCP / API
persistent memory
provider routing
security hardening

这个项目和我们的 Quant R&D Agent 极度相关。

它已经接近我们一直说的闭环:

1
2
3
4
5
6
7
research question
  -> hypothesis
  -> signal engine
  -> backtest
  -> attribution
  -> report
  -> next research

后续 HKUDS002 应该重点拆:

1
2
3
4
5
6
7
8
agent runtime
tool registry
market data loader
backtest engine
Research Autopilot
Alpha Zoo
Shadow Account
MCP/API boundary

它也可以和 LLMQuant 的 data-mcpskillsMagents 做强对比。

nanobot

nanobot 的一句话定位:

1
nanobot = ultra-light personal AI agent shell

README 里说得很直白:

1
open-source, ultra-lightweight personal AI agent you can truly own

它不是专门做交易,也不是专门做 RAG。

它更像一个个人 agent operating layer:

1
2
3
4
5
6
7
8
9
10
11
CLI
WebUI
chat channels
tools
memory
MCP
model routing
automation
deployment
project workspaces
long-running goals

本地顶层目录:

1
2
3
4
5
6
7
nanobot/
webui/
docs/
bridge/
case/
scripts/
tests/

对我们来说,nanobot 最重要的是“轻”。

我们不是一上来就需要一个巨大的 agent platform。

我们需要的是:

1
2
3
4
5
6
能长期陪跑
能接工具
能记忆
能接 Telegram / WebUI / CLI
能处理长任务
能被自己读懂和改造

这正好对应我们的个人生产力系统:

1
Pengyi Personal AI Operating Loop

后续 HKUDS003 应该重点拆:

1
2
3
4
5
6
7
8
AgentRunner
tool call loop
memory system
MCP server/client
provider routing
channel integrations
goal mode
project workspace

First integration idea

如果只看三个项目本身,它们已经很强。

但真正有价值的是把它们合起来:

1
2
3
4
nanobot = personal agent shell
LightRAG = project memory and source-grounded retrieval
Vibe-Trading = quant research and trading workspace
LLMQuant = finance data, skills, knowledge, and ecosystem

组合后可以形成:

1
2
3
4
5
Pengyi AI Scientist Workbench
  -> nanobot runs the personal agent loop
  -> LightRAG stores papers, notes, project docs, CV/PS/RP materials
  -> Vibe-Trading provides trading research and backtest workflows
  -> LLMQuant supplies finance-specific data, skills, and knowledge structures

这就是 HKUDS 系列对我们最直接的价值。

不是“多看几个 repo”。

而是把它们变成自己的系统组件。

Reading order

第一阶段建议这样走:

PostProjectFocus
HKUDS000study mapoverall map and first three priorities
HKUDS001LightRAGRAG / KG / storage / API server / WebUI
HKUDS002Vibe-Tradingtrading agent / Research Autopilot / backtest / MCP
HKUDS003nanobotpersonal agent loop / tools / memory / channels
HKUDS004comparisonLightRAG + Vibe-Trading + nanobot + LLMQuant integration

这条路线足够聚焦。

HKUDS 有 87 个公开 repo,我们现在不需要把 87 个全拆。

先把这三个拆透,就已经能支撑我们做自己的:

1
2
3
AI Scientist Workbench
Quant R&D Agent
Personal Research OS

My takeaway

HKUDS 第一眼看上去很大。

但如果从我们的目标出发,它可以先压缩成三块:

1
2
3
LightRAG     -> knowledge memory
Vibe-Trading -> quant research workflow
nanobot      -> personal agent shell

这三块加上 LLMQuant 的 finance stack,正好组成我们下一阶段要做的系统骨架。

所以 HKUDS000 的结论是:

1
2
3
HKUDS is the AI infrastructure side of our Research OS.
LLMQuant is the finance-domain side.
Our next task is to connect them through our own Pengyi Quant R&D Agent.
This post is licensed under CC BY 4.0 by the author.