FICC006: FICC x AI Research OS - RAG / Agent / Quant Workflow
这是 PENGYI_FICC_MAP 的 FICC006。
前面的路线是:
1
2
3
4
5
6
7
FICC000 -> FICC 总地图
FICC001 -> Fixed Income / Rates / Credit
FICC002 -> Currencies / FX
FICC003 -> Commodities
FICC004 -> Daily FICC Brief Generator
FICC005 -> Event-to-Signal Workflow
FICC006 -> FICC x AI Research OS
这一篇要把前面的所有东西合成一个系统:
1
2
3
4
5
6
7
8
FICC x AI Research OS
= RAG memory
+ Graph knowledge
+ Agent workflow
+ Quant validation
+ Forecast ledger
+ Human review
+ Public-safe artifact layer
公开边界:
1
2
This is educational material and research infrastructure thinking.
It is not investment advice, trading advice, or an actionable alpha note.
中文边界:
1
2
这是公开学习笔记和研究系统设计,不是投资建议。
不包含内部数据、客户信息、未脱敏策略、实盘观点或可交易 alpha。
一句话总览
FICC x AI Research OS 的核心是:
1
用 RAG 管知识,用 Graph 管关系,用 Agent 管流程,用 Quant 管验证,用 Ledger 管复盘,用 Human Review 管质量和边界。
更工程化一点:
1
AI Research OS = memory + workflow + validation + evaluation + governance + artifacts.
它不是一个“自动交易机器人”。 它是一个研究操作系统:
1
2
3
4
5
6
7
8
9
10
research question in
-> retrieve evidence
-> build event context
-> generate hypothesis
-> validate quantitatively
-> diagnose bias
-> write report
-> human review
-> update memory
-> next research plan
为什么是 OS
单篇报告解决的是一次问题。
Research OS 解决的是长期积累问题。
如果没有 OS,每天会变成:
1
2
3
4
5
6
7
读新闻
写总结
产生想法
忘掉想法
重复读新闻
重复写总结
重复产生想法
有 OS 以后,研究变成:
1
2
3
4
5
6
每一次 brief 进入 memory
每一个 event 进入 event store
每一个 hypothesis 进入 research ledger
每一次 validation 进入 evaluation history
每一次 human review 进入 quality feedback
每一次 output 变成可展示 artifact
这就是从“学习”变成“资产”的关键。
总体架构
FICC x AI Research OS 可以拆成九层:
1
2
3
4
5
6
7
8
9
Layer 1: Data and Document Ingestion
Layer 2: FICC Knowledge Base
Layer 3: RAG Memory
Layer 4: Graph RAG
Layer 5: Agent Workflow
Layer 6: Quant Validation
Layer 7: Forecast / Research Ledger
Layer 8: Evaluation and Governance
Layer 9: Artifact and Presentation Layer
完整链路:
1
2
3
4
5
6
7
8
9
10
public data / notes / reports
-> normalized knowledge base
-> RAG and graph memory
-> daily brief agent
-> event-to-signal agent
-> quant validation
-> bias diagnosis
-> forecast ledger
-> public-safe report
-> human review
核心原则:
1
2
3
4
每一个结论都要可追溯。
每一个假设都要可验证。
每一个结果都要可复盘。
每一个公开输出都要可脱敏。
Layer 1: Data and Document Ingestion
第一层是输入。
FICC 的输入天然是多模态、多来源、多频率的:
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
market data:
yields, curves, FX spot, forwards, basis, futures, spreads, vol
macro data:
CPI, payroll, GDP, PMI, trade balance, inventories, central bank rates
policy text:
FOMC, ECB, BoJ, PBoC, speeches, minutes, official statements
commodity reports:
EIA, WASDE, OPEC reports, inventory reports, shipping and weather notes
research text:
daily briefs, public reports, educational notes, sanitized research memos
internal memory:
forecast ledger, signal cards, validation reports, human review notes
第一版不追求豪华数据源。
第一版要保证:
1
2
3
4
5
source clear
timestamp clear
asset clear
unit clear
public/private boundary clear
这比堆数据源更重要。
Layer 2: FICC Knowledge Base
知识库不是简单文件夹。
它应该按研究对象组织:
1
2
3
4
5
6
7
8
9
10
11
12
13
14
Fixed Income:
bonds, rates, yield curve, duration, DV01, credit spread, repo, swaps
FX:
spot, forward, swap, options, carry, basis, central bank divergence, capital flow
Commodities:
energy, metals, agriculture, inventory, futures curve, seasonality, geopolitics
Cross-Asset:
inflation, growth, risk sentiment, dollar liquidity, funding stress, safe haven
Research Methods:
event study, backtest, bias diagnosis, forecast ledger, human review
每个知识条目最好有 metadata:
1
2
3
4
5
6
7
8
9
10
topic
asset_class
asset
region
time_horizon
source_type
public_safety_level
last_updated
related_events
related_hypotheses
这样后面的 RAG 才能精确检索。
Layer 3: RAG Memory
RAG 的作用是把研究记忆调出来。
它要回答:
1
2
3
4
5
6
这个事件以前出现过吗?
