# Alpha Desk（阿尔法投研台）

> Alpha Desk is a preconfigured DeepSeek Harness (DSH) product for China A-share research: a team of 9 financial bots with 35 skills, a market center and dashboard, 6 research report types, scheduled tasks and memory templates — plus a report engine, deep memory, a team knowledge base and agent self-evolution (RSI, in development). It is a research tool / data terminal, not an investment advisory product, and does not constitute investment advice.
>
> Alpha Desk 是为 DeepSeek Harness 预配置的 A 股研究工作台：9 个金融 Bot、35 个技能、6 种研究报告，配有报告引擎、深度记忆、团队知识库与智能体自进化（RSI，研发中）。研究工具，不构成投资建议。

Status: v0.1.15-dev, developer preview. Runs on macOS and Windows. Requires DSH 0.2.0-rc.2+, Python 3.9+ and access to the private registry. The bots collaborate in a group chat: pulling data, running event studies, writing reports, backtesting, watching the market and reviewing decisions; every conclusion states its evidence, window and uncertainty.

## Key concepts

- **Report Engine (报告引擎, quant-report-engine)**: fetch → model → render → independent verify. Pure Python standard library; all numbers come from one metrics.json; output is a self-contained HTML file (inline SVG, no scripts, no external URLs). A separate module that never imports the computing code refetches and recomputes, and verify_html asserts every formatted number in the HTML matches raw data. Unreachable sources are recorded as failures, never filled with made-up data. Six sub-engines: technical compare, event commentary, stock snapshot, earnings review, industry compare, deep dive.
- **Deep Memory (深度记忆)**: two layers. Long-term bot memory (a Markdown MEMORY.md per bot plus topic files and a shared team memory; remember / forget / recall tools; automatic post-chat review; every change journaled in journal.jsonl; secrets, passwords and ID numbers are refused). Causal memory pairs each judgment with its real outcome after the validation window (decision → outcome); reads require as_of so future lessons never leak into past judgments.
- **Team Knowledge Base (团队知识库, Team KB)**: a Markdown knowledge base. Sources are split into pages with source hashes (ingest → draft → review → published → stale); a different reviewer must approve before publishing (writer ≠ reviewer); pages go back to review when a source changes. Chinese-aware BM25 search; bots query it through read-only MCP tools (kb_search, kb_get, kb_context, kb_lint, kb_stats).
- **RSI = 智能体自进化 / Agent Self-Evolution (Recursive Self-Improvement)** — in development, a standalone module not yet shipped in 0.1.15-dev. How it works: the bot team re-tests itself on frozen data snapshots (strict asOf cutoff, future data rejected); each run is scored on outcome 40 + process 45 + cost 15; a candidate may change exactly one file per A/B experiment, with a paired bootstrap 95% confidence interval; it is promoted to rsi/promoted/ only if the mean difference > 0, the CI lower bound > 0 and there are ≥ 10 samples; a human reviews and merges it, and any step can be rolled back. Settled improvements flow into the team knowledge base as drafts that still need independent review. RSI never merges its own results and never changes existing bots, profiles or trading logic; current cases are synthetic fixtures that validate the mechanism only.
- **Research only, not investment advice**: Alpha Desk does not recommend buying or selling specific securities, gives no target prices, ratings, return promises or trading signals, does not connect to brokerages or place orders, and does not resell market data. Data defaults to free astock / global sources; iFinD, Tonghuashun and other sources are bring-your-own.

## Pages

- [Alpha Desk 官网（中文）](https://deepseek-alpha.com/): full product page in Chinese
- [Alpha Desk (English)](https://deepseek-alpha.com/en/): full product page in English
- [Report Engine](https://deepseek-alpha.com/en/#engine): pipeline and verification details
- [Agent Self-Evolution (RSI)](https://deepseek-alpha.com/en/#rsi): loop, scoring, promotion gates and trust layer
- [Deep Memory](https://deepseek-alpha.com/en/#memory): long-term bot memory and causal memory
- [Team Knowledge Base](https://deepseek-alpha.com/en/#kb): review workflow and search
- [FAQ](https://deepseek-alpha.com/en/#faq): common questions
- [Early access / 申请内测](https://deepseek-alpha.com/#waitlist): join the beta waitlist

## Optional

- [llms-full.txt](https://deepseek-alpha.com/llms-full.txt): the complete site copy in Chinese and English
- [Sitemap](https://deepseek-alpha.com/sitemap.xml)