Regulated RAG

Zero-dependency RAG toolkit for regulated industries — BM25 + TF-IDF + RRF, no vector DB, no Embedding model, pip-install-free

Product Manual Q&A · Regulatory Policy Search · Compliance Advisory · Instant Knowledge Base

中文文档 | English

---

Why This Exists

Every RAG tutorial tells you to install LangChain, spin up a vector database, download an Embedding model, configure chunking strategies... and 2 hours later you still haven't asked a single question.

Regulated RAG takes the opposite approach: pure BM25 + TF-IDF with Reciprocal Rank Fusion, Python standard library only, zero configuration. It works out of the box with built-in knowledge bases for financial products and regulatory policies.

Feature	LangChain RAG / LlamaIndex	Regulated RAG
Setup time	30-120 min	30 seconds
External dependencies	Vector DB + Embedding model	None
Configuration	Chunk size, overlap, top-k, model selection	Zero config
Chinese support	Requires specific Embedding model	Built-in, Jieba tokenizer
Included knowledge bases	None	11 documents across 2 domains
Benchmark transparency	Varies	Hit@1=100%, <3ms
Regulated industry focus	None	Banking regulators + product manuals

---

Quick Demo

Product Manual Q&A — Ask About Any Financial Product

cd product-manual-rag
python scripts/rag_cli.py ask "大额存单的起存金额是多少？"

## Answer

大额存单的个人起存金额为20万元，机构起存金额为1000万元。

**Sources:**
- [大额存单] 第3条: "个人投资者认购大额存单起点金额不低于20万元..."

**Confidence:** 0.95 | **Retrieval time:** 1.2ms

Regulatory Policy Search — Query Banking Regulations

cd regulatory-policy-rag
python scripts/rag_cli.py ask "商业银行金融资产风险分为哪几类？"

## Answer

根据《商业银行金融资产风险分类办法》，金融资产分为五类：正常、关注、次级、可疑、损失，其中后三类合称不良资产。

**Sources:**
- [商业银行金融资产风险分类办法] 第四条: "商业银行将金融资产...

**Confidence:** 0.92 | **Retrieval time:** 2.1ms

---

Two Engines

1. Product Manual RAG

Query product manuals (wealth management, credit cards, loans, trusts, etc.) with source citation.

Feature	Detail
Retrieval	BM25 + TF-IDF dual retrieval with RRF fusion
Tokenizer	Jieba (Chinese-first)
Hit@1	100% on test set
Avg latency	< 3ms
Output	Answer + source document + section + confidence score
Included KB	6 product manuals (fund, deposit, trust, credit card, SME loan, wealth management)

2. Regulatory Policy RAG

Query banking and financial regulations from CBIRC, PBOC, and CSRC with compliance advisory.

Feature	Detail
Retrieval	BM25 + TF-IDF dual retrieval with RRF fusion
Tokenizer	Jieba (Chinese-first)
Auto-classification	Detects applicable regulator (CBIRC/PBOC/CSRC)
Output	Answer + policy source + compliance advisory + applicability
Included KB	5 regulatory documents across 3 regulators

---

Architecture

┌──────────────────────────────────────────────────────┐
│                  Regulated RAG                        │
├──────────────────────┬───────────────────────────────┤
│  Product Manual RAG  │  Regulatory Policy RAG        │
│  (6 documents)       │  (5 documents, 3 regulators)  │
├──────────────────────┴───────────────────────────────┤
│              Shared RAG Engine                        │
│  ┌─────────────┐  ┌─────────────┐  ┌──────────────┐ │
│  │  BM25       │  │  TF-IDF     │  │  RRF Fusion  │ │
│  │  Retriever  │  │  Retriever  │  │  + Re-rank   │ │
│  └─────────────┘  └─────────────┘  └──────────────┘ │
│         │                │                │          │
│         └────────────────┼────────────────┘          │
│                          ▼                           │
│              Jieba Chinese Tokenizer                  │
│                          │                           │
│                          ▼                           │
│            ┌─────────────────────────┐               │
│            │  WeCom / Feishu / CLI   │               │
│            └─────────────────────────┘               │
└──────────────────────────────────────────────────────┘