这个变量通常影响哪些资产?
之前我们怎么解释类似市场反应?
有哪些反例?
哪些旧 hypothesis 相关?
哪些 validation report 可以复用?
RAG corpus 可以分四层:
1
2
3
4
5
6
7
8
9
10
11
Knowledge Notes:
FICC000 - FICC006
Daily Memory:
daily briefs and market snapshots
Research Memory:
signal cards, validation reports, bias reports
Reference Memory:
official reports, public documentation, public educational materials
检索方式不要只有 semantic search。
应该组合:
1
2
3
4
5
6
7
keyword search
metadata filter
semantic retrieval
time-window retrieval
asset-specific retrieval
event-type retrieval
ledger retrieval
例子:
1
2
3
4
5
6
7
8
9
Query:
USD strengthens while gold also rallies. How to interpret?
Retrieve:
FICC002 USD liquidity and safe haven section
FICC003 USD and gold section
previous daily briefs with USD up / gold up
risk-off event notes
real-rate counterexamples
这个检索结果会让 agent 不至于只靠当前 prompt 硬猜。
RAG 的输出要求
RAG 不应该只返回长文本。
它应该返回结构化 evidence pack:
1
2
3
4
5
6
7
8
EvidencePack:
query
retrieved_items
key_evidence
counter_evidence
historical_analogies
uncertainty
source_ids
每一个 retrieved item 都要保留:
1
2
3
4
5
6
7
source
date
topic
asset_class
quote_or_summary
confidence
public_safety_level
这对金融研究特别关键。
如果没有 source 和 timestamp,就无法判断证据是否可信。
Layer 4: Graph RAG
普通 RAG 擅长找文本。 Graph RAG 擅长找关系。
FICC 天然适合图谱,因为它是跨资产因果网络:
1
2
3
4
central bank -> policy rate -> yield curve -> FX forward -> carry
oil price -> inflation -> central bank reaction -> rates -> FX
USD liquidity -> EM FX -> credit spread -> risk sentiment
inventory shock -> commodity curve -> producer margin -> credit risk
核心节点:
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
Asset
Currency
Country
CentralBank
MacroVariable
Commodity
Curve
Event
Narrative
Hypothesis
Feature
Outcome
Evidence
Forecast
Report
核心关系:
1
2
3
4
5
6
7
8
9
10
11
12
Event -> affects -> Asset
MacroVariable -> drives -> Narrative
CentralBank -> sets -> PolicyRate
PolicyRate -> affects -> YieldCurve
YieldCurve -> affects -> FXForward
Commodity -> affects -> Inflation
Inflation -> affects -> CentralBankReaction
Currency -> forms_pair -> CurrencyPair
Evidence -> supports -> Hypothesis
CounterEvidence -> challenges -> Hypothesis
ValidationReport -> evaluates -> Hypothesis
Forecast -> reviews -> Claim
Graph RAG 能回答:
1
2
3
4
今天这个事件会影响哪些二阶变量?
这个 hypothesis 依赖哪些宏观机制?
哪个 evidence 支持它,哪个 evidence 挑战它?
哪些旧 validation report 和它相似?
这是 FICC Research OS 的记忆骨架。
Layer 5: Agent Workflow
Agent 负责把流程自动化,但不能失控。
推荐 agent 分工:
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
Ingestion Agent:
收集和标准化输入。
RAG Agent:
根据问题检索知识、证据和历史案例。
Graph Agent:
查找跨资产关系和因果路径。
Daily Brief Agent:
生成 FICC daily brief。
Event Agent:
抽取 structured events。
Hypothesis Agent:
生成 testable research hypotheses。
Quant Agent:
设计 feature / outcome / validation plan。
Skeptic Agent:
检查反例、偏差、过拟合和因果错误。
Report Agent:
生成 markdown / json / table outputs。
Human Reviewer:
审核质量、边界和发布。
关键不是 agent 数量。 关键是每个 agent 的输入输出必须固定。
1
2
3
4
agent output must be schema-valid
agent conclusion must be evidence-grounded
agent forecast must be ledger-tracked
agent public output must be human-reviewed
Agent Harness
FICC Agent Harness 应该规定:
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
Allowed Inputs:
public data
public documents
sanitized notes
approved local research memory
synthetic examples
Forbidden Inputs:
confidential bank data
client information
proprietary strategy parameters
live positions
unredacted alpha notes
Allowed Outputs:
educational explanations
research hypotheses
validation plans
public-safe reports
bias diagnostic reports
Forbidden Outputs:
direct trading advice
position sizing
live order instructions
undisclosed alpha details
client-specific recommendations
Validation rules:
1
2
3
4
5
Every claim needs evidence.
Every hypothesis needs counter-evidence.
Every feature needs timestamp check.
Every validation needs bias report.
Every public report needs redaction check.
这就是 harness 的价值:
1
让 agent 高效,但不越界。
Layer 6: Quant Workflow
Quant workflow 接住 agent 的研究假设。
路径:
1
2
3
4
5
6
7
8
event
-> hypothesis
-> feature
-> outcome
-> event study
-> backtest prototype
-> bias diagnosis
-> research verdict
和 FICC005 的连接:
1
2
3
4
5
FICC005:
Event-to-Signal Workflow
FICC006:
把这个 workflow 放入完整 AI Research OS。
Quant 层不应该直接追求“漂亮收益曲线”。
它先要输出:
1
2
3
4
5
6
7
validation report
bias diagnostic report
regime dependency report
feature version
outcome version
research verdict
next action
研究结论可以分级:
1
2
3
4
5
6
7
8
rejected
weak evidence
monitor only
promising for further study
needs better data
needs regime split
public demo only
private research only
这个分级比“赚没赚钱”更重要。
Layer 7: Forecast / Research Ledger
Ledger 是 OS 的长期记忆。
它至少要记录三类对象:
1
2
3
4
5
6
7
8
Claim Ledger:
daily brief 里的市场解释和预测。
Hypothesis Ledger:
event-to-signal 生成的 research hypothesis。
Validation Ledger:
每次 event study / backtest / bias diagnosis 的结果。
字段:
1
2
3
4
5
6
7
8
9
10
11
12
13
id
date
source
asset_class
claim_or_hypothesis
evidence
counter_evidence
horizon
confidence
validation_status
outcome
review_notes
next_action
Ledger 的核心作用:
1
2
3
4
防止只记住成功案例。
防止每天重复研究同一个问题。
让研究判断可以被校准。
让 agent 有长期记忆。
这也是 AI Scientist 的核心能力:
1
2
not just generate ideas
but remember, test, and improve ideas
Layer 8: Evaluation and Governance
AI Research OS 必须有评估。
评估分四层:
1
2
3
4
RAG evaluation
Agent evaluation
Quant evaluation
Human review evaluation
RAG 评估:
1
2
3
4
5
retrieval relevance
source grounding
timestamp correctness
coverage
counter-evidence retrieval
Agent 评估:
1
2
3
4
5
schema validity
evidence usage
reasoning consistency
overconfidence control
public-safety compliance
Quant 评估:
1
2
3
4
5
6
sample size
out-of-sample behavior
bias diagnosis
regime stability
transaction-cost awareness
feature availability
Human review 评估:
1
2
3
4
5
edit distance
review pass rate
common failure modes
redaction issues
repeated mistakes
Governance 不是为了降低速度。 Governance 是为了让系统能长期跑。
1
2
fast output without review becomes noise.
fast output with review becomes compounding asset.