---

Quick Start

git clone https://github.com/yuzhaopeng-up/regulated-rag.git
cd regulated-rag

# No pip install needed! Pure Python standard library.
# (Optional) Install jieba for Chinese tokenization:
pip install jieba

CLI Usage

# Product manual Q&A
cd product-manual-rag
python scripts/rag_cli.py ask "基金的定投策略是什么？"
python scripts/rag_cli.py ask "家族信托的最低门槛是多少？"

# Regulatory policy Q&A
cd regulatory-policy-rag
python scripts/rag_cli.py ask "理财产品的销售管理有哪些规定？"
python scripts/rag_cli.py ask "反洗钱客户身份识别的要求是什么？"

# Run benchmark tests
python -m pytest tests/

Python API

from product_manual_rag import ProductManualRAG

# Initialize with built-in knowledge base
rag = ProductManualRAG()

# Ask a question
result = rag.query("大额存单的起存金额是多少？")
print(f"Answer: {result.answer}")
print(f"Source: {result.source_document} - Section {result.section}")
print(f"Confidence: {result.confidence:.2f}")
print(f"Retrieval time: {result.latency_ms:.1f}ms")

---

Build Your Own Knowledge Base

Regulated RAG is designed for any regulated industry, not just banking:

from rag_engine import RAGEngine

# Create a new RAG instance
engine = RAGEngine(knowledge_base_dir="my_kb/")

# Add documents (Markdown or plain text)
engine.index_documents(["policy_2026.md", "regulation_v3.md"])

# Query
results = engine.search("What are the reporting requirements?", top_k=3)

Applicable Industries

Industry	Example Use Case
Banking	Product manual Q&A, regulatory compliance
Insurance	Policy document search, claims guideline
Healthcare	Clinical guideline lookup, drug interaction
Legal	Case law retrieval, contract clause search
Government	Policy interpretation, citizen FAQ
Telecom	Service agreement Q&A, tariff comparison

---

Benchmark Results

Metric	Product Manual RAG	Regulatory Policy RAG
Hit@1	100%	100%
Hit@3	100%	100%
Avg retrieval latency	1.8ms	2.3ms
Knowledge base size	6 documents	5 documents, 3 regulators
Zero-shot accuracy	95%+	92%+

Run the benchmarks yourself:

cd product-manual-rag && python -m pytest tests/ -v
cd regulatory-policy-rag && python -m pytest tests/ -v

---

Comparison with Vector-based RAG

Dimension	Vector RAG (LangChain/LlamaIndex)	Regulated RAG
Setup	Install vector DB + download Embedding model	Clone and run
Dependencies	chromadb/faiss/qdrant + sentence-transformers	Python stdlib + jieba
Config required	chunk_size, chunk_overlap, embedding_model, top_k, score_threshold	None
Chinese support	Needs multilingual Embedding model	Native Jieba tokenization
GPU required	Recommended for Embedding	Never
Deterministic	No (Embedding variance)	Yes (BM25+TF-IDF are deterministic)
Reproducible	Depends on model version	Bit-for-bit reproducible
Best for	Large corpora (10K+ docs), semantic similarity	Regulated domains, exact-match compliance

Why BM25+TF-IDF wins in regulated domains: Compliance queries demand precise term matching ("反洗钱" must hit "反洗钱", not a vague semantic neighbor). Deterministic retrieval means audit-friendly, reproducible results — critical when regulators ask "how did you find this answer?"

---

Ecosystem

Repo	Description
financial-ai-skills	104 financial AI skills (rule engines)
soe-compliant-office	20 SOE-compliant office skills
skill-framework	L0-L4 skill governance framework
fintech-h5-demos	57 zero-dependency H5 demos
regulated-rag (this repo)	Zero-dependency RAG for regulated industries

Contributing

PRs welcome! Please ensure:

1. No company-internal information in code or knowledge bases
2. New knowledge bases use generic document names (e.g., BankX-Product.md)
3. Run python -m pytest tests/ before submitting
4. BM25+TF-IDF only — no vector DB or Embedding dependencies

License

MIT License — Free to use, modify, and distribute with attribution.