Layer 9: Artifact and Presentation Layer
最终要沉淀成可展示资产。
公开层可以展示:
1
2
3
4
5
6
7
8
9
10
learning notes
system architecture
public-safe demo
sanitized templates
synthetic examples
validation framework
bias checklist
research reports
GitHub README
website posts
私有层保留:
1
2
3
4
5
6
unredacted signal cards
private data experiments
real parameter choices
private research ledger
collaboration notes
career transition materials
这就是 public credit 和 private alpha 的分离。
公开展示的重点不是:
1
我发现了什么可交易秘密。
而是:
1
我能把复杂金融研究问题工程化、系统化、可审计化。
MVP 设计
第一版 MVP 可以很小:
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
ficc_ai_research_os/
data/
sample_market_snapshot.csv
sample_events.csv
docs/
ficc000.md
ficc001.md
ficc002.md
ficc003.md
memory/
daily_briefs/
signal_cards/
validation_reports/
ledger/
forecast_ledger.csv
research_ledger.csv
src/
ingest.py
retrieve.py
graph.py
brief_agent.py
hypothesis_agent.py
validate.py
review.py
outputs/
daily_brief.md
signal_cards.md
validation_report.md
MVP 不要求接入真实交易系统。
MVP 只要求完成闭环:
1
2
3
4
5
6
7
8
9
input sample events
-> retrieve related FICC knowledge
-> generate daily brief
-> generate hypotheses
-> create signal cards
-> run simple validation placeholder
-> produce bias report
-> update ledger
-> create public-safe report
这个闭环足够展示:
1
AI + Finance + Quant + Research Engineering
System Interface
第一版接口可以是 CLI:
1
2
3
4
5
python -m ficc_os ingest --date 2026-07-03
python -m ficc_os brief --date 2026-07-03
python -m ficc_os hypotheses --date 2026-07-03
python -m ficc_os validate --hypothesis-id H001
python -m ficc_os review --date 2026-07-03
也可以是一个 pipeline:
1
ficc-os run-daily --date 2026-07-03
输出:
1
2
3
4
5
outputs/daily_brief_2026-07-03.md
outputs/signal_cards_2026-07-03.md
outputs/validation_report_2026-07-03.md
ledger/forecast_ledger.csv
ledger/research_ledger.csv
核心不是 CLI 多漂亮。 核心是每一步都有 artifact。
Pseudocode
系统主流程:
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
def run_ficc_ai_research_os(date: str):
inputs = ingest_public_safe_inputs(date)
knowledge = load_ficc_knowledge_base()
memory = load_research_memory()
ledger = load_ledgers()
evidence_pack = retrieve_relevant_context(
inputs=inputs,
knowledge=knowledge,
memory=memory,
)
graph_context = build_cross_asset_graph_context(
inputs=inputs,
evidence_pack=evidence_pack,
)
daily_brief = generate_daily_brief(
inputs=inputs,
evidence_pack=evidence_pack,
graph_context=graph_context,
)
events = extract_events(daily_brief)
hypotheses = generate_hypotheses(events, evidence_pack, graph_context)
signal_cards = build_signal_cards(hypotheses)
validation_reports = []
for card in signal_cards:
report = validate_signal_candidate(card)
report = diagnose_bias(report)
validation_reports.append(report)
reviewed_outputs = human_review(
daily_brief=daily_brief,
signal_cards=signal_cards,
validation_reports=validation_reports,
)
update_ledgers(reviewed_outputs)
publish_public_safe_artifacts(reviewed_outputs)
关键要求:
1
2
3
4
No artifact without source.
No hypothesis without mechanism.
No validation without bias report.
No public output without review.
和前几篇的关系
FICC004 是:
1
Daily Brief Generator
它回答:
1
2
3
今天发生了什么?
跨资产叙事是什么?
证据和反例是什么?
FICC005 是:
1
Event-to-Signal Workflow
它回答:
1
2
3
如何从事件生成可验证 hypothesis?
如何定义 feature / outcome?
如何验证和诊断偏差?
FICC006 是:
1
AI Research OS
它回答:
1
如何把 RAG、Graph、Agent、Quant、Ledger、Review 和 Artifact 合成长期运行的研究系统?
三者关系:
1
2
3
FICC004 = daily memory
FICC005 = research engine
FICC006 = operating system
面试怎么讲
如果被问:
1
你说的 FICC x AI Research OS 是什么?
可以回答:
1
我把它理解成一个面向 FICC research 的研究操作系统。底层是 public-safe 的 FICC knowledge base 和 research memory;中间用 RAG 和 Graph RAG 做证据检索和跨资产关系建模;上层用 agent workflow 生成 daily brief、抽取事件、提出 hypothesis、设计 feature 和 outcome;然后通过 quant validation、bias diagnosis、forecast ledger 和 human review 完成闭环。它不是自动交易系统,而是把金融研究过程结构化、可追踪、可验证、可复盘的系统。
如果被问:
1
RAG 在这个系统里有什么用?
可以回答:
1
RAG 用来把当前事件放回历史和知识上下文。比如今天出现 USD 和 gold 同涨,RAG 会检索以前的 safe haven、real rate、USD liquidity、risk-off 案例,返回 evidence 和 counter-evidence。它不是替代判断,而是给 agent 和研究员提供可追溯的证据包。
如果被问:
1
Agent 在这个系统里有什么风险?
可以回答:
1
主要风险是幻觉、过度自信、因果误判和越过金融合规边界。所以我会用 harness 约束 agent:固定输入输出 schema,限制数据源,要求每个 claim 有 evidence 和 counter-evidence,要求每个 validation 有 bias report,公开输出必须经过 human review,不允许直接输出交易建议、仓位或执行指令。
如果被问:
1
Quant workflow 在里面怎么接?
可以回答:
1
Daily brief 产生 structured events,Event-to-Signal workflow 把事件变成 testable hypothesis,再定义 feature、outcome 和 horizon,做 event study 或 backtest prototype,然后输出 validation report 和 bias diagnostic report。最后结果进入 forecast / research ledger,用于后续复盘和下一轮研究计划。
对我们的意义
这一篇对我们很关键。
因为它把几条主线统一了:
1
2
3
4
5
6
7
8
FICC domain knowledge
AI agent harness
RAG / Graph RAG
Quant research
RD-Agent style automation
Forecast ledger
Credit OS
Public website artifacts
这就是我们要做的能力组合:
1
2
3
4
5
6
懂金融场景。
懂 AI 系统。
懂 quant 验证。
懂工程交付。
懂公开呈现。
懂边界控制。
它不是银行里的重复流程。 它是可以迁移到 AI lab、quant team、research engineering、AI for finance 的系统能力。
下一步
FICC006 之后,最自然继续:
1
2
3
4
5
6
7
8
FICC007:
FICC Forecast Ledger and Evaluation
FICC008:
FICC Interview Playbook
FICC009:
FICC AI Research OS MVP Spec
如果按工程路线,下一篇可以做:
1
FICC007 -> Forecast Ledger and Evaluation
因为没有 evaluation,OS 只是自动生成内容。 有 evaluation,OS 才能真正进化。
当前结论
FICC x AI Research OS 可以压成:
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
RAG:
retrieve evidence and historical context
Graph:
model cross-asset relationships
Agent:
automate structured research workflow
Quant:
validate hypotheses and diagnose bias
Ledger:
remember claims, outcomes, and lessons
Human Review:
control quality, safety, and public boundary
Artifact Layer:
turn research into public-safe credit
一句话收束:
1
FICC x AI Research OS turns market knowledge into a compounding research machine.
中文:
1
FICC x AI Research OS 把市场知识变成能够持续积累、验证、复盘和展示的研究机器。